Attention over vulnerable brain regions associating cerebral palsy disorder and biological markers

Graphical abstract

Highlights

• •Associating body composition trends as biological traits with the brain structures of healthy and CP children.
• •Selection of appropriate MRI contrast for underlying BTs estimation and classification.
• •Introducing specialized learning modules to capture BTS-associated vulnerable regions.
• •Assisting Radiomics through our presented results and visuals obtained through specialized attention mechanisms.

Abstract

Introduction:

Cerebral palsy (CP) is a neurological disorder caused by cerebral ischemia and hypoxia during fetal brain development.Early intervention in CP favors medications and therapies; however, monitoring early brain development in children with CP is critical. It is essential to thoroughly examine brain-vulnerable regions associated with biological traits (BTs).Variations in BTs were evident in children with CP; however, it is critical to explore the BTs’ impact on the brains of healthy controls (HC) and CP-disordered children.

Objective:

This study associates BTs with HC and CP children.This study investigates the neurodevelopment of HC and CP underlying BTs. This study establishes a benchmark for the association of BT with HC and CP children.

Method:

The proposed AWG-Net is composed of customized spatial-channel (CSC) and multi-head self (MHA) attentions, where CSC blocks are incorporated at the first few stages, MHA at later stages, and cumulative-dense structures to propagate susceptible regions to deeper layers. The training samples include T1-w, T2-w, Flair, and Sag, annotated with age, gender, and weight.

Results:

The significant results for HC and CP are age (HC: MAE = 1.05, MCS₁₀=85.63, R₂=0.844; CP: MAE = 1.16, MCS₁₀=84.79, R₂=0.717), gender (HC: Acc = 82.98%, CP: Acc = 82.00%), and weight (HC: MAE = 4.65, MCS₁₀=56.30, R₂=0.78; CP: MAE = 2.85, MCS₁₀=70.24, R₂=0.82). Vulnerable regions for age are the cerebellar hemisphere, frontal, occipital, and parietal bones in HC and inconsistent in CP. HC and CP commonalities are in the frontal bone and cerebellar hemisphere with HC and discrepant in the occipital and temporal bones for CP. Similarly, gender differences are found for HC and CP.

Conclusion:

Age and gender are marginally less affected by the brain regions vulnerable to CP than weight estimation. T1-w is appropriate for age, weight, and gender. The learned trends are different for HC and CP in brain development and gender while slightly different in the case of weight.

Introduction

Cerebral palsy (CP) is a neurological disorder of motor function in the brain that affects body movement, balance, and posture of a person [1]. It affects the outer layer of the brain’s motor region, called the cerebral cortex, which controls muscular actions. The term CP is a broad one, and it is often believed to be caused by injuries to the brain in preterm infants or near childbirth. It is the most common motor disability in childhood, caused by abnormal brain development and affecting subjects’ ability to control their muscles. The etiology of CP can be both genetic and environmental factors, such as viral and bacterial intrauterine infections, intrauterine growth restriction, antepartum hemorrhage, oxygen deprivation, placental complications, complicated and prenatal exposure to toxins among others [1]. Besides the problems related to movement and posture, CP patients may also experience seizures, spine changes, joint problems, and intellectual disability, all affecting quality of life. Some children with CP may also be accompanied by epilepsy, behavior disorder, and visual and hearing impairment, which badly affects their daily life [2]. An estimated 3 to 10 out of every 1,000 infants develop some form of CP [1]. The incidence of CP is 10% to 15% in preterm infants and 0.2% to 0.25% in all children [3], and it is reported that 10% of the total population is estimated to be affected by some types of CP [4]. These symptoms feed into a negative cycle of declining mobility and may lead to developing chronic diseases [5].

Early diagnosis and the role of biological traits

It is vital to study the relevance of any biological traits (BTs) in brain images such as MRI to carry out comprehensive examination and therapy of CP in clinical settings [6], [7], where BTs can include age, weight, gender, or all these factors as whole. Initially, age, weight, and gender were selected to explore their association with the brain of healthy individuals and those affected by neurological disorders, specifically CP. Exploring the effects and influence of BTs on the brain of healthy and diseased children remains challenging [8]; however, altered body compositions as traits were evident in children with CP [9]. Szkoda extrapolated that BTs, including height, weight, body mass index (BMI) z-score, and BMI percentile, were reported to be associated with neurological disorders in children [9]. Among BTs, age has a significant role in early treatment for CP. Early age signs of CP in children usually include delays in achieving motor or movement milestones, in addition to more age-specific symptoms. Therefore, early-age diagnosis, specific trait estimation, and identifying developmental processes are much demanded, and these can lead to early therapy to prevent lifelong complications [10]. Early identification together with traits’ effects is critical to reduce complications in CP diagnosis and therapy [11]. Other traits, including gender [7], [12], [13] and weight [14] are also crucial for guidance diagnosis, medications, supportive treatments and physical therapy, and surgical procedures to alleviate symptoms and improve motor skills [15]. Therefore, this study aims to explore the role and effect of BTs with healthy and CP children from their brain MRIs to assist radiomics in diagnosis, prognosis, treatment, medication, and therapy.

Brain development

CP results in irreversible damage to fetal brain development [7], but early and accurate identification can minimize long-term consequences. Previous studies on CP have primarily examined early diagnosis, brain injury characterization, and neurodevelopmental outcomes in infants [7]. However, none of these studies have explored the patterns of brain development in both CP and healthy subjects, specifically regarding the progression of brain lesions and other regions of interest (RoIs) [7]. Additionally, brain imaging can reveal the brain age gap [16], which is the difference between biological and chronological ages and can be linked to neurological disorders [17]. Research [7] discovered that brain age development in patients with CP may be affected by the early age superior post-stroke outcomes and focal damage.

Studying early brain development and its associations may facilitate treatments and further examination [15]. Sumita et al examined the difference between predicted age and brain age and associated the difference with the risk of CP [18]. The infants later diagnosed with CP were more likely to have an estimated younger brain age in ’Early’ scans, where the estimated age deviations were significantly different compared to those infants who were not diagnosed with CP. Linking brain development or aging specifically at an early stage is crucial, because children have a greater potential for recovery from neural disorders [11]. Atlas-based methods from brain MR images have proven useful for assessing pathologies in fetal brains and estimating biological age and its deviation from the nominal gestational age [19]. After estimating brain age, another important factor is the difference between the predicted age and the actual age. According to [20], studying the brain and its regional predictors reveals that older brain age is linked to neurodegenerative diseases and increased mortality rates.

Neuroimaging can estimate brain age based on structural and functional information exploring the brain health, which either appears older, younger, or showing signs of neurological disorder [6]. A positive correlation suggests reduced mental and physical fitness [21], cognitive impairments [22], and traumatic brain injuries [23]. Negative correlation is associated to neurological disorders [24]. Individuals with abnormal BTs often exhibit variations in brain age and brain health due to gray and white matter atrophy, which can be associated with CP [25]. The optimal diagnosis age for CP is between one and six months after birth, but the patients are too young to cooperate with doctors [26]. Abnormal brain development is common in children with CP, but there is no recent report on the actual brain age of children with CP [7]. Concerning the description of brain development in patients with CP, it is more common for researchers to segment white matter damage and gray matter atrophy in patients with CP through MRI [27]. Still, it does not specify the exact brain age of patients with CP after brain atrophy. It is hypothesized that the trajectory of brain development in CP patients varies according to the level of brain damage as a lesion [27]. Therefore, it is essential to study brain health with age progression in both healthy and neurologically disordered subjects.

Body composition as weight and association with brain visuals

Brain morphology also alters in subjects with obesity [28]; however, it is still unclear if a significant correlation between BMI and brain morphology exists. Traditionally, obesity or overweight has been associated with cardiovascular disease, diabetes, and hypertension [29], where recently, obesity has been attributed to significant brain atrophy and cognitive impairment [29], which leads to the reduction in brain volumes [29]. It is necessary to investigate the association between BMI and brain structure for healthy subjects and with those subjects diagnosed with neurological disorder. Steegers provided evidence for an association between BMI and cortical morphology in a large sample of school-aged children drawn from the general population [28]. Cortical thickness is positively associated with BMI, providing evidence that a normal BMI during childhood is associated with more typical measures of brain surface morphology, which may equate with more optimal brain development. The findings provide evidence that overweight is associated with smaller subcortical gray matter volumes and impacts brain physiology at multiple levels, including cortical and subcortical abnormalities [30]. These studies provided consistent evidence of smaller cortical thickness or reduction in the gray matter volume in obese; however, the investigated brain regions vary across different studies [31]. A lower cortical thickness, lower volume of cortical regions, and affected prefrontal cortex were also found to be associated with obese children [32]. Hamer measured obesity and found it related to lower gray matter volume and brain volumes, including caudate, putamen, pallidum, and nucleus accumbens, where no apparent associations between obesity and white matter were detected [33]. .

The BMI has been associated with an increased risk of accelerated cognitive decline and dementia, causing neurobiological changes. The associations between BMI and brain structure, including overall and regional brain volumes and white matter microstructure, were assessed via MRI in a sample of the general population [34]. Children and adolescents with high BMI also exhibit decreased volume of frontal and limbic cerebral gray matter regions and decreased hippocampal volume compared to children with healthy BMI [35]. Hidese investigated the correlation of BMI and bilateral cerebellum exterior volumes and is found lower in the obese group, where bilateral cuneus and calcarine cortex, left cuneus, and left precuneus volume are lower in the underweight group [36]. In general, bilateral frontal and temporal areas, basal nuclei, and cerebellum are more commonly involved. BMI was positively associated with ventricular, pallidum, and amygdala and negatively with pallidal volumes [37]. In addition, obesity was associated with white matter microstructure [38], which needs further research on how and why deficient and additional weight influence the brain structure and association with neurological disorders. The study found structural differences in the hypothalamus of obese people, where the hypothalamus is involved in controlling appetite [39]. Brown showed the relation between obesity and impaired circulation to and in the brain, brain atrophy, and decreased cognitive functioning [40]. Changes in adipokines, gut hormones, and gut microbiota partly mediate the relationship between body weight and the brain. Individuals with bipolar disorders frequently suffer from obesity, yet the effects of obesity on brain structure in bipolar disorders are under-researched [37].

