The role of texture analysis of T1 weighted images in diagnosis of chronic kidney disease

Abstract

Chronic kidney disease (CKD) is real and growing, life threatening challenge for modern societies. Nowadays, kidney biopsy is the gold standard for the diagnosis of renal, structural changes, even though it is an invasive procedure with some contraindications and complications. Several studies suggest that microstructural changes during development of kidney disease may give detectable differences in magnetic resonance (MRI) signal, however not always accessible for human perception. The aim of our study was to determine the possibility of using deep neural network algorithms for the analysis of T1-weighted MRI images in the assessment of renal structural changes, particularly in estimating the presence of an active and chronic phases of disease process. The study is based on images of MRI examinations consisting only dixon-based T1-weighted sequence of 52 patients who underwent kidney biopsy and 8 healthy volunteers. The volunteers with no history of any kidney disease formed group “1” and the patients were divided into two groups: group “2”—with active phase of CKD and group “3”—with non-active, advanced phase of CKD, basing on histopathological outcome. The acquired, sorted images were than presented to deep neural network algorithms. Balanced accuracy of verified algorithms in differentiating the study groups were as follows: custom – 83.1%, AlexNet − 81.1%, ResNet-50–93.1%, ViT – 73.5%. The presented algorithms give promise to be efficient, fast and reliable diagnostic tool for patients not suited for kidney biopsy procedure. Additionally, they have potential to be independent of the type of MRI scanner.

Introduction

Chronic kidney disease (CKD) is real and still growing challenge facing modern national healthcare systems. At present, it is estimated that 10.6% of European population is suffering from CKD¹. According to KDIGO (Kidney Disease Improving Global Outcomes) five stages of CKD are distinguished. This classification is based on the estimated glomerular filtration rate (eGFR; stages G1-G5) and the severity of proteinuria (stages A1–A3)². The early diagnosis and proper treatment may prevent developing of end stage kidney disease and CKD complications, however, in some cases it is difficult to differentiate active from chronic, irreversible changes.

Radiological examinations, including ultrasound, computed tomography and, more recently, magnetic resonance imaging (MRI), are used in radiological assessment of the urinary system, including the kidneys. They visualize morphological changes as well as the causes of non-renal impairment^2–7. Indirectly, they can provide indicative information about advanced stages of kidney disease like reduced kidney size, change of cortical-spinal differentiation, reduction in the thickness of the renal cortex. However, in most cases, the standard examinations mentioned above do not answer the question of what type of nephropathy we are dealing with, how advanced the process is or whether there is a chance of its reversibility. Most untreated nephropathies lead to irreversible changes in the form of glomerular fibrosis and renal tubular atrophy⁸ and, consequently, to impaired renal function up to advanced stages of CKD resulting in the need for renal replacement therapy.

Nowadays, kidney biopsy^9,10 serves as the gold standard for the diagnosis of structural changes. However, it is an invasive procedure, fraught with many complications and not always possible to perform due to the clinical condition of the patient.

Progress in understanding and treating CKD is accompanied by development of diagnostic tools, including diagnostic imaging. In present scientific discourse the profound analysis of two functional MRI sequences – diffusion-weighted images and blood oxygen level dependent MRI are hoped to be milestones of capturing real time microstructural and functional state of the kidney^11–14.

However, functional MRI techniques are not yet validated in CKD and they are not widely available making it the tool reserved for specialised clinical centres only. On the other hand, some studies suggest that not only functional MRI sequences can reveal damage of renal parenchyma on microscopic level; CKD may give detectable differences in signal also in basic MRI sequences as T1- and T2-weighted images^15–17. This is an particularly promising prospect, especially since extremely dynamic development of resonance diagnostics together with web-based image analysis algorithms provide a perspective of insightful evaluation of organs, including their microstructure.

The aim of our study was to determine the possibility of using such algorithms for the analysis of T1-dependent MRI images in the assessment of renal structural changes, particularly in estimating the presence of an active and chronic phases of disease process.

