MGB-Unet: An Improved Multiscale Unet with Bottleneck Transformer for Myositis Segmentation from Ultrasound Images

Abstract

Myositis is the inflammation of the muscles that can arise from various sources with diverse symptoms and require different treatments. For treatment to achieve optimal results, it is essential to obtain an accurate diagnosis promptly. This paper presents a new supervised segmentation architecture that can efficiently perform precise segmentation and classification of myositis from ultrasound images with few computational resources. The architecture of our model includes a unique encoder-decoder structure that integrates the Bottleneck Transformer (BOT) with a newly developed Residual block named Multi-Conv Ghost switchable bottleneck Residual Block (MCG_RB). This block effectively captures and analyzes ultrasound image input inside the encoder segment at several resolutions. Moreover, the BOT module is a transformer-style attention module designed to bridge the feature gap between the encoding and decoding stages. Furthermore, multi-level features are retrieved using the MCG-RB module, which combines multi-convolution with ghost switchable residual connections of convolutions for both the encoding and decoding stages. The suggested method attains state-of-the-art performance on a benchmark set of myositis ultrasound images across all parameters, including accuracy, precision, recall, dice coefficient, and Jaccard index. Despite its limited training data, the suggested approach demonstrates remarkable generalizability by yielding exceptional results. The proposed model showed a substantial enhancement in accuracy when compared to segmentation state-of-the-art methods such as Unet++, DeepLabV3, and the Duck-Net. The dice coefficient and Jaccard index obtained improvements of up to 3%, 6%, and 7%, respectively, surpassing the other methods.

Introduction

Inflammatory myopathies (i.e., also referred to as myopathies) are chronic illnesses primarily characterized by muscle inflammation and muscle weakness [1]. Nevertheless, myositis is widely adopted as a shorthand to designate the usual origins of inflammatory muscle diseases or idiopathic inflammatory myopathies. Muscle inflammation can arise from multiple sources, including infections, the negative consequences of drugs or pharmaceuticals, physical trauma, cancer, and, especially, autoimmune disease. After an episode of inflammation, muscle fibers experience degradation. This affects the functionality of muscle tissue, leading to fatigue and restricted mobility [2, 3]. Myositis encloses various forms, including polymyositis (PM), dermatomyositis (DM), and inclusion body myositis (IBM) [4]. These categories exhibit distinct aetiologies, manifestations, and therapeutic approaches. Each of them shows muscular weakness. PM is an autoimmune disorder characterized by the immune system’s assault on the muscle fibers, resulting in weakening and discomfort [5]. On the other hand, DM can impact the respiratory system, cardiovascular system, and gastrointestinal system. IBM is an uncommon and progressive disorder characterized by muscular weakening and atrophy. IBM impacts the musculature of the wrists, fingers, thighs, and feet [6].

The diagnosis is dependent on the clinical examination, which evaluates the pattern of muscle weakness, together with laboratory testing, including measurement of creatine kinase (CK) levels and autoantibodies, performance of electromyography (EMG), and evaluation of the histology of the skeletal muscle [7]. Magnetic resonance imaging (MRI) of the skeletal muscle is beneficial for finding an appropriate muscle for biopsy and for visualizing the extent of muscle damage that may not be apparent by clinical observation [8]. While disorders like DM, PM, and NM usually have a favorable response to immunosuppressant therapy, IBM frequently has poor responsiveness to these drugs. Misdiagnosis can result in treatments that are ineffective or hazardous [9]. IBM often exhibits resistance to immunosuppressant medication, a treatment that can be advantageous for PM and DM but may worsen IBM symptoms and raise one’s vulnerability to infections [10]. In addition, IBM is occasionally misidentified as alternative types of myositis or muscle illnesses, such as polymyositis, dermatomyositis, or muscular dystrophy. This might lead to inappropriate or ineffective interventions that may worsen the condition or cause adverse outcomes [11].