Gomeze underlined the significance of neuroinflammatory and neurodegenerative mechanisms linked to obesity [31]. Impaired neurological function in the central nervous system may be a consequence of structural changes in the brain and decreased cerebral integrity, particularly in the hippocampus [41]. A negative association was found between obesity and neurocognitive functioning, such as executive functioning, attention, visuospatial skills, and motor skills. Yilmaz investigated the body weight percentile of the children and determined the association of the body weight percentile of children with CP [14]. Semsek aimed to examine the relation between BMI and functional level and health-related quality of life in children with CP [42]. A significant difference was found in children with CP compared to normal body weight [42]. BMI affects functional independence and health-related quality of life in children with CP [42]. Children with CP have demonstrated higher rates of overweight and obesity than their typically developed peers [43]. Kuula found more structural abnormalities in very low birth weight participants [44]. Preterm, very low birth weight adults had a higher prevalence of brain abnormalities, including absolute brain volumes and gray matter structures against their term-born siblings [44]. It is evidenced from the findings that a neurobiological interaction between obesity and brain structure exists [30]. Being overweight negatively affects brain function and structure. However, the mechanisms involved are not entirely understood. The exact underlying mechanisms remain unknown, and further research should be performed and elaborate evidence to the relevancy of brain structure with weight [45]. The existing studies were not carried out using brain visuals to associate the brain structural variations in both health controls and CP patients with weight. This study aims to carry out CP association with BT as weight and visualize the vulnerable regions from brain structure using deep learning (DL) approaches.

Gender impact on neurodisorder

Besides age and weight associations with the structure of the brain MRI of children, it is also of high importance to elaborate on gender differences in both healthy and CP subjects. The brain age and overweight estimation are specifically studied for subjects with different genders, and the researchers found slower diffusion along the nerve fibers [46]. Among these, the female participants are found to have increased movement across the fibers. The combination of overall and central obesities is higher for men compared to women, which has been linked with gray matter levels [33]. A study using voxel-based morphometry and gender-related stratification showed that BMI was significantly and negatively related to bilateral cerebellum exterior volumes only in women [36]. Since women are more susceptible to obesity than men, it seems plausible that neural correlates may also be different and might be reflected in the brain’s white matter structure [46]. The stuyd [46] showed systematic sex-related differences; as yet, it needs to be clarified which microstructural changes are actually present.

Since there is evidence for sex-related differences in cognition and functional connectivity in youngs [47], however, no gender-related differences were observed in children with diplegia or quadriplegia, both for motor and cognitive functions [48]. It is pointed out that gender might influence differently the psycho-motor development of children with hemiplegia and those with a more severe clinical involvement such as diplegia and quadriplegia [48]. The sex differences in cognitive profiles are related to multivariate patterns of resting-state functional connectivity MRI [47]. The cognitive profile from functional connectivity data regarding sex has been attempted by using support vector machines [47]. However, no such distinguished patterns were observed. In addition, the brain age difference for male and female subjects was found with minor differences [49]. Jonssan introduced a 3D CNN model borrowing ResNet design and fused gender information for the prediction of brain age from T1-weight (T1-w) MRI[50].

Brain aging occurs in patients with CP; however, it is inconsistent with gender differences. In patients with CP, the brain tolerance of female patients to damage factors is higher than that of male patients [7]. CP is more frequently seen in males [51] belonging to different ethnicities [52]. Most children diagnosed with CP have the spastic variety [53]. The existing studies have attempted to associate brain age and gender with CP using a 2D convolution operation from T1-w contrast [7]. However, no inclusion of children below 5 years has been made where children below 5 years have high potential from brain lesions. The CP prevalence is associated with factors such as gestational age, weight, and birth weight [43]. Open investigated weight, gender, and age associations; however, no study relates BTs with CP patients and healthy subjects [30]. The existing studies obtained lower train accuracy and higher testing accuracy, which may reflect over-fitting in a 2D DL model training. The 2D CNN may not fully reflect the age, weight, and gender effects of neurological disorder (CP) and also needs to differentiate such effects for health controls and CP. Hence, besides age and weight as BTs, this study proposed a DL model to learn and reflect gender-related differences in normal and CP subjects using structural brain MRI.

Brain MRI as single modality and contrast selection

Brain MRI is one of the techniques to identify early markers of motor or cognitive outcome [17] and critical to finding CP-specific outcomes [54]. MRI can provide images with various contrasts, and the shared information among these contrasts can help determine lesions and affected regions [55]. The MRI modality is the leading imaging method for brain aging, diagnosing neurological disorders, and clinical pediatrics, providing insights into brain pathology and clinical findings [26]. MRI is useful for diagnosing genetic conditions and understanding the causes of CP [56]. Hyperintense punctuation of CP lesions and their correlations with clinical outcomes are critical [57]. MRI is used to locate and evaluate the severity of brain lesions, aiding in the identification of patterns in vulnerable brain development [58]. MRI can help monitor brain lesions and treatment effects with variations in BTs in healthy and CP children [59]. The brain MRI can be used as a single modality for the estimation of BTs, including age, gender, and weight, while excluding additional medical tests [50]. Recent studies have demonstrated that MRIs can be utilized to associate different BTs and brain structures of subjects with reasonably good accuracy [21]. However, the MRIs of children may experience additional noise and artifacts due to the uncooperative movements, which make the therapy and diagnosis challenging [60]. Therefore, it is critical to investigate and capture brain structural development or vulnerable regions to assist an effective therapy. Therefore, this study adds a variety of explicit noises and artifacts to train a robust and generalized DL model.

Different MRI contrasts can be acquired, such as T1-w, T2-weighted (T2-w), Sagittal (Sag), and Flair, to capture white and gray matters in the brain, which can be utilized to learn structural information regarding the brain lesions and associations [55]. Not all desired image contrasts may be available due to limitations in scan time, signal-to-noise ratio, and image artifacts [61]. Optimal contrast selection highlights the importance of studying the association of CP with BTs. New MRI techniques have revealed insights into neurological diseases; however, choosing the most relevant contrasts for a specific pathology remains challenging. The optimal selected contrast provides valuable insights into human tissue, pathology, and brain diseases and is extensively used in clinical diagnosis [62]. Selecting the most informative MR scan is crucial to avoid long acquisitions, motion artifacts, and high costs, benefiting the developmental process [20]. Different contrasts have been utilized to investigate abnormalities and association, such as T1-w and T2-w for age estimation [63], Flair for pathology and abnormalities distinguishing [15], and Sag for CP diagnosis [64]. Different contrasts have a different association with the structural composition of the brain; particular contrast appropriateness and selection would ease future research [50]. Selection of the most relevant contrasts for a given pathology remains challenging. Therefore, in this study, contrast selection has been carried out appropriately for each BT association.

BTs estimation using recent deep learning models

Recently, DL models have been employed in a wide range of applications, including neurological disorder recognition [65], disease classification [66], [67], associating neurological with BTs [8]. The machine learning-based investigation may determine what a healthy brain normally looks like at different life levels, assisting medical professionals in examining the severity of a disease [8], [68], [69]. The traditional way to estimate BTs is to extract features from brain MRIs, followed by classification or regression analysis [23]. The drawback of feature extraction methods is the loss of information since the features are likely designed differently for extracting information relevant to BTs. Recently, the DL-based methods have garnered much interest [70] by introducing convolutional neural networks (CNNs) [70] to be applied for BTs estimation specifically for aging [23]. Therefore, the learned DL models provide clinical information regarding BTs and respective deviations from typical brain development caused by neurological disorder [49]. Traditionally, DL methods require several specialized and often sophisticated preprocessing steps, making them impractical for broad use in clinical settings.

The recent DL strategies achieved outstanding performance in clinical practices and improved diagnostic accuracy to fully assist physicians in diagnosis and therapy[71], [72]. Some of the DL models demonstrated that MRIs can predict BTs as chronological age with reasonably good accuracy [21]. In addition to the 2D-CNN based DL models for prediction [73], 3D-CNN models received much attention due to the suitability for learning gray and white matters from 3D brain MRI health against BTs [63]. The 3D CNN models can accurately associate BTs in healthy individuals from structural brain MRI [74]. Normal brain development as BTs is accompanied by patterns of neuroanatomical compositions and changes that DL models can capture. The neuroscience community has widely adopted the construct of MRI-derived brain as an informative biomarker of brain health at the individual level [23]. The DL models have offered numerous other MRI-based biomarkers of neuroanatomical changes [49]. Because of their high sensitivity, the DL models smartly capture brain changes that occur with vulnerabilities, which is useful for early diagnosis of brain health in healthy subjects and with subjects having neurological disorders, specifically CP patients.

Every child with CP has a unique composition of neurological symptoms, motor severity, and accompanying impairments, constituting their functional profile (patterns) [54] where such patterns can be efficiently learned through machine learning [75], [76]. The signs and symptoms of CP usually appear in the early months of life, but the average age for diagnosis is around 2 years [10]. However, McIntyre study included children older than 5 years [10]; however, it is extremely challenging to associate brain developmental processes below the age of 5 due to the high potential of children’s recovery. Zhang predicted the brain age of children with CP but observed with overfitting [7]. The overfitting may caused by the employment of 2D CNN for the 3D MRI modality. The 2D CNN may not fully capture BTs’ effects on the brain structure of CP subjects. Similarly, Jang proposed an M3T-based 3D classifier with significant results; however, the underlying structure is limited to equal dimensions (i.e., L = H=D) because available datasets may contain variations in dimensions [77]. Traditional ways rely on the extracted features from brain MRIs [23]; however, they lack sufficient visual explanations of BTs association in healthy and CP. The brain MRI can demonstrate the physical compositions of the brain, which plays important roles in region identification and BTs association [78]. The DL models learn such compositional information from MRI, whereas clinicians can utilize it to observe minute details regarding lesions, which would be impossible. Therefore, an intelligent system is demanded to capture such composition.