Material and methods

Renal MRI was conducted on 3T Magnetom Vida (Siemens Healthcare GmbH, Erlangen, Germany) and the scanning protocol consisted of dixon-based T1-weighted images (in-phase, out-phase and water-only images; TR = 4ms, TE₁ = 1.26ms, TE₂ = 2.4ms, TA = 14.78; voxel size: recon − 1.4 × 1.4 × 2.0 mm, acquisition – 1.56 × 1.41 × 4.00 mm; FOV = 450 mm; number of slices − 64) aligned to the long axis of each kidney.

The MRI examination took place a day before or a day after kidney biopsy. Images of both kidneys were analysed in patients who had an MRI before biopsy whereas in patients who had a biopsy prior to MRI, only images of the unbiopted kidney were taken into analysis.

The images with very low SNR ratio or containing motion and chemical shift artifacts were excluded and finally we included examinations of:

1. Healthy volunteers, without history of any kidney disease – 8 patients (16 kidneys) – Group 1.

2. Patients who underwent kidney biopsy – 52 patients:

   1. Patients with active, potentially reversible phase of CKD demonstrated by visible histopathological inflammatory infiltration signs – 34 patients, i.e. 14 patients with scanned two kidneys and 20 patients with one (48 kidneys) – Group 2.
   2. Patients with non-active, advanced and irreversible phase of CKD with absent or minimal inflammatory infiltration and histopathological features of kidney fibrosis – 18 patients, i.e. 9 patients with scanned two kidneys and 9 patients with one (27 kidneys) – Group 3.

The precise characteristics of all groups is presented in Table 1. All of the participants are of the Caucasian race.

The patients were suffering from several diseases such as focal segmental glomerular sclerosis (2 group – 5 patients, 3 group – 7 patients), vasculitis (2 group – 17 patients, 3 group − 2 patients), lupus nephritis (2 group – 4 patients, 3 group − 0 patients), tubulointerstitial nephritis (2 group – 2 patients, 3 group − 0 patients), tubular nephropathy (2 group – 1 patients, 3 group − 0 patients), IgA nephropathy (2 group – 5 patients, 3 group − 4 patients), diabetes related nephropathy (2 group – 0 patients, 3 group − 2 patients), 3 patients had end-stage kidney disease.

The histopathological examination was performed by two independent nephropathologists unrelated to the study. The chronic, irreversible changes in kidney biopsy specimens were diagnosed on the basis of the presence: pathologic global glomerulosclerosis with involvement of > 50% glomeruli with interstitial fibrosis covering > 50% of the area and tubular atrophy > 50% tubules⁹.

Our trial and experimental protocol was approved by Bioethical Committee associated with Medical University of Lodz (decision RNN/206/20/KE, dated 08/09/2020). All methods used in our study were carried out in accordance with all relevant guidelines and regulations. Written informed consent was obtained from all participants before enrolling to this prospective study.

Background

Texture analysis (TA) has a long record of successful applications in classification of medical images. For a comprehensive overview, see e.g.^18,19. However, examples of studies devoted specifically to MR renography are relatively scarce^20,21.

In ²², the efficacy of TA was assessed for classifying renal dysfunction into three stages: healthy control, moderate, and severe states. The research involved 166 CKD patients categorized based on GFR levels using two cutoff thresholds (30 and 60 ml/min/1.73 m²). Various magnetic resonance imaging examinations, including T1-weighted, diffusion-weighted imaging (DWI), and blood oxygen level-dependent (BOLD) scans, were conducted. Texture features extracted from these scans included first- and second-order statistics, such as histograms, grey-level co-occurrence matrix, run-length matrix, and grey-level dependence matrix. The model construction process included sequentially selecting the most significant features and training a range of classifiers, such as decision trees, support vector machines (SVM), and linear discriminant analysis. The highest accuracy, reported at 82.8% (macro average), was achieved with the SVM classifier using the radial basis function kernel.

In another paper²³, where TA was used for renal performance assessment, T2-weighted images of diabetic patients were analyzed. The subjects were divided into three groups: normal, mild and severe renal impairment. Texture parameters were calculated using the robust features model. Then, the feature vectors were classified by the SVM algorithm. Again, the achieved results (area under the curve value equal to approx. 0.85 for each class) proved feasibility of classical texture analysis in differentiation of various renal tissue states.