Medical image segmentation is one of the most critical steps in analyzing medical images. Before considering the region of interest, it is necessary to eliminate the background and other structures in the image. This process provides more precise information about the area of interest which can improve the accuracy of medical image analysis [12]. By reducing the overall complexity of the image and drawing attention to the specific area of interest, segmentation improves the visualization of medical imaging. Medical image segmentation can reduce the amount of data needed for processing, which speeds up the analysis process [13]. Moreover, by providing more detailed information about the target area, segmentation can improve the accuracy of diagnoses. Medical image segmentation can use various methods, including region expansion, edge detection, thresholding, and machine learning-based approaches [14].

Despite the impressive results achieved by deep learning approaches in medical image segmentation, they do encounter several obstacles. Deep learning models may not perform optimally when faced with medical images due to low quality, noise, artifacts, or the absence of labels. Medical images often exhibit regions or things that are notably larger or more abundant than others [12].

In recent years, there has been increased research on the importance of early detection for the automated diagnosis of IIMs. Automatic detection helps with early treatment, which improves health and wellness by reducing regressions and halting the progression of irreversible damage [15]. In patients with myositis, segmentation can assist in identifying and measuring the degree of muscle inflammation, injury, or atrophy [16]. Moreover, this can support tracking the course of the disease, how well it responds to therapy, and the outlook. Differentiating between myositis types, such as DM, PM and IBM, can be aided by segmentation. Each kind of myositis has different patterns of muscle involvement, which can be identified by segmentation [17]. Segmentation can aid in locating and ruling out additional sources of anomalies in the muscles, such as tumors, infections, or artefacts that might resemble or exacerbate myositis. This can lessen the chance of incorrect diagnoses or pointless treatments. In this research, we provide a model for automated segmentation and classification of inflammatory myopathies from ultrasound images. The following are the main highlights of this paper:

• We provide a new approach, MGB-Unet, for segmenting several stages of myositis from ultrasound images using an enhanced UNet architecture. Our system incorporates a novel residual block called the Multi-Conv Ghost switchable bottleneck Residual Block (MCG-RB) module, which enables multi-level feature extraction.

• The Integration Bottleneck transformer is utilized in conjunction with the MCG-RB module to minimize the semantic disparity between the encoder and decoder features of the Unet Architecture.

• We extensively validate the proposed model using ultrasound images that include normal, polymyositis (PM), dermatomyositis (DM), and inclusion body myositis (IBM) muscles. In addition, we evaluate the effectiveness of the model by comparing it to other state-of-the-art networks such as Unet, DeepLabV3, and Ducknet.

Related Work

Medical image segmentation is one of the most crucial phases of medical image analysis. Before analyzing the region of interest, it is necessary to eliminate the backdrop and other structures in the image. Segmentation can improve medical image analysis precision by providing more detailed information about the area of interest. Segmentation enhances medical imaging visualization by highlighting the precise region of interest and lowering the overall complexity of the picture. Furthermore, segmentation can improve diagnostic accuracy by providing specific information about the target area.

A lot of studies have been done in the field of myositis disease prediction and early detection because of how important it is. Several machine learning methods have been used for various picture visualization procedures; these methods have been used to develop computer-aided disease diagnostic methodologies. Because of its ability to simultaneously show a detailed picture of soft tissue disease in a large quantity of muscle, magnetic resonance imaging (MRI) was one of the main reasons it was used. MRI simply shows morphological abnormalities in terms of size, kind, and location.

In their study, Nagawa et al. [18] used an open-source program called ITK-SNAP version 3.8.0 to extract the ROI of the muscle from the MRI image. Various features were derived from the ROI using the help of techniques such as the gray-level run-length matrix (GLRLM), the gray-level size zone matrix (GLSZM), the neighboring gray-tone difference matrix (NGTDM), and the gray-level co-occurrence matrix (GLCM). After that, three traditional classification methods, k-nearest neighbors (k-NN), random forest (RF), support vector machine (SVM), and multi-layer perceptron (MLP), were used to group these attributes. The hardships of extracting features and the complexity of choosing the right features from the extracted ones caused the model to be too complex to be employed in real clinics.