BTs related brain sensitivity and visualization maps

The use of data-driven learning techniques resulted in improved accuracy for identifying CP [79]. Several studies have classified CP and captured CP-related characteristics using brain MRI [80]. Identifying vulnerable regions as brain compositions from brain MRI can assist clinicians in the diagnosing and rehabilitation process in neuro-disorders at an early stage [81]. Children with neurological disorders may experience extended delays in brain development compared to their healthy siblings. Previous studies have utilized CNN to estimate BTs (Age) directly from MRI scans in adults [82]. However, the application of these techniques in assessing the risk of CP in preterm infants remains unexplored. Levakov produced ”explanation maps” by assigning pixel/voxel-wise contributions to the prediction in an image [20]. In [20], a threshold was applied to the explanation map to identify regions that had the greatest impact on the model’s prediction. Therefore, attention has a key role in the intended propped AWG-Net model, not only to learn BTs’ associated regions but also to assist in the generation of visual maps. Attention mechanisms in deep learning are valuable for identifying important areas and generating explanation maps. Children with CP have individual neurological symptoms, motor severity, and accompanying impairments that make up their functional profile and compositions [54]. Such compositions as vulnerable patterns are helpful in training DL towards robust results [58]. Limited methods exist for the early identification of visual compositions associated with brain regions, specifically in CP patients. Therefore, it is necessary to use a sophisticated DL model with customized attention mechanisms.

Proposed study

This study primarily set a benchmark to associate BTs (i.e., age, gender, and weight) with brain structural composition to assist in the examination and diagnosis of both healthy and neurologically disordered children. There are a number of neurological disorders, among which we only focus on CP and its association with BTs. The entire processes, including preprocessing, contrast selection, training, and picking the right evaluation strategy, are outlined in Algorithm 1. The proposed generalized model, namely AWG-Net, trains on BTs labeled datasets with four dataset versions corresponding to the four contrasts, as depicted in Fig. 2. Our study employs DL training to estimate the age weight, classify gender from brain MRIs, and identify the sensitive regions to these BTs both in healthy and CP children. We employed spatial channel attention with customization (C-SCA) towards generality and capturing sensitive regions at distinct depth levels. In parallel, the input 3D brain MRIs pass through two ways of learning, with the inclusion of C-SCA block and direct (dense) connections (Fig. 2(a)). The details of C-SCA are shown in Fig. 2(b-c), where details are described in the methodology section. Our introduced C-SCA is a specialized attention approach based on the utilization of spatial and channel-wise attention and a fully convolutional network, consist of two further phases to capture prominent regions and avoid learning complexities. In the deeper layer, the multi-head attention (MHA) robustly captures the sequential information along the slices with proper embeddings as required (Fig. 2(c)). We examined the captured sensitive brain regions linked to BTs and their susceptibility to CP. We visualize the learned features for BT effects and their flow from upper to lower levels along the slices for each BT. Additionally, we illustrate the BTs-based distribution of susceptible brain regions. The following are the key contributions of our research:

• •Associating BTs with brain structures of healthy and CP children.
• •Selection of appropriate MRI contrast for underlying BTs estimation and classification.
• •Introducing specialized learning modules to capture BTS-associated vulnerable regions.
• •Assisting Radiomics through our presented results and visuals obtained through specialized attention mechanisms.

Fig. 2.

The proposed AWG-Net architecture. (a) Single contrast block (SCB) receives MRI and passes to customized spatial channel attention (C-SCA) and deeper SCBs. Multi-head attention (MHA) has been utilized at the deeper level. MHA outputs and dense connections are merged. The stacked SCBs are followed by fully connected (F.C) and dropout layers and end with an N number of neurons. (b) C-SCA is composed of two phases (P-1 and P-2). P-1 captures vulnerable regions utilizing GA-P, GM-P, and AA-P. P-2 starts by transforming 3D slices into 2D slices and then uses fully convolutional (F.C) as an encoder-decoder structure embedded with a max-pooling and global average pooling block (MAB). (c) Each MAB is made up of GA-P and GM-P along the channels, followed by a sigmoid function and an elementwise sum. (d) The MHA block. e) The 3D structure is divided into 2D slices with integration of separation (sep), classifier (cls), and positions (pos).

Methodology

This study aims to estimate the BTs of both healthy and CP children using brain MRI contrasts. The underlying MRI includes T1-w, T2-w, Flair contrasts, and a Sagittal (Sag) view of the brain. The entire study flow diagram has been shown in Fig. 1 composing of five major steps. The study comprised contrast selection and training of an appropriate DL model, which is illustrated in Algorithm 1.

Algorithm 1

Pseudocode of outlining the entire framework

Fig. 1.

Flowchart of the proposed study. The whole study is divided into five steps. Step 1 receives data from the database. Step 2 prepares the dataset. Step 3 performs dynamic augmentation. Step 4 carries out the training of AWG-Net over the input training split. Finally, Step 5 evaluates the output for prediction, classification, and visualization. The detailed implementation is outlined in Algorithm 1.

The proposed algorithm outlines the entire intended study. The pseudocode receives data (Dt) and produces output (X) with arguments (CT, TrT, and St) correspondingly for contrast type (CT), trait type (TrT), and healthwise subject distribution as status (St). First of all, we align all the MRI datasets (Dt$\in R^{\mathit{HWDC}}$) into the same dimensions in terms of Height (H), Width (W), Depth (D), and Channel (C) into new dimensions ($r-\mathit{Dt}\in R^{n.\mathit{HWDC}}$) as $n.H,n.W,n.D,n.C$, and n detonating height, width, depth, channel, and dimensional value, respectively. Similarly, we employed a run-time augmentation method to increase the generalizability and avoid overfitting; however, the original copies of both training and testing datasets were also retained. The dataset was passed for noise addition, flipping left-to-right and upside-down, as well as being rotated with an angle ($\angle$) in the range −180$°$-to-180$°$. For the augmentation, we passed Dt with order (L), resizing mode (K), and noise mode (M). The four functions sequentially call in the first loop over the dataset, including Resize($r-\mathit{Dt}$, L, K), Noise($r-\mathit{Dt}$, M, Clip), Rotate($r-\mathit{Dt},\angle$), and Flip($r-\mathit{Dt},{\mathit{mathfrakD}}_r$). The second outer loop (Lines 14 to 24) represents the contrast and trait selection and corresponding network for training and learning. The training process runs for a given number of epochs (Algorithm 1(Line-14)). The calculation of the number of iterations (iter) involves dividing $\mathit{rtextrm}-\mathit{Dt}$ by batch size ($B.\mathit{size}$), where $B.\mathit{size}$ depends on the GPU capacity available. A particular CT subject-wise dataset (SWD), as well as St and TrT selections, have been carried out in Lines 15–18. This pseudocode will be a generalized mechanism for choosing CT, St, and TrT in Lines 19, 20, and 21, respectively. A permutation in the form of $S.C\bowtie \mathit{Tr}\bowtie \mathit{St}$ is presented before calling the actual deep learning model. The proposed AGW-Net model runs on $S.C,\mathit{Tr}$, and St as passed parameters. Finally, the pipeline outputs the prediction and classification of traits and the visualization of mutually sensitive regions. In this research, we focus on age, weight, and gender estimation and visualization of vulnerable regions as explainable maps.

DL modeling for BTs estimation and association with brain MRI

The proposed DL model, namely AWG-Net, trains on single contrast MRI to learn variations in brain structure according to the given BTs. The proposed AWG-Net model associates BTs with structural brain MRI (Fig. 2). The detailed filter size, kernel size, and dimensions at the corresponding layers are outlined for ease of understanding in Fig. 2. The model’s overall structure is formulated as follows:

$$\mathit{AWG}-\mathit{Net}={\displaystyle \sum _{j=0}^n}{\mathfrak{R}}_j\left\{\right(P^{\to m}:[C-{\mathit{SCA}}^{m-a},{\mathit{MHA}}^{m-b}]),{\mathrm{\Psi }}^{m-c})\text{;}\ell \left({\omega }_{\lambda },{\beta }_{\lambda }\right)\}\text{.}$$ (1)

The model generates an estimate of a specific BT by repeatedly integrating the results of all the modules for n number. $\mathfrak{R}$ represents the structure of AWG-Net. The model architecture is composed of C-SCA, MHA, and auxiliary structure $\mathrm{\Psi }$. The symbol $\mathrm{\Psi }$ denotes all dense structures, 3D and 2D convolutions, FC, dropout layers, and the number of neurons for BTs estimation. Similarly, learning weights and biases are correspondingly denoted through ${\omega }_{\lambda }$ and ${\beta }_{\lambda }$ at a particular network level ($\lambda$) confined by the function $\ell$. The frequency of these modules is not consistent; thus, each module has been superscripted with frequency. The number of modules (P) ranged in $1\to m$, where each particular module existing counts by subtracting $a,b$, and c from the maximum number m representing the total modules. For instance, there are two C-SCA ($m-a$) modules where a single MHA ($m-b$) is employed, and variations as $m-c$ can be seen in the occurrence of $\mathrm{\Psi }$.

Customization and embedding spatial-channel attention

The C-SCA module captures sensitive regions of BTs from input brain MRI at earlier levels. The C-SCA blocks are embedded in two positions based on the model’s performance. The learning from the upper to deeper layers continues, however, with a decrease in dimensions (HWD) and an increase in feature maps (C). The structure of C-SCA is further formulated as follows:

$$\begin{array}{cc}C-\mathit{SCA}\hfill & ={\overbrace{\{n.\mathit{Dt}:(\mathit{GA}-P,\mathit{GM}-P)\to \oplus \mathit{AA}-P\otimes n.\mathit{Dt}\text{;}{\mathrm{\Psi }}_{P-1}\}}}^{P-1},\hfill \\ \hfill & {\overbrace{\{n.\mathit{Dt}\in R^{{\mathit{HWDC}}_{3d}}:\to \mathit{nn}.\mathit{Dt}\in R^{{\mathit{HWDC}}_{1d}}\text{;}\mathit{MAB}\}}}^{P-2},\hfill \\ P-1\hfill & =n.\mathit{Dt}:(\mathit{GA}-P,\mathit{GM}-P)\to \oplus \mathit{AA}-P\otimes n.\mathit{Dt}\text{;}{\mathrm{\Psi }}_{P-1},\hfill \\ P-2\hfill & =n.\mathit{Dt}\in R^{{\mathit{HWDC}}_{3d}}:\to \mathit{nn}.\mathit{Dt}\in R^{\frac{\mathit{HWDC}}{r}}\text{;}\mathit{MAB}\oplus \mathit{nn}.\mathit{Dt}\hfill \\ \hfill & \Rightarrow n.\mathit{Dt},\hfill \end{array}$$ (2)

The Phase-1 (P-1) of C-SCA receives 3D MRI as $n.\mathit{Dt}$ and forwards it through three parallel units: global average pooling (GA-P), global max pooling (GM-P), and adaptive average pooling (AA-P) to get features, compress them, and excite them (Fig. 2(b)). The structures of GA-P and GM-P are only discrepant in max and average pooling. The kernel size in GA-P and GMM-P varies from early to deeper layers for extracting appropriate BTs associated with vulnerable features. GA-P and GM-P are both used with different kernel initializers to keep the data’s variability across layers. They both use a single feature map to reduce the input n.Dt$\in R_{3d}^{\mathit{HWDC}}$ to nn.Dt$\in R_{1d}^{1\times 1\times 1\times C}$. Each unit is followed by FC and dropout sequentially to discard the less vital features prior to merging layers and capturing adjacent cross-channel local information. The merged form is passed through a sigmoid nonlinear function and elementwise multiplied with the original receiving input. Similarly, the AA-P unit receives n.Dt$\in R_{3d}^{\mathit{HWDC}}$ and flattens into nn.Dt$\in R_{1d}^{1\times 1\times 1\times C}$ using F.Conv layers using stride convolutions. AA-P first compresses along channel-wise, followed by 2D convolutions embedding with a dropout layer to reduce the learning complexities and excite the explicit learning weights along the channel. A 2D convolution is applied, passed through a sigmoid function, and elementwise multiplied back with the receiving input (n.Dt). The joint outcome of GA-P, GM-P, and the output of AA-P are elementwise summed and forwarded to phase 2 (P-2).