In this study, we developed the custom network architecture but also studied the standard backbones pretrained on ImageNet1K dataset²⁴. In the former case, it was necessary to augment the available collection of images, as the number of patients included in the study was relatively small compared to the number of network parameters. Without artificially increased variability of the training examples, the DL model cannot generalize to test data. In the latter case, on the other hand, the model development utilizes transfer learning mechanism. It is assumed that the pretrained network backbone is already capable of extracting features sensitive to various image patterns and only the network head has to be fine-tuned, which releases the pressure for data augmentation.

Data preparation

As a first step to prepare images for training we performed segmentation of the kidney in the quasi-coronal plane, i.e. the acquisition plane aligned with the oblique orientation relative to the orthogonal body directions (see Fig. 1). The segmentation was accomplished manually by an expert in medical image analysis using the editing tools available in the Slicer 3D software. The quality check of the segmentations was performed by an experienced radiologist. For each patient, the kidneys were delineated in several middle cross-sections (from 5 to 10 slices depending on the kidney size), embracing renal cortex and medulla. Since due to the Dixon protocol inherent characteristics, in-phase (IP), opposed-phase (OP) as well as water-only (WO) images are automatically registered, the segmentation masks were drawn only in one of the channels. We chose opposed-phase image for that purpose as in the analyzed material it exhibited the best contrast between renal tissues and the surroundings. Then, the kidneys were masked-out from the original 2D slices by multiplying the corresponding images and segmentations. Next, the images submitted to further analysis were limited to the region occupied by the kidney. Moreover, it was required by the custom network architecture, that the size of the input images was also kept constant, so that the size of classification head consisting of the fully connected layer could be adjusted accordingly. Hence, in each slice, a kidney was cropped by a window of size 112 × 112 pixels centered over a segmentation mask. The results of the above-described process are presented in Fig. 2, which shows examples for all three classes considered in this study. These examples are presented in pseudo-colors, as IP, OP and WO channels for visualization purposes were concatenated and stored together as an RGB file (24-bit encoding, 8 bits per channel). However, the training images were normalized to 0–1 range and saved as NumPy arrays of 32-bit float numbers in order to preserve original data bit depth.

Fig. 1.

Visualization of imaging plane orientation.

Fig. 2.

Examples of masked image cross-sections analyzed in the study. In-phase, opposed-phase and water-only channels stacked together forming a pseudo-color RGB image.

The last step in the data preparation was image augmentation. The designed transformation function was composed of shifting (limited to 5% of the image width), scaling (5% limit), rotation (up to 15 degrees), horizontal and vertical flip, and brightness alteration. These effects were randomly applied to input images, variable number of times to obtain balanced distribution of classes in the final augmented dataset. Table 2 summarizes the cardinalities of the image subsets in particular CKD classes.

Table 2.

Cardinalities of datasets prepared for training.

Patients group	Number of images	Total number of slices	No. augmented slices (no. of augmentations)
1	14	179	2685 (15)
2	34	398	2786 (7)
3	26	247	2717 (11)

Table 1.

Presentation of analyzed groups’ basic characteristics.

Group	N	Gender		Mean age (years)	Mean eGFR	Mean proteinuria (g/24 h)
		Females	Males
1	8	3	5	44.50	60 <	–
2	34	16	18	50.94	40.18	4.83
3	18	5	13	55.06	38.23	2.54

Eventually, it must be noted that patient information was retained for all processed slices. The training and evaluation was performed in the leave-one-patient-out scenario, i.e. we trained a series of models, each time separating one patient from both train and validation sets, and then tested the performance of a given classifier on this selected patient data.