To divide ROI regions using V-Net, Wang et al. [19] suggested a deep learning model they dubbed Deep Anatomical Federated Network (DAFNE). In order to distinguish between myositis and facioscapulo-humeral dystrophy (FSHD1), Fabry et al. [20] described using a Lipschitz deep learning model in combination with whole-body magnetic resonance imaging (MRI).

In spite of the fact that MRI has proven it can produce high-resolution images, the method is still quite expensive and is only used occasionally. Recent advances in ultrasonic (US) imaging, however, have made it possible to depict muscle morphology with more accuracy by making soft tissue characteristics more clearly visible [21]. Muscle morphology visualization has also been enhanced by this. The fact that it is cheap, simple, easy to use, safe, and does not involve ionizing radiation is one of the US’s advantages [11].

Nodera et al. [22] used conventional methods, like Random forests, to distinguish between different types of myositis illnesses by extracting characteristics from ultrasound images. Using both traditional and deep learning techniques, Burlina et al. [23] built a publicly available ultrasound dataset and assessed the accuracy of the classifications. Three separate binary class scenarios were generated by Burlina and his colleagues based on the findings of the model evaluation. The ability of the models to differentiate between healthy and unhealthy individuals was the primary metric for evaluation in the first case. In the second scenario, the models were tested to see how well they could differentiate between IBM patients and those who were healthy. As for the third scenario, we put the models through their paces to see how well they could distinguish between IBM patients and PM and DM patients. Because IBM is making modest but steady progress, this was done. This leads to a high rate of incorrect diagnoses, such as diabetes and PM in individuals with IBM. There were a lot of methods proposed for myositis categorization, but the supplied model was flawed when it came to automatic segmentation. The region of interest was segmented manually by skilled medical personnel, despite the great job done by Burlina et al. [23].

Furthermore, the offered methods just utilized three distinct ways for binary classification, distinguishing between healthy and unhealthy patients. In addition, Uccar et al. [24] used VGG16 and VGG19 to build a deep-learning model for myositis classification. In order to evaluate the model’s performance and reliability, the three scenarios previously mentioned in [23] were used. The model also differentiated between several types of myositis disorders and healthy individuals in a multiclass classification trial. Moreover, a method called Generative Adversarial Networks (GAN) augmentation was proposed by Tan et al. [25] to improve the performance of the standard VGG16 network for classifying myositis.

Marzola et al. [26] dug further into a technique that uses the grey level of the cross-sectional area to diagnose dysfunctional muscles. The approach begins by utilizing a cluster of convolutional neural networks to divide the cross-sectional area (CSA). After that, we find out if the muscle is unhealthy or abnormal by calculating the average grey level z-score of the segmented area and comparing it to the z-score that measures the level of muscle health. Although this research approach has improved upon other studies, there is still room for optimization in the segmentation of muscle cross-sectional area, since the accuracy of the classification result of aberrant muscle is approximately 91.5%.

In [27], the paper proposes an automated approach for analyzing transverse musculoskeletal ultrasound images. The method combines multi-task learning (MTL) with attention mechanisms to achieve two primary tasks: muscle cross-sectional area (CSA) segmentation: accurate delineation of muscle boundaries. Abnormal muscle classification: identifying neuromuscular diseases (e.g., myasthenia gravis, myotonic dystrophy). The proposed MMA-Net (Multi-Task Model with Multi-Scale Fusion and Attention) effectively integrates information from different scales and enhances feature extraction. Experimental results demonstrate superior performance in both classification and segmentation tasks.

Moreover, in [28], the study aims to analyze muscle MRI scans in patients with neuromuscular diseases (NMDs). The authors employed deep learning approaches (including U-Net 2D, U-Net 3D, TransUNet, and HRNet) for muscle segmentation. The focus is on muscle cross-sectional area (CSA) segmentation. The study evaluates the impact of fatty infiltration on segmentation accuracy. HRNet achieves the best performance, recognizing every muscle in the database. However, accuracy decreases significantly for the most infiltrated muscles (>20% fat fraction).