P-2 receives input as 3D structure ($n.\mathit{Dt}\in R^{{\mathit{HWDC}}_{3d}}$) and translates into $\mathit{nn}.\mathit{Dt}\in R^{\frac{\mathit{HWDC}}{r}}$ into the deeper layer using F.Conv operations. This structure continues into the bottleneck layer and back to its original form while building an encoder-decoder structure. The encoding process is carried out by compressing with a ratio (r) and decompressing to the same level as that of the direct connections. Each level is connected through a direct connection where each direct connection is passed through a max-pool and average-pool block (MAB) (Fig. 2(c)). The direct connection in parallel with F.Conv passes to the MAB block. The similar feature weights merge back to normal 3D maps with dimensions of $H,W,D,C$ for further and deeper level network learning. The normal and MAB outputs are fused and summed elementwise to produce output with focused regions in the brain. The receiving structures from the parallel unit may have different structural and viewing information, and therefore, optimal fusion and sharing take place when translating 2D the 3D representations [83]. All units capture distinct features and pass the cumulative interaction score to the next level.

Employment of multi-head self attention

The output from the C-SCA block is passed to a transformer block (MHA block) for learning further sensitive regions at deeper levels and included with dense structures. Besides the C-SCA block, MHA has been utilized at a deeper level to learn useful information variants among slices and facilitate BT estimation. The employment of the MHA block empowers the model to identify regions with the highest contribution to the prediction. The formulation of the MHA block is illustrated as follows:

$$\begin{array}{cc}\mathit{MHA}\hfill & =\mathit{PaE}\Rightarrow \mathit{TrB}\hfill \\ \mathit{PaE}\hfill & =[F_{\mathit{cls}},{\overbrace{D_1,D_2,..,D_n}}^{F_1},{\overbrace{D_1,D_2,..,D_n}}^{F_2},\dots ,{\overbrace{D_1,D_2,..,D_n}}^{F_n}].{\mathit{Pos}}_{\times F_n}\hfill \\ \mathit{TrB}\hfill & =\mathit{TrE}\left(\mathit{PaE}\right)\Rightarrow \mathit{LrN}\oplus \mathit{MHA}\hfill \\ \mathit{TrB}\hfill & =\mathit{ReD}[\mathit{LrN}+\mathit{MHA}]\in R_{\mathit{Ln}}^{\mathit{HWDC}}\hfill \end{array}$$ (3)

The self-multi-head attention (MHA) plays a significant role in capturing sensitive regions along the slices and their mutual relation in order to estimate BTs. Therefore, we employed MHA at the deeper layers after running a wide range of experiments for the potential number of slices. The MHA obtains the output by passing through a position and embedding (PaE) block followed by a transformer block (TrB). The multiple 3D structures ($F_1,F_2,\dots ,F_n$) are sliced, separated (sep), and labeled with class ID (cls). The sliced and positioned embedded structures pass to a transformer encoder (TrE) followed by layer normalization (LrN). Similarly, the MHA applies and concatenates back with LrN to the next level. Finally, the resize dimensions (ReD) block applies to bring back the attentive slices into the same dimensions as the corresponding layer level (Ln). The ultimate estimation takes place as follows:

$$\mathit{Est}=\mathit{TrB}(m.\mathit{Dt})\oplus {\mathrm{\Psi }}_{\mathit{Dense},3\mathit{DCon}.,\mathit{FC},\mathit{DO}},$$ (4)

The output of TrB and other structures ($\mathrm{\Psi }$) merges and produces the intended results. Finally, produced a robust and generalized DL model for BTs association with both healthy and neurological disordered subjects.

Dataset curation

The available datasets in previous studies on computer vision are inadequate for training deep learning models regarding CP and do not cover all four contrasts. The sample collection was carried out at Shenzhen Hospital with approval from the ethics committee, with Reference No. 202004105. The recruited patients’ age range spans from 1 month to 17 years, with a mean age of 2.98 years. The subjects contain both males and females with a minimum weight of 3 kg and a maximum of up to 30 kg. The data set was collected from 2013.1.1 to 2022.10.31. The guardians were notified with written consent to agree to conduct the study. All methods were performed per relevant guidelines and regulations. The inclusion and exclusion criteria have been described in Table 1. The dataset is composed of 716 children, with 327 being patients and 389 being controls (Table 2). Each subject has four contrasts: T1-w, T2-w, T1-sag, and Flair. Each contrast has three versions of datasets, including only controls (OC), only patients (OP), and joint forms of both controls and patients (P$\oplus$C). The dataset was obtained from MRI machines manufactured by Skyra, GE, and Philips, which were located in the stationed hospital. The echo-planar imaging parameters are as follows: TR  = 2000 ms, ET  = 30 ms, flip angle  = 90$°$, matrix size  = 64x64, 32 axial slices, field of view  = 24x24 cm², slice thickness  = 3 mm with no gap. The 3D-MPRAGE structure has the following settings: T1 relaxation time (TR) of 2300 ms, echo time (TE) of 2.26 ms, one average slice thickness of 1.0 mm, and a field of view (FOV) of 256 mm. The DICOM MRIs were converted to NifTi (.nii) format using MRIcron. As the collection has been made using different manufactured machines and producing variations in dimensions, the images are transformed to have the same dimensions, including depth (volume), using Scikit. Specifically, all MRI images were aligned to a size of 320x320x17x1, with the corresponding height (H), width (W), volume/depth (D), and channel (C). Augmentation is applied to dynamically modify the training splits, enhancing the model’s ability to generalize and preventing overfitting during training. The augmentation process involves randomly rotating an image within a range of 180$°$ and −180$°$, as well as adding various types of noise, including Gaussian, Localvar, S&P, Speckle, and Poisson noises. The order and mode for resizing, including nearest neighbor, bi-linear, bi-quadratic, bi-cubic, and bi-quartic, are randomly selected. The tensor placeholder (Tn) for AWG-Net (Eq. 1) is represented by five dimensions as follows:

$$\mathit{Tn}=T(B_n,H,W,D,C),$$ (5)

where $B_n,H,W,D$, and C denote batch size, height, width, depth, and number of channels, respectively. The collected dataset is the property of the Radiology Department of Shenzhen Children's Hospital and will be available upon request by the reviewers and the research community.

Table 1.

The inclusion and exclusion criteria for carrying out the study.

Subject	Criteria	Description
CP	Inclusion	1. CP diagnosis was confirmed using the international consensus criteria (i.e., a permanent, nonprogressive motor impairment resulting from a perturbation that occurred in the fetal or infant brain that may be associated with a range of comorbidities including, but not limited to, cognitive, visual, auditory or communicative impairments, along with feeding difficulties or epilepsy.
		2. Complete routine magnetic resonance examination.
	Exclusion	1. With history of other neurological disorders, traumatic brain injuries, and systemic illnesses.
		2. With metallic or motion artifacts and poor image quality.
Controls	Inclusion	1. Apgar score >= 8 at 1 min and 7 min after birth. Full term and of normal weight.
		2. Complete routine magnetic resonance examination and structural MRI showed no abnormality.
	Exclusion	1. With history of perinatal asphyxia, intrauterine distress, or any neurological disease, traumatic brain injuries, or systemic illnesses.
		2. With metallic or motion artifacts and poor image quality.

Table 2.

Training and testing splits.

Dataset	Health controls	Patients	Health controls & Patients	T1-w	T2-w	Flair	Sag
Training Split	330	325	615	✓	✓	✓	✓
Testing Split	59	42	101	✓	✓	✓	✓
Total	389	327	716	-	-	-	-

MRI slice distributions

Among the four contrast (scans), T1-w, T2-w, and Flair are taken from the axial view (Figure S1 (a)), where the Sag is captured from a sagittal view (Figure S1(b)). The axial view can demonstrate different brain tissues depending on the slice depth, where the middle-depth slices are annotated for explanation in Figure S1(a) and Sag slices shown in Figure S1(b). The regions covered by a slice for a subject are significantly important to machine learning.

Axial and sagittal view based slices distributions

Different ranges of slices cover brain regions, some of which may be deeply associated with CP, where BTs include but are not limited to age, weight, and gender. The detailed slice distributions in terms of covering regions are listed both from the axial (Supplementary Table S1) and Sag view (Supplementary Table S2). Four columns corresponding to each group outline the underlying brain structure in the slice range. The axial view scans are obtained from inferior to mid and superior, while the Sag contrasts are captured from left to middle and then to the right side of subjects.