Custom deep neural network

The first DL model examined within this study was a relatively simple network composed of convolutional layers in the feature encoding backbone and fully connected layers in the classification head. The 3-channel input image is passed through the sequence of 3 convolutional blocks formed by a set of 2-dimensional convolutional filters, ReLU activation function, max pooling operation and batch normalization layer. In each block, max pooling operator reduces the input resolution by a factor of 2. The number of output channels in the first layer is equal to 16 and doubles in each consecutive stage. The sizes of the convolution filters are set to 5 × 5 with stride = 1 in the first 2 layers, and 3 × 3 with stride = 2 in the third layer. The padding is adjusted so that the resolution of the extracted feature maps are the same as their corresponding layers input. As a result, the input to the classification head is a 64-channel image embedding map of size 7 × 7, which is then flattened and submitted to the cascade of 3 fully connected layers, the first two of which have again ReLU activation function, whereas the output of the last one is passed to the SoftMax function to obtain final class prediction.

Pretrained models

In the second part of the experiments we included 3 well-known architectures which proved effective in various image recognition applications. These were AlexNet, ResNet-50, and Vision Transformer, which are briefly summarized below.

The first utilized model was a classic AlexNet model²⁵, which could be viewed as a similar but more complex structure than our proposed one. It has 5 convolutional layers and 3 fully connected layers. It uses ReLU activation functions, and the dropout technique to mitigate overfitting. It also uses local response normalization to improve generalization. Despite its simplicity, AlexNet is still widely used in the medical field for the classification of e.g. tumors and other diseases^26–28. Although, as said, there are no significant architectural differences between AlexNet and our custom model, we decided to include it in the study due to availability of weights pretrained on ImageNet dataset.

ResNet-50²⁹ is composed of so-called residual blocks built out of overall 50 convolutional, pooling, and fully connected layers. It uses a bottleneck design, which reduces the number of parameters and computations by placing 1 × 1 convolutions before and after 3 × 3 convolutions. It follows two basic design rules: (1) the number of filters in each layer is the same depending on the size of the output feature map, and (2) if the feature map size is halved, the number of filters is doubled to maintain the time complexity of each layer. ResNet-50 is divided into five stages, each with a different number of residual blocks and output feature map sizes. The first stage has one convolutional layer and one max pooling layer, and the last stage has one average pooling layer and one fully connected layer. The stages in between have different numbers of bottleneck residual blocks, as shown in Table 3.

Table 3.

Layers configuration in the ResNet-50 model.

Stage	Output size	Number of blocks
1	112 × 112	1
2	56 × 56	3
3	28 × 28	4
4	14 × 14	6
5	7 × 7	3

We employed ResNet-50 to leverage its core architectural component, i.e. residual connections between input and output feature maps of a block, which mitigates the problem of vanishing gradients. Therefore, on hand it is possible to exploit capability of a very deep network to extract embeddings suitable for complex, non-linear scenarios, and on the other hand, it allows avoiding the degradation of classification accuracy.

Lastly, the Vision Transformer (ViT)³⁰ model employs a transformer-like architecture which has recently gained popularity. In this design, the image is first partitioned into a sequence of fixed-size patches, each of which is linearly embedded. Patch embeddings are then concatenated with the positional encodings and the resulting sequence of vectors is fed to a standard Transformer encoder.

The ViT architecture consists of a series of transformer blocks. Each transformer block consists of two sub-layers: a multi-head self-attention layer and a feed-forward layer. The self-attention layer allows the model to attend to different parts of the input sequence, while the feed-forward layer applies a point-wise fully connected layer to each position independently. The output of each transformer block is passed through a layer normalization layer before being fed to the next transformer block. The final output of the last transformer block is passed through a multi-layer perceptron (MLP) to produce the final classification output. The number of transformer blocks differes among the ViT variants. In the base version, utilized in this study, there are 12 such blocks.

In comparison to ResNet, ViT architecture is more computationally efficient, requiring fewer parameters. More importantly, ViT models comprise of the self-attention mechanism allowing the model to attend to different parts of the input sequence, making it easier to understand how the model is making its predictions.