Myositis is a heterogeneous group of diseases that cause inflammation and damage of the skeletal muscles. Segmentation of myositis lesions from ultrasound images can help to assess the severity and progression of the disease, as well as the response to treatment. However, there is no existing model that can perform automatic segmentation of myositis disease from MRI data. This is a challenging task due to the variability of the lesions, the low contrast between muscle and inflammation, and the lack of annotated data. Therefore, developing a robust and accurate model for myositis segmentation is an important and open research problem.

Proposed MGB-Unet Architecture

In this section, we introduce our proposed segmentation MGB-Unet Architecture. The proposed model is shown in Fig. 1. The proposed methodology employed the UNet as the basis for the suggested network architecture for myositis segmentation. By merging their features, we incorporate the BOT module, serving as a transformer-style attention module that narrows the gap between the encoding and decoding stages. In addition, the Multi-Conv Ghost Switchable Residual block (MCG-RB) module is utilized to extract multi-level features during both the encoding and decoding stages. This module combines multi-convolution with ghost switchable residual connections of convolutions.

Fig. 1.

Overview of the proposed MGB-Unet for myositis segmentation approach from ultrasound images

During the encoding phase, the network is fed with input ultrasound images. The encoding stage comprises four layers: the fourth layer only extracts the features, while the previous three layers extract features and downsample them. The MCG-RB module is employed in the encoding path to extract comprehensive feature representations for the myositis ultrasound pictures. It has the ability to extract more profound characteristics while preserving finer details compared to conventional convolution.

In each layer, the resulting encoder feature map keeps the number of channels the same. Afterward, the feature map is down-sampled using a convolution with a step size of 2, doubling the number of channels called DownConv. In the decoding phase, the feature map is upsampled by transposed convolution with a step size of two in the decoder stage, resulting in half as many channels as the original by UpConv. Subsequently, the input is incorporated into the feature map of the relevant encoder level, which the convolutional block has processed. The multiconv block later performs feature extraction, resulting in the generation of a decoder feature map. Ultimately, the number of feature channels is reduced from 64 to 2, and then the classified class is computed using the softmax activation layer.

BottleNeck Transformer (BOT Module)

In order to bridge the semantic gap between the features from the encoding and decoding stages of the same resolution level, the BOT module is used prior to performing decoding. This is because the decoding features differ significantly from the encoding features after multiple sampling and convolution operations.

The features from the BOT module of the same level, the features from the encoding levels, and the features from the prior decoding layers make up the three components that make up the input to each decoding layer. The proposed MCG-RB module in each decoding layer first concatenates these characteristics before decoding them. The design specifications for the MCG-RB and BOT encoder and decoder will be covered in detail below.

The BOT module shown in Fig. 2 was embedded in the skip connection between the encoder and the decoder. An essential component of the suggested hybrid network is that it has multi-head self-attention (MHSA) [29]. The self-attention architecture can process and aggregate the feature map data in order to support the CNN in managing long-distance dependencies. Self-attention is a machine-learning technique that assigns weights to different parts of an input sequence based on relevance. In the transformer model, self-attention generates parallel context-aware representations for each token. Multi-head attention extends the concept of self-attention by allowing the model to learn multiple ways to attend to different parts of the input. This involves multiple heads focusing on distinct aspects of the sequence, capturing diverse patterns and dependencies. Components of Multi-Head Self-Attention: The attention layer takes three input parameters: Query (Q), Key (K), and Value (V).

Fig. 2.

a BottleNeck Transformer, b MHSA Attention Mechanism [29]

As shown in Fig. 2, Q represents the current token and its context, K represents other tokens in the sequence, and V provides information about each token. Multi-head self-attention splits these parameters into multiple heads (N heads), allowing for parallel computations. The attention scores for each head are computed by taking the Q and K dot product, followed by a softmax function to obtain normalized weights. Weights are then multiplied with V to get the attended representation. Finally, the outputs from all heads are combined to form the final multi-head attention representation.