Models training

After considerable training experiments of DL models on the four MRI contrasts, AWG-Net was reported with robust results and stood as the most appropriate DL model train on T1-w. Therefore, AWG-Net(Fig. 2) and the underlying state-of-the-art (SOTA) [77], [84], [85], [86], [87], [88], [89], [90] models are trained and evaluated on the same training and testing splits. Adam optimizes the network weights with the initial learning rate of ${10}^{-5}$, and ${\beta }_1=0.9$ and ${\beta }_2=0.999$ for 1000 epochs, where after every 50 epochs, the learning rate drops by 10 decimals. The list of network parameters is illustrated in Table 3 The softmax and categorical cross-entropy functions compute the prediction probability and classification accuracy for CP patients and controls. The training and testing splits contain 615 and 101 subjects, totaling 716 (Table 2). The testing split contains 42 patients and 59 controls. The trained models were evaluated via both quantities and qualitative-based metrics. For the quantitative metrics, we employed MAE, MSE, MCS_i, confusion matrix, positive predictive value (PPV), negative predictive value (NPV), F₁, accuracy, linear regression, boxplot, violin plot, and joint plot. In subjective evaluation, we outline the affected regions vulnerable to BTs in healthy children and those suffering from neurological disorders (CP) useful in clinical practices.

Table 3.

Network parameters analysis.

Parameter	Values	Description
Learning rate	$1\times {10}^{-5}$	Critical for determining the step size at each iteration towards the minimum loss
Momentum	${\beta }_1=0.9,{\beta }_2=0.999$	Accelerating gradients vectors for faster converging
Learning decay factor	$1\times {10}^{-1}$	Adjusting learning rate over time for better performance
Epoch	1000	The number of times the entire dataset is passed forward and backward through the neural network
Batch size	4	Number of training sample used in a single iteration
Optimizer	Adam	Optimization algorithm
Learnable Parameters	1.645197 M	Network weights and biases learn during training
Objective function	Softmax and categorical cross-entropy	Quantification the difference between the predicted and the actual values in the training data

It is critical to measure and report the time spent on specific phases within the experiment. Time measurement for an experiment is critical regarding replicating the experimental and reproduction and provides insights into the optimization of the experiment. Thus, incorporating time spent as part of the experimental results is an essential aspect of thorough scientific reporting. Notably, there are four dataset versions (T1-w, T2-w, Flair, and Sag) on which the underlying models are trained (Table 2). However, considering the appropriateness of T1-w, the proposed AWG-Net and SOTA models are trained on T1-w sequences (Table 4). The time spent on the underlying experiment can be divided into 1) data collection, 2) data curation, and 3) running experiments and evaluating results.

• •The data collection was carried out from 2013.1.1 to 2022.10.31.
• •Two to three months were taken to organise the data for training and evaluation purposes.
• •Each model was run for 1000 epochs. We had four GPUs where four deep-learning models could run in parallel. The batch size was 4. Every 100 epochs was run for around 2 h, considering the 615 MRIs training split. Therefore, around 40 h are taken for a total of 1000 epochs to ensure the optimal training parameters and hyperparameters learning.

Table 4.

The underlying SOTA and proposed AWG-Net networks are trained on the same parameters and evaluated through mean absolute error (MAE), mean square error (MSE), mean cumulative score (MCS₂₋₁₀) [92], and correlation coefficient (R₂) metrics. The proposed model outperformed counterparts.

Scan	MAE	MSE	MCS2	MCS3	MCS4	MCS5	MCS6	MCS7	MCS8	MCS9	MCS10	R2
Model-1 [86]	2.01	8.78	48.01	56.93	63.86	68.89	72.70	75.68	78.10	80.09	81.86	0.46
Model-2 [84]	2.50	11.57	37.78	48.01	55.54	61.30	65.91	69.67	72.82	75.39	77.49	0.29
Model-3 [87]	2.06	8.82	29.78	39.34	47.21	54.64	60.53	65.53	69.03	72.13	74.66	0.54
Model-4 [93]	2.20	9.68	31.51	42.20	50.69	57.34	62.80	66.95	70.29	73.06	75.42	0.21
Model-5 [89]	2.06	7.73	33.06	44.46	53.44	60.10	65.10	69.26	72.49	75.08	77.27	0.49
Model-6 [90]	2.59	13.93	25.74	35.64	44.75	51.98	57.70	62.19	65.19	69.05	71.73	0.21
Model-7 [85]	2.06	7.73	33.06	44.46	53.44	60.10	65.10	69.26	72.49	75.08	77.27	0.29
Model-8 [88]	2.16	7.10	29.23	40.16	49.34	56.55	62.41	65.45	70.76	73.68	76.08	0.43
Model-9 [91]	2.00	8.69	35.47	45.28	54.05	60.64	65.62	69.55	72.71	75.34	77.49	0.45
Model-10 [77]	2.12	10.79	38.03	48.23	56.00	61.56	66.05	69.55	72.54	74.94	76.89	0.26
AWG-Net (Proposed)	1.07	3.69	69.64	75.89	80.21	82.86	85.35	86.97	88.23	89.30	90.25	0.75

Results

It is not straightforward to compare the accuracy of our AWG-Net method to the existing SOTA models for the estimation of BTs because (1) the existing methods are trained on either of the MRI contrasts such as T1-w, T2-w, sagittal, or Flair, (2) all the existing datasets may not contain all BTs such as age, gender, and weights together all possible scans. However, for the baseline study and comparison with SOTA models, we train a few existing 3D CNN models on our collected datasets for both estimation of age, weight, and classification of gender. We employed both healthy and CP patients to train three times for each healthy, control, and joint form controls and CP through four MRI contrasts. The proposed AWG-Net and the underlying SOTA models [77], [84], [85], [86], [87], [88], [89], [90], [91] are trained and evaluated on the same training and testing splits as tabulated in Table 4. Training parameters and hyper-parameters for the SOTA models are employed according to their original studies. The underlying SOTA models received four contrasts, where T1-w was reported substantially for BT estimation; consequently, we also trained all the selected SOTA models on the same MRI contrast. The AWG-Net outperformed the counterparts with higher prediction accuracy after running for fivefold cross-validations. The demonstrated results validate the proposed models’ efficiency and emphasize using T1-w MRI for BT (Age) estimation. The low MAE and MSE scores indicate better performance, whereas the high scores of MCS₂, MCS₃, and R₂ show better model performance. The experimental results of AWG-Net against the SOTA models are compared and depicted in Fig. 3. The plotting was based on the MCS score. From the depiction, it can be seen that AWG-Net scores for MCS₂₋₁₀ are higher than competitors’ models.

Fig. 3.

Experimental plot of the comparison of the proposed model against the SOTA models in terms of MCS scores.

Age estimation

Quantitative evaluation

We trained AWG-Net for age estimation over the four contrasts with three possible subject-wise datasets such as only controls (OC), only patients (OP), and both controls and patients (BCP) as illustrated in Table 5. The results are evaluated via MAE, MSE, MCS₂₋₁₀ [92], boxplot, and linear regression as quantitative scores and qualitative evaluation is carried out in Section 3.1.2. Overall, the age prediction accuracy of T1-w (Avg. MAE: 1.17) is higher than Sag (Avg. MAE: 1.233), Flair (Avg. MAE: 1.79), and T2-w (Avg. MAE: 2.03). Similarly, control subjects’ brain age prediction accuracy is higher than that of CP patients, with a small margin, reflecting a minor influence on the regions vulnerable to neurological disorders and associated with brain age development. Notably, in Table 5, the prediction accuracy of controls is higher than the counterparts CP patients (T1-w: C = 1.05, P = 1.16; Sag: C = 1.24, P = 1.26; Flair: C = 1.16, P = 2.37; T2-w: C = 1.29, P = 2.37). Similarly, we also utilized a mix of controls and patients to develop and elaborate the joint training effort, which was found with decreasing inaccuracies (T1-w: P$\oplus$C = 1.11; Sag: P$\oplus$C = 1.23; Flair: P$\oplus$C = 1.79; T2-w: P$\oplus$C = 2.03).

Table 5.

Quantitative performance measurements AWG-Net. Three dataset versions included only controls (OC) and only patients (OP), and both controls and patients P$\oplus$Controls are utilized to estimate age. T1-w, T2-w, Flair, and Sag MRI sequences. Lower MAE, MSE, MCS₂₋₁₀ and higher R₂ show result significance (bold).

Scan	Subject	MAE	MSE	MCS2	MCS3	MCS4	MCS5	MCS6	MCS7	MCS8	MCS9	MCS10	R2
	OC	1.05	2.49	51.97	61.55	68.60	73.65	77.41	80.24	82.43	84.19	85.63	0.844
T1-w	OP	1.16	4.43	53.33	63.63	70.18	74.54	77.66	80.00	81.82	83.45	84.79	0.717
	P$\oplus$C	1.14	3.89	50.58	60.66	67.97	72.84	76.52	79.28	81.42	83.14	84.55	0.747
	Avg.	1.117	3.603	51.96	61.947	68.917	73.677	77.197	79.84	81.89	83.593	84.99	0.769
	OC	1.24	3.17	46.59	56.98	64.94	70.43	74.50	77.68	80.16	82.15	83.77	0.80
Sag	OP	1.26	5.28	51.51	61.81	68.00	72.72	76.10	78.63	80.60	82.36	83.80	0.66
	P$\oplus$C	1.20	3.88	47.78	58.91	66.15	71.21	75.02	77.97	80.26	82.09	83.09	0.74
	Avg.	1.233	4.11	48.627	59.233	66.363	71.453	75.207	78.093	80.34	82.2	83.553	0.733
	OC	1.16	2.93	47.31	58.33	66.02	71.50	75.42	78.36	80.64	82.58	84.16	0.81
Flair	OP	2.37	16.05	41.81	50.45	56.36	60.90	64.67	67.95	70.70	73.27	75.37	0.516
	P$\oplus$C	1.85	9.01	40.79	51.22	58.46	63.75	68.13	71.67	74.51	76.92	78.89	0.41
	Avg.	1.793	9.33	43.303	53.333	60.28	65.383	69.407	72.66	75.283	77.59	79.473	0.337
	OC	1.29	3.61	45.87	55.91	63.65	69.17	73.42	76.74	79.33	81.39	83.08	0.77
T2-w	OP	2.37	16.00	41.81	50.45	56.36	60.90	64.67	67.95	70.70	73.27	75.37	0.489
	P$\oplus$C	2.45	13.51	31.46	41.78	49.51	55.71	60.51	64.51	67.90	70.76	73.29	0.25
	Avg.	2.037	11.04	39.713	49.38	56.507	61.927	66.2	69.733	72.643	75.14	77.247	0.334

In order to provide a more comprehensive explanation using quantitative evaluation of the age prediction and mutual relationship between controls and patients, we employed a variety of plots to show the performance of AWG-Net (Fig. 4). The performance evaluation was carried out on the three dataset versions. The drawing has been made for T1-w contrast in column-1 (Fig. 4)Column-1), T2-w (Column-2), Flair (Column-3), and Sag (Column-4). The first row of Fig. 4 plots the results as a boxplot for BCP corresponding to the four contrasts (T1-w, T2-w, Flair, to Sag). The second row plots the results for control subjects, whereas the third row draws the patient subjects only. There are variations in spreading in the boxplot for the four contrasts (Fig. 4First row)). The drawn boxplot represents the network’s internal structure as mean weights. The boxplot for the T1-w scan has narrow distributions, little distinction in sizes, and closely matches median lines between the actual and predicted ages. The statistical analysis reveals the significance of training the AWG-Net model on the T1-w contrast in relation to the concepts of median lines, overlapping, and distribution prediction metrics. It can be revealed that the T1-w contrast has the potential to be utilized for age estimation in BCP as the benchmark until discovery. In comparison, the statistical results for T2-w, Flair, and Sag scans are less significant when estimating the age of children.