The rationale for selecting the above-described models was to probe the capabilities and biases of different DL architectural families on T1-weighted MR images for chronic kidney disease diagnosis. Thus, we deliberately chose four clinically and methodologically diverse models that collectively cover a broad spectrum of inductive biases and learning paradigms. The custom CNN provides a baseline that reflects the practical limits of training from scratch with a compact capacity, allowing us to assess whether constrained representational power or limited data hinder performance relative to more parameter-rich networks. One of the main advantages of a custom architecture is that it can be tailored to the specific needs of the problem at hand. In this study, e.g., we manipulated with the size of convolutional filters in order to obtain optimal representation of the kidney tissue appearance in the latent embedding space. Moreover, we expected that the custom model would better reflect specific data patterns inherent in the analyzed images. Eventually, our design is lightweight as compared to ResNets or Vision Transformer. Consequently, it can be retrained more efficiently, which is crucial for hyper-parameter tuning and when only new data become available.

Pretrained models, on the other hand, are trained on large, general-purpose datasets, which yields generic image descriptors in the embedding space. AlexNet, a well-established CNN with a deeper architecture than the custom model, affords a historical and empirical point of comparison to contemporary architectures, helping us gauge whether modest architectural advances translate into meaningful gains after fine-tuning on grayscale MR data. ResNet-50 embodies the prevalent modern CNN paradigm with residual connections, enabling deeper feature hierarchies while maintaining stable gradient flow; this choice tests whether depth plus skip connections—and the associated transfer-learning benefits from ImageNet pretraining—translate into robust discrimination of complex tissue textures and structural patterns in CKD assessment. Vision Transformer (ViT) introduces a fundamentally different inductive bias: self-attention over image patches can capture long-range dependencies and global context that CNNs may miss, providing a crucial counterpoint to convolution-based models and revealing whether data scale and domain adaptation are sufficient for transformers to outperform traditional CNNs in a medical-imaging setting. Together, these four models enable a systematic decomposition of performance drivers—depth and residual learning, architectural efficiency, transfer learning from natural images, and transformer-based global modeling—while offering practical breadth for fair comparison under a modest medical dataset, including grayscale adaptation and consistent training protocols. This design also aligns with widespread conventions in medical imaging literature, facilitating interpretability, replicability, and generalizability of the results.

Results

As mentioned above, models training was executed separately for each patient, each time by taking a given patient images only to the test set. The rest of the data were split into train and validation folds in proportion 80% : 20%. For every model, the training loss was calculated using the cross-entropy function and the training was launched for 100 epochs. The final model was the one which performed best on validation loss and it was later used to calculate the prediction on a test patient. Weights optimization procedure was controlled by the Adam algorithm. In the case of custom network, images were processed in the original resolution, for AlexNet and ResNet-50, they were upscaled to 224 × 224 pixels using bi-cubic interpolation, whereas for ViT to the size of 384 × 384 pixels.

Secondly, for each test patient an output class label was determined based on the majority of classes inferred for its individual image slices. To avoid ambiguous predictions, we set a threshold of 70% and at least 5 slices indicating the same class to determine a final label. Otherwise, a patient would be marked unclassified. These parameters were chosen to balance diagnostic reliability with practical constraints. The 70% threshold reflects a pragmatic compromise between the ideal of perfect agreement and the inherent variability in renal morphology and texture patterns across disease categories. Similarly, requiring a minimum of five slices ensures sufficient evidence for a robust decision while accommodating cases of advanced CKD where fewer slices may be available for classification. Based on these assessments, we calculated with respect to each of the classes the sensitivity or true positive rate TPR, the specificity or true negative rate TNR, and the mean and balanced accuracy scores mACC and bACC. The obtained results are shown in Table 4.

Table 4.

Evaluation metrics obtained for the tested neural network models (best scores shown bold).

		mACC	bACC	Sensitivity	Specificity
Custom	1	82.8%	83.1%	83.3%	100.0%
	2			93.8%	90.0%
	3			72.2%	97.7%
AlexNet	1	78.4%	81.1%	75.0%	96.0%
	2			90.6%	90.0%
	3			77.8%	95.5%
ResNet-50	1	88.9%	93.1%	91.7%	97.9%
	2			100.0%	92.9%
	3			87.5%	100.0%
ViT	1	65.9%	73.5%	75.0%	94.0%
	2			84.4%	86.7%
	3			61.1%	93.2%