When working with highly detailed images as myositis dataset, the self-attention in MHSA can help the network better understand the relationships between different regions and improve the segmentation accuracy. Furthermore, the MHSA shown in Fig. 2 with enough heads is at least as expressive as any convolutional layer. The MHSA enhances the deep model’s performance in learning the most detailed features by creating multiple attention maps and embedding features from an image to encode rich information. The BoT block, which gains its advantages from the MHSA, can assist the network in improving segmentation performance.

Multi-Conv Ghost Switchable Bottleneck Residual Block (MCG-RB) Module

The Ghost-Block is also utilized in the proposed MCG_RB Module as shown in Fig. 3. This architecture produces representative features with a low computational cost.

Fig. 3.

Multi-Conv Ghost switchable bottleneck Residual Block (MCG-RB) module

Myositis images have variety of fine details because of the tissues and fats of the ultrasound image. The traditional Unet segmentation algorithm used two successive 3x3 convolution blocks causing it to be unable to obtain the features of tissue details of the muscle. To improve the segmentation accuracy, the proposed block used three different convolutions with different kernel sizes to acquire multiscale feature information maps and integrate them.

As demonstrated in Fig. 3, the ghost residual bottleneck is utilized with switchable normalization (GBS). The GBS extracts indicative feature maps at low computational performance. It outperforms other effective techniques because linear operations on feature maps have substantially lower computational complexity than conventional convolution. The Ghost Block expands the features and channels by using linear operations that are more affordable after first generating intrinsic feature maps using conventional convolution. In order to ensure stability and expedite the training process, it is recommended to include a normalization layer. The batch size limits the efficiency of Ghost-Block-BN due to the instability and reduced accuracy of batch normalization (BN) at small batch sizes. Instead, the suggested model utilizes switchable normalization (SN) [30], which effectively adapts to different batch sizes. SN computes layer-wise, minibatch-wise statistics, channel-wise, and layer and instance normalization in order to recognize their crucial weights to determine their ideal integration, guaranteeing network accuracy and performance stability in the case of small batch sizes.

To minimize the weight of the suggested model, the computed parameters are reduced by employing a Depthwise-Convolution (DwConv) block instead of the conventional one [31]. DwConv not only reduces the computational complexity of the model but also mitigates overfitting. The process of depthwise convolution involves two distinct steps: first, a pointwise convolution is used to create a linear combination of the output from the depthwise convolution. Then, the depthwise convolution applies a single convolutional filter to each input channel. Nevertheless, the standard convolution method carries out both channelwise and spatial-wise computations simultaneously, resulting in excessive parameters.

Materials and Evaluation Measures

Dataset and Preprocessing

The proposed model was tested using the muscle ultrasound dataset presented by Burlina et al. [23]. This dataset has a total of 3214 ultrasonography images from 80 patients, including 19 with IBM, 14 with PM, 14 with DM, and 33 with normal conditions. Seven muscle groups were scanned on both sides of each patient: the gastrocnemius, flexor digitorum profundus, flexor carpi radialis, deltoids, rectus femoris, and tibialis anterior. For every muscle except the rectus, with a depth setting of 6 cm, the depth was set at 4 cm. To consider changes in echo intensity caused by minor adjustments to the probe’s position, adjust the muscle by acquiring up to three independently acquired B-mode pictures of the cross-sectional view of every muscle. A few patients recorded just two images for many muscles. This ultimately resulted in the capture of 3214 muscle images. When imaging at 4 cm and 6 cm depths, the generated input images had pixel resolutions (width$\times$height) of 476$\times$499 and 318$\times$499.

The Johns Hopkins Myositis Clinic in Baltimore, Maryland used a GE Logiq E machine to take the ultrasound images. In addition, the images are meticulously organized and labeled according to the standards set by the European Neuromuscular Centre (ENMC). Furthermore, the dataset contained output masks for the significant surrounding muscles and adipose tissues, carefully separated by medical experts, as depicted in Fig. 4 [32].