Fig. 4.

The presented figure illustrates the comparative analysis of ages obtained from T1-w brain MRI scans, specifically focusing on the correlation between actual ages and predicted ages. First row: from left-to-right, depicts boxplot between actual and predicted age for the contrasts T1-w, T2-w, Flair, and Sag represented in Column 1 -to- 4. The depiction is drawn for BCP. Second row: this row shows the linear regression plot for the health control subjects. Third row: this outlines the linear regression plot for CP patients. The line along the diagonal shows significant outcomes, while skewed-away lines show poor results.

The second and third rows have a vital role in figuring out the appropriate scan and subject associations (the first and second rows of Fig. 4). The second row represents control subjects; therefore, the linear regression lines are closely along the diagonal for all the scans, especially for the T1-w contrast. However, the regression lines for CP patients are found to be a bit diverged from the diagonal position. Yet, the age prediction results are substantial for T1-w and Sag compared to T2-w and Flair. This comparison implies that in children suffering from neurological disorders, such as CP, the brain developmental-related regions are influenced by the palsy disorder. The visualization of such regions is shown in Fig. 5. .

Fig. 5.

The figure depicts the AWG-Net model’s learned trajectories against controls and patients across different age groups. The input MRI slices are shown on the top. The subjects are categorized into eight groups, from 0–1 year to 15–18 years. Similarly, the upper row represents the early layers (a) of the model, where the middle row (b and c) represents the middle layers, and so on for the deeper layer. Each group has been appended with an MAE score at the bottom.

Qualitative evaluation

The learned features from the trained network structure on brain MRIs are observed to be significant, representing distinguishable trends for patients and controls. The visuals may provide logical explanations for clinical practitioners and radiologists. The incorporation of attention mechanisms has an impact on the network with distinction in weights. Through the attention-embedded DL strategy, the important areas are highlighted in the brain, and their flow has been shown spreading from higher-level slices to deeper layers. All the contrast combinations with significant results are taken by default from an axial view (Figure S1). Such a viewing position aims to depict the key regions vulnerable to CP and is useful for CP’s association with BTs. To qualitatively evaluate the results, we acquired the learned knowledge from AWG-Net and visualized it to clinically understand the vulnerable regions associated with aging slicewise (Fig. 5). Grouping has been made based on age while training the model; however, while evaluating and understanding the results, we arranged the OC and OP into eight groups (Fig. 5). The subjects fall in the category of 0-to-1, 1-to-3, 3-to-5, 5-to-8, 8-to-11, 11-to-13, 13-to-15, and 15-to-18 years. AWG-Net is composed of deep learning layers, and therefore, we selected four layers, from upper-to-deeper layer, for acquiring the learned weights (Fig. 5(a-d)). Each group is illustrated with an MAE score at the bottom. For the health controls, the model accurately predicts the age from the brain MRI in the early age (0.48 years) and poor results for the later year (15–18 years). Similar result patterns can be observed in the patients in terms of MAE score, as the model is composed of attention mechanisms that play a key role in capturing sensitive regions regarding BTs (specifically for age).

The visual trends are the average values of the slices along the depth dimension obtained from a particular layer, proceeded by the attention mechanism. The depiction in Fig. 5 conveys details regarding the difference in learning patterns between health controls (OC) and CP patients (OP), from early age (0–1 year) to later age (15–18 year) while learning the network weights as vulnerable regions from upper to deeper layers. Overall, for the OC, the deeper layers emphasize capturing consistent features from hard and soft tissues. The upper model layers mostly focus on the bony tissues and external skull structure, including the frontal bone. However, the deeper layers are sensitive to the cerebellar hemisphere in control subjects. Similarly, the model also learns the external skull as an associated sensitive area; however, it found inconsistency across the categorywise age range and layerwise (i.e., along the depth slices of brain MRIs) trends in patients. The skull bone includes the frontal bone, occipital bone, parietal bone, and soft tissue such as the cerebellum. Notably, the sensitive regions for control subjects are consistent while inconsistent for patients across the age progression.

Weight

Quantitative evaluation

In addition to age as BTs, weight estimation is also the key direction in this study. As a benchmark, AWG-Net has also been trained on the MRI slices regarding extrapolating human weight association to brain MRI of healthy controls and neurological disordered children, specifically CP. From the quantitative statistics (Table 6), weight estimation is also found significant in the training results through T1-w MRI compared to the counterpart contrasts. The significant MAE (2.85), MSE (15.69), MCS₂ (26.91), and R₂ (0.82) were reported for weight estimation through T1-w MRI sequence. The patient subjects were reported to have lower MAE and MSE, higher MCS, and R₂ scores; thus, it represents that brain region information is found to be more variant in patients than in healthy subjects. Moreover, poor results in terms of MAE (6.31), MSE (85.36), and MCS₂₋₁₀ (18.18) are reported for T2-w contrast, which implies that T1-w is more appropriate for studying weight as BT from brain MRI. The evaluation parameters have significant scores for patients underlying T1-w and Sag, while there is a minor difference in the case of Flair and T2-w scans. Overall, a higher correlation (R₂) has been reported between the brain MRIs of HC and CP subjects and weight (see last column of Table 6).

Table 6.

Tabulation of statistical results AWG-Net for training on brain MRI labeled with weight as BT. The model was trained on four MRI contrasts, including T1-w, Sag, Flair, and T2-w. The results are illustrated in terms of MAE, MSE, MCS₂₋₁₀, and R₂ scores. The subject-wise categories contain health controls only (C), patients only (P), and health controls and patients (P$\oplus$C) followed by cumulative score. The lower the MAE and MSE scores and the higher the MCS₂₋₁₀ and R₂ values, the more the results are significant. Overall, T1-w shows significant results, specifically for patients, whereas T2-w shows poor results.

Scan	Subject	MAE	MSE	MCS2	MCS3	MCS4	MCS5	MCS6	MCS7	MCS8	MCS9	MCS10	R2
	C	4.65	39.80	17.92	23.65	29.24	34.94	40.24	44.75	48.86	52.79	56.30	0.78
T1-w	P	2.85	15.69	26.06	36.81	45.45	51.81	56.88	61.13	64.64	67.63	70.24	0.82
	P$\oplus$C	4.40	40.05	22.14	28.32	34.68	40.32	45.15	49.56	53.22	56.36	59.31	0.76
	Avg.	3.967	31.847	22.04	29.593	36.457	42.357	47.423	51.813	55.573	58.927	61.95	—
	C	5.56	66.11	18.63	25.26	30.96	36.02	40.24	44.35	47.67	50.75	53.56	0.63
Sag	P	4.01	32.70	24.24	31.36	36.36	41.21	46.23	50.68	54.34	57.45	60.33	0.62
	P$\oplus$C	5.12	54.14	18.18	24.47	31.18	36.71	41.15	45.10	48.64	52.02	55.24	0.68
	Avg.	4.923	50.983	20.35	27.03	32.833	37.98	42.54	46.71	50.217	53.407	56.377	—
	C	5.47	59.17	17.92	23.11	28.81	33.69	37.94	42.20	45.87	49.13	52.29	0.67
Flair	P	5.75	59.41	18.78	23.18	23.18	28.36	33.03	37.40	41.36	44.64	47.45	0.50
	P$\oplus$C	5.93	72.54	14.45	20.45	26.57	31.70	36.46	40.82	44.67	48.18	51.36	0.57
	Avg.	5.717	63.707	17.05	22.247	26.187	31.25	35.81	40.14	43.967	47.317	50.367	—
	C	5.53	56.37	14.33	19.89	25.59	31.00	35.79	40.18	43.18	47.31	50.73	0.69
T2-w	P	5.75	59.41	18.78	23.18	28.36	33.03	37.40	41.36	44.64	47.45	50.08	0.32
	P$\oplus$C	6.31	85.36	18.18	23.77	28.95	33.68	38.16	41.69	44.91	47.76	50.34	0.49
	Avg.	5.863	67.047	17.097	22.28	27.633	32.57	37.117	41.077	44.243	47.507	50.383	—

After the statistical result illustrations (Table 6), we outlined the proposed model performances in terms of probabilities density function for T1-w, T2-w, Flair, and Sag in Fig. 6, where the first to fourth columns show the corresponding contrast. Similarly, the violin plots depict the joint form of controls and patients in the first row, the regression line for health controls in the second row, and the third row presents the actual and predicted weights of CP patients. In the first row, among the four contrasts, T1-w contrast plots between actual and predicted weights were found with close similarities in terms of center, spread, and distribution (Fig. 6First row)). Specifically, the median points and population distribution across the medians, spreading of the box, interquartile range (IQR), and density (frequency and width) have a resemblance to the actual and predicted weights. However, major discrepancies have been observed in density, first quartile, maximum, and medians for T2-w and Flair. However, the spreading and medians are resembled with minor discrepancies for the Sag contrast. Similarly, the regression plot drawn for the health controls is spread less uniformly along the diagonal line, showing poor results (Fig. 6Second row)). However, the regression plots for the patient subjects show significant results between the actual and predicted weights, especially for the T1-w contrast. Such comparison implies that the model learns more vulnerable regions or variations in regions associated with weight in CP patients compared to health controls. However, the results for the rest of the three contrasts are not that significant, representing widely scattered points around the regression line.

Fig. 6.