Discussion

It can be observed that apart from ViT, all the models achieved high level of accuracy. The best performing architecture occurred the ResNet-50 while the proposed design came in second place. The superior performance of ResNet-50 over the custom CNN, AlexNet, and ViT can be attributed to a combination of architectural and data-regime factors. Residual connections in ResNet-50 enable training of a deeper network with stable gradient flow, which supports learning richer hierarchical representations needed to capture the texture and structural patterns characteristic of T1-weighted MR images. The bottleneck design enhances parameter efficiency, allowing effective feature extraction without overfitting given our dataset size. Pretraining on ImageNet provides robust low- to mid-level feature detectors, which, after appropriate fine-tuning and grayscale-channel adaptation, translate well to MR imaging features. In contrast, the ViT typically requires larger datasets or domain-specific pretraining to realize its potential, and the custom CNN trained from scratch lacks both the depth and optimization stability to match the representations learned by ResNet-50. AlexNet, while fine-tuned, remains shallower with an older architectural design that is less capable of capturing the nuanced patterns present in T1-weighted MR images. Collectively, these factors explain why ResNet-50 achieved the best performance in our experiments.

It must be noted, however, that the training experiments in the case of the custom network revealed significant dependence of the final classification score on the initial weights state. We experimented with several initialization methods, such as He and Glorot initializations, with no significant effect on the stabilization of the loss optimization process. Thus, although it occurred feasible to differentiate CKD classes based on image texture content, the patterns responsible for this discrimination are subtle and more extensive experiments will be needed to confirm our conclusions.

In radiological approach to CKD, the first stage focused on features of T1-dependent imaging that could be biomarkers of progression of structural changes in kidney parenchyma. Initially, correlations between MRI signal intensity on T1 images or time values on T1-dependent parametric maps and GFR values were sought. Several researches groups have demonstrated the presence of this relationship in animal models^31,32 as well as in studies in patients with CKD^33–35. However, the fundamental limitation of these methods was the lack of versatility of measurements, which is due to differences resulting from the technical variations of scanner models and protocol settings in imaging centres. This limitation forces the establishment of specific standards for each centre exclusively. Hricak et al.^36–38 proposed an interesting solution to the above problem by introducing a ratio of cortical-core differentiation calculated on the basis of signal T1 intensity. However, in the cited studies, although the association of imaging parameters with renal function was proven, the calculations made did not allow to determine the degree of renal failure, much less to assess the activity of the inflammatory process.

Another very interesting step in the search for a method that could replace kidney biopsy was the introduction of texture analysis of T1-dependent images to determine the stage of CKD development. One of the first reports in this field was the promising but methodologically limited work of Yu et al.³⁹ They proved the possibility of differentiating healthy volunteers from patients with stage 3 CKD in the course of diabetes on the basis of several selected texture features. Further, Zhang et al.⁴⁰ proposed an effective method to qualify patients into four separated groups of renal CKD regardless of the disease entity causing CKD. The above reports confirmed the possibility of linking texture analysis with GFR value, but still did not allow to assess the degree of kidney disease activity, what determines further clinical management and prognosis.

In our knowledge, this report centered to create texture analysis algorithm capable to differentiate the degree of disease activity in attempt to replace histopathological examination is the first in the literature. For our study, we decided to include the widest possible range of features available on T1-dependent images – both morphological i.e. kidney size (length, organ width, parenchyma width) and their shape, as well as texture parameters (including vector characteristics: direction, turn, pixel brightness, contrast or entropy). We did not make any selection of attributes to provide the analysis as comprehensive, objective and precise as possible; the neural network had access to all the features without the researcher’s influence on their selection. Also, we decided not to limit our study group by specific etiology of CKD. According to the authors, development mechanisms of CKD are mostly similar in various pathologies and include arteriolar remodelling, hypoperfusion, thickening of basement membrane, degeneration of podocytes, glomerular remodelling and finally renal fibrosis^41–46. Thanks to this approach, it became possible to develop an algorithm that was able to reveal the presence or absence of an active CKD phase and the algorithm is universal, the reasons causing CKD - both primary and secondary, do not affect its performing.