Fig. 4.

Example from the muscle ultrasound dataset with mask representation a PM, b DM, c IBM, and d Normal

The proposed model cannot only segment the area of interest with high accuracy, but also it can be used to classify the segmented area, whether it is diagnosed as PM, DM, IDM, or normal with no myositis diseases. The following preprocessing steps are made before data enters the proposed model. Firstly, medical masks and ultrasound images were resized to 256$\times$256. Then, transform the mask into a tensor encoded in one-hot format. In this case, there are four disease classes and an extra one for the background class, so five classes in total. Thus, the mask tensor will contain five values, each belonging to each class. One-hot encoding [33] is beneficial for medical picture masks since it can reduce the calculation and evaluation of the masks by decreasing the number of dimensions and the intricacy of the data. Moreover, it can also improve the efficiency and precision of the models by eliminating the uncertainty and prejudice associated with the ordinal or numerical encoding of the classes. Additionally, it can enhance the visualization and perception of the masks by utilizing diverse colors or symbols for each class.

The model could generalize better after adding more data to the training set. Dropout and other regularization approaches are rendered obsolete. Enhancements are put into place via the albumenations library. The method uses a random approach to modify training images, leading to noticeable changes. This aids the model’s ability to generalize to fresh data. Before each epoch, training input augmentations were applied at random. The techniques used are CLAHE filter [34], flipping, color jitter effect, and affine transformation. Each ultrasound image is flipped horizontally and vertically. Moreover, a random color jitter effect is employed by randomly choosing brightness, hue, and saturation values. Affine transformation is used to rotate and scale the ultrasound image. The dataset before augmentation contained 3214 images only; however, after augmentation, the dataset contained 8926 images, as demonstrated in the following Table 1.

Table 1.

The table presents the dataset size utilized to evaluate our proposed model with and without augmentation for each class

Disease	Without augmentation	With augmentation
Normal	1313	2179
IBM	794	2070
DM	814	1837
PM	553	2840

Experiments and Results

Evaluation Metrics

Thorough evaluations utilizing specific measures like the dice coefficient (DC), the Jaccard index (JI), Precision (Pre), Recall, and Accuracy (Acc) are required to ascertain the efficacy of various neural network architectures. These five indicators serve this function admirably. The DC, which is also known as the F1_score, is a way to demonstrate how similar two sets are to one another. The value is zero if there is no overlap and one if there is. By comparing the sizes of the intersection and union of two sets, the JI finds the degree to which the two sets overlap. A harmonic mean of recall and Pre, the F1_score has a version called the F2_score. For situations where recall is more important than Pre, the F2_score is advantageous since it assigns more weight than Pre. On medical diagnosis, the primary focus may be on accurately detecting all patients with a disease rather than prioritizing avoiding false alarms.

Ablation Experiments

The first step in evaluating the proposed architecture is to test each contribution stage independently, as shown in Table 2. It presents the results of each step separately when evaluated using the augmented dataset. The six evaluation measures show that the proposed architecture outperformed the traditional UNet. Precisely, the BOT module improves Pre by 10%, while the MCG_RB module enhances the JI by 12%. Moreover, the ablation investigation shows that the BOT and MCG_RB modules work well together to improve segmentation performance. Both the Pre and Recall are significantly enhanced by 21%. The approaches with bold numbers are considered the best ones, as determined by computing the p values of paired t-tests on the DC. At the 5% ($\mathit{p}<0.05$) significance level, the values were found to be substantially different.

Table 2.