Weight estimation and depiction using brain MRI of healthy and CP children. These plots aim to highlight weight distributions of brain structures for normal and neurologically disordered subjects. First row plots violin plots for both health controls and patients. Second row plots regression plot for health controls. Third row plots estimation between actual and predicted weights of CP patients. The first to fourth columns outline MRI contrasts correspondingly for T1-w, T2-w, Flair, and Sag.

Qualitative evaluation

The visualization of vulnerable brain regions from the network learned weights as saliency or explanation maps on children’s brains is crucial for studying and examining neurological disorders. The depiction has been made in Fig. 7. On top, the input MRI slices received to the DL model. The following visuals in Fig. 7(a) are captured from the model network structure and grouped according to OC, OP, and BCP. Each group has been depicted with three layers of network depth. Similarly, the corresponding histograms are shown under each group (Fig. 7(b)). In Fig. 7(c), the merged form of the three histograms is shown for comparison purposes. In health controls, the DL model focuses both on hard and soft tissues while learning about weight association with brain regions. These tissues contain the frontal, occipital, temporal, and cerebellar hemispheres. The CP patient subjects have almost found similar tissue relevancy except for the occipital and temporal bones. Interestingly, the joint dataset version of patients and controls has similarities in the frontal bone and discrepancies in the parietal lobe. The variations in frontal, temporal, and cerebellum areas are associated with subjects’ weight as previously articulated [30]. Moreover, the weight prediction accuracy in terms of MAE score is significant (lower) compared to controls presented via histograms (Fig. 7(b-c)). However, both control and patient subjects with ages above 13 years are observed to have fewer weight variations, which leads to poor weight prediction results (higher MAE score).

Fig. 7.

The above figure presents the sensitive regions to weight in children’s brains and their association with healthy and CP patient subjects. The age ranges for the corresponding age category are 0-to-1, 1-to-3, 3-to-5, 5-to-8, 8-to-11, 11-to-13, 13-to-15, and 15-to-18 years. On top, the input MRI slices to the DL model are presented. The subjects are categorized into 8 groups based on their age affiliation. (a) A list of visuals is categorized for health controls, CP patients, and joint versions of controls and patients (BCP). Each group has been depicted with three layers of depth corresponding to the DL model structure. (b) The corresponding histograms are shown under each group of visuals. (c) The joint form of the three histograms is shown for comparison purposes, where A to H denotes groups 1 to 8, respectively.

Gender

Quantitative evaluation

Similar to Age and weight as BTs, we also trained AWG-Net on the brain MRIs labeled with gender information. For gender classification, there are four types of MRI contrast where each contrast has three dataset versions, including BCP, OP, and OC, as illustrated in Table 7. Overall, gender prediction has satisfactory classification accuracy in terms of confusion metrics; however, it can be improved by incorporating further intelligence. Interestingly, the evaluation results elaborate that T1-w also plays a key role in the identification of BT as gender and relevant vulnerable regions. Besides, Sag contrast also found a key source of connecting brain regions in children. In the training on the isolated patients and control datasets, the classification accuracy is higher for patient subjects, implying that patients have a more clear reflection of gender-related variation on their brain visuals compared to their counterparts’ health controls. The Flair scan is found to have poor performance regarding gender classification.

Table 7.

Gender classification from the four contrasts using AWG-Net model. Among, T1-w shows significance in performance compared to counterparts modalities.

Scan	Subject	TP	FP	FN	TN	Specificity	Sensitivity	PPV	NPV	F1	Accuracy
	P$\oplus$C	81	17	8	24	0.653	0.717	0.827	0.80	0.6847	81.88
T1-w	P	29	2	7	12	0.881	0.704	0.946	0.649	0.783	82.00
	C	55	15	3	15	0.533	0.795	0.816	0.889	0.638	82.98
	P$\oplus$C	71	11	21	35	0.761	0.67	0.866	0.625	0.713	76.81
Sag	P	25	6	5	14	0.700	0.641	0.806	0.737	0.669	78.00
	C	41	10	14	23	0.697	0.641	0.804	0.622	0.668	72.73
	P$\oplus$C	61	21	22	34	0.62	0.64	0.74	0.61	0.63	68.84
T2-w	P	20	10	7	12	0.55	0.625	0.67	0.63	0.58	65.31
	C	39	12	12	25	0.68	0.61	0.76	0.68	0.64	72.73
	P$\oplus$C	68	14	34	22	0.61	0.76	0.83	0.39	0.68	65.23
Flair	P	21	9	7	12	0.57	0.64	0.70	0.63	0.60	67.35
	C	144	9	69	42	0.82	0.77	0.94	0.38	0.80	70.45

To outline the sample distributions of males, females, and joint form, we plot boxplot for health controls, patients, and both controls and patients from the trained result on T1-w (Fig. 8). Overall, the male and female subjects are found distinct from smaller to large number variations in terms of median lines overlapping, distributions, width, and outliers. For T1-w (Fig. 8First column)), the median lines, outliers, width, and distribution reveals the significance to use a particular contrast (T1-w). However, less overlapping is observed, which indicates minor discrepancies between male and females of the control group (OC). However, the male-to-female differences for the patient group (OP) under T1-w contrast are higher (Fig. 8(Second column)), indicating that gender-vulnerable regions on the brain would also be effected with CP disorder. The joint of controls and patient (BOC) with respect to male, female, and male&female also show variations in their boxplots (Fig. 8Third column)). The plotting for the rest of the contrasts is excluded from the drawing here because of the poor performance. More details regarding gender variations along the age progression are depicted in the Supplementary section.

Fig. 8.

Gender classification evaluation for T1-w contrast for the three versions of datasets, including controls, patients, and joint of control and patients. From left to right, the boxplot depicts controls, patients, and joint form of controls and patients (controls$\&$patients), respectively. Similarly, each boxplot contains results corresponding to male, female, and their joint dataset version.

Qualitative evaluation

It is of high importance to visualize the DL model’s learned weights, specifically from attention mechanism, against male and female in the group of health controls (Fig. 9(a)), CP patients (Fig. 9(b)), and Patients-Controls (Fig. 9(c)). There are cumulatively sixteen slices from T1-w brain MRI with heatmap regions representing sensitive regions to gender in each group. The red squares show major differences, whereas the blue ones show minor differences between the corresponding slices of male and female. As the DL model is embedded with attention mechanisms, particular regions are focused on learning gender as BT. The common focused regions include frontal bone, occipital bone, temporal bone, basal ganglia, and dorsal thalamus for the three groups, including OC, OP, and BCP. However, the health controls are focused on the parietal lob and internal capsule. Moreover, from the slice-by-slice comparison, it is challenging to dig out variations in the bony structure, but in terms of brain parenchyma, slice 8 (Fig. 9(b)) has a red forehead. The area of the lobe between (a) and (b) for slice 8, occipital lobe, and basal ganglia seems to be larger in OP compared to OC. Similarly, in the 9th slice between (a) and (b), the frontal lobe and occipital lobe of the group (b) seem to be larger, whereas in the 10th slice, the frontal lobes and occipital lobes of (b) are smaller than those of (a). On the 14th and 15th slices, the parietal lobes of (b) seem to be fuller than those of (a), but on the 16th floor, the parietal lobes of group b are smaller than (a). It also has high significance in analyzing male-to-female variations in each group. The DL model focuses on the frontal bone of female subjects deeper along the slices compared to male subjects in OC. Similarly, the female subjects have consistency in the gyrus regions in deeper layers and focus early on the internal capsule compared to males in OC. The major differences for male to female subjects of OP are found in slices 7 to 10, with variations observed in the occipital lobe, basal ganglia, and parietal lobes. The model can be equipped with more smart attention to capture detailed sensitive regions in order to capture gender differences in both OC and OP.

Fig. 9.

The model learned features regarding gender separately in health controls and patients, and both health and patients are visualized. The matching pairs inside the red rectangle show major differences; however, the blue color shows minor differences. There are three groups of dataset versions for which the depiction has been drawn. (a) Visualization for health controls having males and females in the first and second row. (b) Depiction of patients suffering from CP where (c) draw the learned weights for the DL model trained on the joint form of controls and CP patients.

To further evaluate the network’s learned weights as visual maps and explanations, we grouped the OC and OP based on their age into seven categories: Group-A (0-to-1), -B (1-to-3), -C (3-to-5), -D (5-to-8), -E (8-to-11), -F (11-to-14), and -G (14-to-18) years (Fig. 10). Each OC and OP has male and female visual representations. The visuals are the accumulative weights from the network structure at deeper layers. This analysis aims to elaborate on the variation of brain-sensitive regions in male and female subjects in both OC and OP. The visual trend variations have been observed in both intra and inter-group subjects. The intra-subject as males (Fig. 10Column 1–3)) and females (Fig. 10Column 4–6)) show hot region spreading with age progression in both OC and OP. A few of the samples from male subjects from both groups are pointed out using red color rectangles. The highlighted areas reveal that different brain regions activate gender and subjects with different brain mapping in their neurological appearance from their brain MRI. To further elaborate along the age progression, the males and females in the early age group are found with significant accuracy (lower MAE); however, later age groups are found to have higher MAE scores retaining smooth inclination (Fig. 10(a)). However, inconsistency in male and female subjects is found in terms of MAE scores along the aging. It reveals signs of influencing the CP vulnerable regions more compared to the health controls.

Fig. 10.

The figure draws the visuals obtained from the network internal structure for Males and Females in (a) health controls and (b) CP patients. The subjects are grouped with respect to their ages, starting from Group A (0-to-1), -B (1-to-3), -C (3-to-5), -D (5-to-8), -E (8-to-11), -F (11-to-13), -G (13-to-18) years. Each row depicts the accumulative scores for the corresponding group, and their graph plots are shown below. The bar graphs shown in (b) represent Males, Females, and Male&Female in terms of accumulative mean values along the y-axis. The x-axis ranges the agewise groups from A to G.

Ablation study

We ran a number of DL models with the inclusion and exclusion of various components and selected AWG-Net as the optimal one. The training has been carried out on the four MRI contrasts, including T1-w, T2-w, Sag, and Flair, by adding the underlying BTs as labeled information. Among all the DL models, AWG-Net has been found significant for all the underlying contrasts utilizing the given BTs. However, on the four dataset versions as four brain MRI contrasts, AWG-Net showed significant results for T1-w contrast. Therefore, AWG-Net has been selected as the benchmark model to compare with the underlying SOTA counterparts (Table 4).