It should be emphasized that the proposed algorithm is potentially independent of the type of MRI scanner, because it is not based on the features of the MRI signal, but on the characteristics of pixels of the image, similarly to corticomedullary differentiation in estimation of renal function⁴⁷ or characterising adrenal masses with chemical shift indexes^48,49. However, this hypothesis needs multi-center trials on larger patients groups. Never-the less, this types of algorithms are easy to use and they take only 5 min of a radiologist` work at the diagnostic station. Then, in a relatively short time we get an answer whether the disease affecting the patient is at that moment active or not.

Algorithm detecting changes in kidney microstructure, significantly earlier than the change in their functioning creates the perspective of early detection of alterations before it reaches an irreversible stage. Therefore, from a clinical perspective, having a fast and effective diagnostic test, based on performing a single T1-dependent sequence in standard MRI scanners, not requiring special preparation of the patient, taking about 2–3 min, with the report available almost immediately after the examination is very looked-for from the point of view of the clinician. It allows for answering the diagnostic question - whether we are dealing with potentially reversible changes or whether the process is advanced and unfavourable in prognosis, especially in cases of contraindications to kidney biopsy. Although it must be admitted that MRI will not replace a kidney biopsy in the matter of assessing the type of glomerulopathy.

The leave-one-patient-out (LOPO) evaluation strategy adopted in this study was primarily motivated by the limited dataset and the need to ensure patient-level independence during model assessment. While this approach involves training multiple models during experimentation, it does not reflect the intended clinical deployment workflow. In practice, a single consolidated model would be trained on all available data following rigorous validation and hyperparameter optimization. The LOPO results provide a conservative estimate of the model’s ability to generalize to unseen patients, which is critical in medical imaging applications where inter-patient variability can be substantial. Nevertheless, we acknowledge that true generalizability requires validation on larger, multi-center datasets encompassing diverse imaging protocols and patient demographics. Future work will focus on aggregating such datasets and performing external validation to confirm robustness and clinical applicability. Additionally, regulatory considerations and computational feasibility strongly favor a unified model rather than patient-specific training, ensuring scalability and practicality in real-world clinical settings.

Our study has several limitations that relate to group size imbalances. The modest overall sample stems from the selective biopsy indications for CKD, leading to disproportionately larger disease groups relative to the control group. This imbalance, together with within-group heterogeneity (e.g., diverse GN subtypes in group 2), can bias model learning toward features characteristic of the majority classes and reduce sensitivity and precision for the underrepresented group, increasing the variance of performance estimates in that subgroup. The single-center design and predominant use of a limited set of scanners and field strengths further constrain external validity and generalizability to other clinical settings. To mitigate these issues, future work should pursue (i) larger, balanced cohorts acquired across multiple centers, (ii) harmonization of imaging protocols (across manufacturers and field strengths), and (iii) methodological approaches to address imbalance during training (e.g., class-weighted losses, balanced sampling, or focal loss) alongside reporting per-group performance with confidence intervals to convey uncertainty.

Conclusion

Analysis of T1-dependent images using texture features and patterns may become a sensitive diagnostic tool in patients with CKD to differentiate between active and non-active, chronic changes of kidney parenchyma with potential to replace an invasive kidney biopsy.

Author contributions

Conceptualization-M.M., I.K., A.K.-Data curation-M.M., K.S.-Formal analysis-M.M., A.K.-Funding acquisition-I.K., L.S.-Investigation-M.M., I.K., A.K.-Methodology-M.M., I.K., A.K.-Project administration-M.M.-Resources-M.M., K.S., I.K.-Software-M.M., A.K.-Supervision-L.S., I.K.-Validation-M.M., A.K.-Visualization-M.M., A.K.-Writing—original draft-M.M.-Writing—review and editing-M.M., A.K., I.K.

Funding

The authors received no funding for this work.

Data availability

The code used in the analyses is available in https://github.com/aklepaczko/kidney-texture-analysis.git.

Declarations

Competing interests

The authors declare no competing interests.