Ablation experiments results for our presented model. The highest results are highlighted in bold

Model	Acc	Pre	Recall	Specificity	DC	JI	P value (Dice)
UNet	0.809	0.752	0.759	0.759	0.755	0.672	$3.3e^{-}11$
UNET + BOT	0.874	0.859	0.852	0.861	0.855	0.726	$7.6e^{-}17$
UNET + GhostBottleNeck	0.867	0.841	0.849	0.852	0.844	0.693	$2.9e^{-}15$
UNET+ MCG_RB	0.886	0.867	0.862	0.869	0.865	0.793	$3.6e^{-}17$
Proposed model	0.932	0.916	0.914	0.919	0.914	0.851	$1.1e^{-}19$

In addition, experiments have been done on changing the number of MHSA heads, as shown in Table 3. Networks can perform better when the number of heads in multihead self-attention increases. The computational cost and overfitting danger, though, might go up if this happens. Taking multihead self-attention and reducing the number of heads can simplify the network and use less memory. However, it might worsen some jobs, particularly ones that need careful attention to the input sequence. Multiple different numbers of heads are applied, seeking the best results. When there are four heads, our network performs at its best. The experiments stopped at four heads only, so the complexity of the model would increase dramatically. After calculating the p values of paired t-tests on the DC and JI, the strategies with bold numbers are deemed the best. The values were found to be considerably different at the 5% ($\mathit{p}<0.05$.), according to the statistics results.

Table 3.

Ablation experiments according to number of MHSA heads of our presented model. The highest results are highlighted in bold

Number of MHSA Heads	Acc	Pre	Recall	Specificity	DC	JI
1	0.901	0.894	0.903	0.899	0.898	0.836
2	0.918	0.908	0.912	0.907	0.911	0.842
3	0.926	0.915	0.911	0.918	0.912	0.846
4	0.932	0.916	0.914	0.919	0.914	0.851

Experimental Results

The effectiveness of the proposed strategy was evaluated by comparing the model’s output with alternative methods documented in the existing literature. The dataset was evaluated using Unet, Unet++, DeepLabv3, and DUCKNet, as indicated in Table 4. The results were outstanding regarding the DC and JI, demonstrating an improvement of 6% and 7%, respectively. UNet++ employs deep supervision, which applies the network’s loss function to several intermediate outputs. This improves the network’s capacity for learning and lessens the gradient vanishing problem. Nevertheless, it has more layers and parameters than U-Net, and it needs more memory and processing power. Complex and heterogeneous medical images, including those with overlapping structures, noise, low contrast, or artifacts, can be too much to handle. Additionally, it might be unable to identify the connections and semantic linkages between various classes or regions in the images, which can be crucial for making medical diagnoses. Even though Deeplabv3 is used for semantic image segmentation in medical applications such as tumor, organ, or lesion detection, complex and heterogeneous medical images. It can be exhausting for the system to handle images that include overlapping structures, noise, low contrast, or artifacts. Moreover, it might not be able to depict the dependencies and semantic links between various classes or regions in the images, which can be important for segmenting myositis medical ultrasound images due to its complex structure. Our proposed model with bold numbers is thought to be the best ones, as shown by the p values of paired t-tests on the JI and the DC, respectively. Statistically, the proposed approach performed better than the rest. In addition, the findings demonstrated that, at the 5% significance level ($\mathit{p}<0.05$), values were substantially distinct.

Table 4.

Comparison between the proposed model and the previous state of arts. The highest results are highlighted in bold

Model	Acc	Pre	Recall	Specificity	DC	JI	F2_score	P value
UNet [35]	0.809	0.752	0.759	0.759	0.755	0.672	0.757	$3.3e^{-}11$
UNET++ [36]	0.821	0.795	0.792	0.805	0.792	0.743	0.792	$4.8e^{-}14$
DeepLabV3 [37]	0.857	0.824	0.815	0.836	0.813	0.762	0.816	$3.6e^{-}16$
DuckNet [38]	0.899	0.863	0.865	0.879	0.861	0.792	0.864	$2.7e^{-}17$
Proposed model	0.932	0.916	0.914	0.919	0.914	0.851	0.914	$1.1e^{-}19$

Figure 5 demonstrates the outputs of each model. UNET, UNET++, and DeepLabV3 failed to detect the suitable class of the IBM because it is the most challenging class in features yet the most important class to be accurately classified. DuckNet uses a proprietary convolutional block that combines numerous tiny convolutions with variable dilation rates to capture and process picture information at multiple resolutions. Thus, this increases the segmentation Acc while preserving the images’ contextual and geographical information. DuckNet successfully detected the suitable class of IBM through the mask. However, it has a very complex architecture with a high computational cost. Moreover, the JI was only 86%. On the other hand, UNet, UNet++, and Deeplab v3 could not detect the suitable class of IBM because of the complex structure of the IBM disease shown in the ultrasound image, which even some specialists diagnose wrongly. Nevertheless, the proposed model demonstrated its efficacy in accurately diagnosing the respective classifications of each disease.

Fig. 5.

Ultrasound image with mask extracted of each model

Figure 6 shows the validation and training loss in each epoch; the two curves overlapping together along the graph show no signs of overfitting. While Fig. 7 demonstrates the Pre and recall graphs of the proposed model along the 200 epochs. Table 5 displays the classification results of each disease. The Acc of the IBM is the lowest due to its complex structure, and it is not easy to diagnose even with specialists. However, the proposed model achieved 90% in F1_score and F2_scores, showing its efficiency in detecting even the most challenging disease. According to the results of paired t-tests on the DC and JI, respectively, the p values were all below 0.001. Moreover, the test showed that the results were found to be significantly different at the level of 5% ($\mathit{p}<0.05$).

Fig. 6.

Train and val loss graph

Fig. 7.

Precision and Recall graph

Table 5.

Classification results for each class separately

Myosistis disease	Acc	Pre	Recall	F1_Score	F2_score
DM	0.922	0.914	0.916	0.915	0.915
PM	0.915	0.913	0.915	0.913	0.914
IBM	0.906	0.905	0.901	0.903	0.901
Normal	0.922	0.928	0.928	0.928	0.928
Average	0.916	0.915	0.915	0.915	0.914

Conclusion

Idiopathic inflammatory myopathies (IIM), such as DM, PM, and IBM, provide a risk of severe and sudden muscle inflammation that often affects organs outside the muscles and leads to significant disability. Therefore, it is becoming more and more important to diagnose and treat these disorders early on. There is no other disease quite like the idiopathic inflammatory myopathies. A critical reduction in the efficacy of much-needed care could result from an incorrect diagnosis, leading to the administration of inappropriate medicine. Segmentation is an essential initial step in assessing diagnostic pictures. Medical image analysis becomes more accurate with segmentation since it allows for more specific data collection on the area of interest. Segmentation can also improve medical image visualization by reducing the amount of data that needs to be processed, which allows for a more explicit focus on the area of interest. Also, medical picture analysis is made much faster with segmentation. As a result, we suggest a better segmentation method. In addition to significantly reducing computational cost, the proposed method achieves the best outcomes according to the evaluation metrics. Our model was compared to several architectures in terms of accuracy, specificity, precision, dice coefficient, and Jaccard index. Compared to the most famous traditional segmentation techniques, the suggested model achieved an average of 92% in most metrics and 85% in Jaccard index scores. Although the proposed model achieved the highest performance, yet it is noticeable that IBM diseases cause the accuracy of the model to drop, it is due to the complexity of the disease. IBM is a progressive disease, so it usually misdiagnosed as PM or DM in the beginning. Also, quantifying muscle changes (e.g., muscle thickness) may lack specificity for IBM diagnosis.

Data Availability

Burlina et al. provided the dataset, "Automated diagnosis of myositis from muscle ultrasound: Exploring the use of machine learning and deep learning methods," which is accessible online at https://github.com/jalbayd1/myopathy_US.

Declarations

Competing Interests

The authors have no relevant financial or non-financial interests to disclose. The authors have no conflicts of interest to declare that are relevant to the content of this article. All authors certify that they have no affiliations with or involvement in any organization or entity with any financial interest or non-financial interest in the subject matter or materials discussed in this manuscript. The authors have no financial or proprietary interests in any material discussed in this article.