During the experimental process, the model passes through a number of developmental phases with the inclusion and exclusion of modules as attention plays a key role. AWG-Net is equipped with two types of attention, namely C-SCA and MHA. The model was first run on brain MRI with the absence of C-SCA and MHA attentions; however, it was not found significant for the estimation and association of BTs. Later on, C-SCA was included at different locations and occurrences, among which the embedding at the early level, after the first and second layers, received significant outcomes. As the number of channels increased in the network depth, MHA was experimented with, which was found to be a useful outcome. However, the isolated version, which included MHA, was found to have poor results for estimation and classification.

In the attention mechanism, both types of attention mechanisms have been modified from their original versions following the needs of the BTs’ estimation and classification. The original SCA was based on squeeze and excitation along the spatial and channel dimensions; however, it is inappropriate for the existing scenarios to have a four-dimensional tensor. Therefore, we proposed two phases based on customized SCA (Fig. 2(b)). Similarly, in the original application of MAH in [77], we customized based on feeding of single view positioned MRI simultaneously (Fig. 2(c)). The statistics of such an ablation study have been illustrated in Table 8 where the statistical illustrations for the rest of modalities, inclusion, and exclusion of components are tabulated in a supplementary section (Table S3). Among the numerous experiments, T1-w contrast with the given structure (Fig. 2) outperformed the counterparts models as having different network structures and thus proved to be the optimal DL model regarding the current scenario.

Table 8.

Selection of an appropriate contrast, C–CSA level, and MHA position based on their performance towards an ultimate DL model. The performances are evaluated using MAE, MSE, and R₂ scores. The details of the model are shown in the supplementary section.

Contrast				C–CSA Level			MHA Position			MAE	MSE	R2
T1-w	T2-w	Flair	Sag	L0	L1	L2	L0	L1	L2
✓	X	X	X	✓	X	X	✓	X	X	1.14	3.89	0.747
✓	X	X	X	X	✓	X	✓	X	X	1.86	7.39	0.531
✓	X	X	X	X	X	✓	✓	X	X	1.93	8.52	0.652
✓	X	X	X	✓	X	X	X	✓	X	2.11	9.41	0.413
✓	X	X	X	X	✓	X	X	✓	X	1.89	8.39	0.617
✓	X	X	X	X	X	✓	X	✓	X	2.01	8.86	0.445
✓	X	X	X	✓	X	X	X	X	✓	2.16	9.44	0.401
✓	X	X	X	X	✓	X	X	X	✓	2.01	9.66	0.394
✓	X	X	X	X	X	✓	X	X	✓	2.21	9.89	0.396
X	✓	X	X	-	-	-	-	-	-	-	-	-

Discussion

This study aims to extrapolate the association of BTs, as BTs are associated with children’s brains using recent technologies. Age, weight, and gender are included as preliminary studies in the study of BTs; however, it can be extended to more types of BTs; however, it can be extended to more types of BTs. Furthermore, the brain visuals as brain MRI that are taken in this study are specifically studied for children with healthy brains and those suffering from neurological disorders. A wide range of studies have employed DL models for BTs’ association with brain MRI but not particularly with CP. Moreover, this study not only aims to associate the brain MRI of healthy and CP patients using DL models but also attempts to employ customized attention mechanisms to reduce the range of underlying modalities into a single MRI and specify a particular contrast. This study laid the foundation as a benchmark study in terms of the DL model with C-SCA and MHA attentions (Fig. 2), BTs, neurological disorder, dataset modality, contrast and view selections (Table 5). Interestingly, this study not only estimates BTs but also describes explanation maps and visuals for clinical practices. The estimation and classification of CP, association with BTs, and appropriate contrast selection at early stages is challenging owing to the factors of gross motor functions and higher potential of recovery from neural disorders in children that are challenging to investigate [60], [11]. It is demanded not only to estimate associations but also to extrapolate the affected and vulnerable regions that are of high importance to radiomics. There are limited approaches available to the joint study of neurological disorder, BTs, brain structural MRI as contrasts and single modality, and visualizing their associations as explanation maps from brain visuals using DL models.

The results obtained by this study emphasize the selection of T1-w with the underlying DL model; however, there needs to be better results for the rest of the modalities. The poor results may imply that future studies with intelligent attention embedding can further improve results not only on T1-w but also on the rest of the modalities. In the literature study, few of the studies show significant results based on T2-w [63], Sag [64], and Flair [55]. However, these contrast models produced poor results, embarking on future work with more robust attention and training more DL models on these contrasts. The poor results on these contrasts may point to the extrapolation that T2-w [63], Sag [64], and Flair [55] have little association with such neurological disorder or brain structural representations. Still, it may lead to the conclusion that the proposed model is the appropriate choice for T1-w. However, future studies can be carried out on the rest of the brain’s visual modalities.

The clinical practices are limited by the time required to acquire multiple contrasts. Therefore, the selection of the most informative MR contrasts is pivotal to avoiding lengthy acquisitions, lowering the possibility of motion artifacts, and ensuring cost-effectiveness. All the desired contrasts may not be available because of the limited scan time, suboptimal signal-to-noise ratio, and image artifacts [61]. In addition, an available contrast may limit robust machine training by having sufficient training examples, leading to poor performance in clinical practices. Therefore, we dynamically add blurriness, noise, and artifacts to provide a generalized DL that can be learned from both cleaned and poor-quality versions. Moreover, every child’s brain has a unique composition [54], and such composition may be variant in different structural MRIs, which can be challenging to capture through naked-eye examination in clinical practices. Therefore, future studies must be included in further research on the poor performance contrasts.

The incorporation of C-SCA into AWG-Net is the appropriate choice, considering the sensitivity at the early layers and passing to deeper layers for further learning. At the deeper layer, the number of channels increases, where the fed MRI contains slices along the fourth dimension. Therefore, MHA at the deeper level separates the channels for each slice with different class numbers and learns the sensitivity for each class. Therefore, embedding MHA is the best option, and it can be extrapolated for further improvements or reduce the attention to an optimal level. Besides, this study innovatively elaborates on the deep association between healthy and CP children, which has a vital role in clinical practices. Clinicians and radiologists can assist in determining the correlation between control and patient subjects. The age estimation accuracy in terms of MAE and MSE is significant for control subjects and poor for CP patients in most of the cases while using AWG-Net (Table 5). Interestingly, CP patients are found to be more deeply associated with brain visuals in the case of weight estimation (Table 6). Similarly, the patient subjects have high accuracy when classifying gender as a trait.

Conclusion

We proposed AWG-Net in this study. It is a deep learning model comprising specialised SCA and MHA blocks and a cumulative dense structure that helps send useful data to deeper layers. The three main investigations conducted in this study are: 1) the analysis of model trains for BTs, taking into account traits such as age, weight, and gender; 2) the examination of four contrasts, namely T1-w, T2-w, Sag, and Flair; and 3) the investigation of different dataset versions, including only controls (OC), only patients (OP), and a combined dataset of controls and patients (BCP). Three BTs are considered to be investigated in this study, including age, weight, and gender, based on the optimal MRI contrast chosen. In a list of single contrast MRIs, T1-w contrast was found to be best suitable for age, weight, and gender estimation (Table 5, Table 6, Table 7). However, T2-w was observed with poor performance for age and weight estimation (Table 5, Table 6), where Flair for gender classification (Table 7). To distinguish health controls and CP patients in terms of BTs, we generated three data splits, each containing OC, OP, and BCP. Overall, the accuracy for OC subjects in terms of age (T1-w: MAE = 1.05, MCS₁₀=85.63, R₂=0.844), gender (T1-w: Accuracy = 82.98%), and weight (T1-w: MAE = 2.85, MCS₁₀=70.24, R₂=0.82) compared to their counterparts. This deduced that age and gender are less affected by the brain regions associated with motor function (CP) compared to weight, which has a bigger influence on the brain regions. To further elaborate on the findings, we aim to visualise the vulnerable regions captured by AWG-Net as network weights from specialised attention modules. From the visual depictions, age-associated vulnerable regions are associated with the frontal bone, cerebellar hemisphere, frontal bone, occipital bone, and parietal bone for health controls (OC), where these areas are inconsistent for CP (Fig. 5). Notably, the signs and symptoms of CP usually appear in the early months, where the average diagnosis age is around 2 years; therefore, including subjects with early age has a vital role in studying brain development in neurological disordered children. Similarly, for weight, the OC and OP share similarities in focusing on the frontal bone and cerebellar hemisphere regions where discrepancies exist in the occipital bone and temporal bone (Fig. 7). Moreover, for gender and slice-wise analysis, the main difference between OC and OP is observed in the occipital lobe and basal ganglia (slice 8), frontal lobes and occipital lobes (slice 10), and parietal lobes (slices 14-to-16) (Fig. 9). For males and females, the DL model focuses on the frontal bone of female subjects deeper along the slices compared to male subjects in OC. One big difference, though, is that the occipital lobe, basal ganglia, and parietal lobes (7 to 10 slices) are different for OP (Fig. 9, Fig. 10). In addition to providing quantitative estimates, this research establishes the groundwork for visually illustrating the critical and susceptible sites associated with BTs, which may be valuable for therapeutic applications.

Limitations and future directions

The proposed DL architecture contains numerous parameters and has a complex structure because it deals with 3D MRI. Similarly, the model takes many epochs to converge to a good accuracy rate because a small learning rate should be chosen while training, which can be improved in the future direction. The accuracy of weight and gender estimation can be improved by delving into more MRI scan selection. It is urged that the association between CP as a neurological disorder and other BTs, such as genetic and behavioral traits, be investigated for future research. Further investigation is needed to fuse different contrast MRIs to utilize symmetric and asymmetric features to estimate BTs efficiently. In addition, different neurological disorders can be included to elaborate on optimal MRI contrast selection and trait association. Investigating this unexplored direction would help radiologists, clinicians, and therapists during early intervention, diagnosis, and treatment. Moreover, more specialized multi-head self-attention mechanisms can be incorporated to achieve cost-effective, robust, and reliable network architectures with enhanced outcomes

Ethical permission

The sample collection was carried out at Shenzhen Children’s Hospital with approval from the corresponding committee, with Reference No. 202004105. The recruited patients’ age range lies from 1 month to 17 years, with a mean age of 2.98 years. The guardian was notified with written consent to agree to conduct the study. All methods were performed per relevant guidelines and regulations.

Declaration of Competing Interest

The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.

Supplementary material

The following are the Supplementary data to this article: