Self-Supervised Deep Learning Classifies Skeletal Muscle Pathology Without Manual Labels

Researchers have introduced SSL-HistoNet, an annotation-free self-supervised deep learning framework that accurately identifies disease-related morphological alterations in skeletal muscle tissue without relying on manual expert annotations. Tested on wheat germ agglutinin-stained skeletal muscle sections harvested from mouse models of amyotrophic lateral sclerosis and Type I diabetes, the model attained a precision of 0.98, a recall of 0.98, and an area under the receiver operating characteristic curve of 0.98. By training directly on unannotated microscopy images, the self-supervised approach matched or surpassed the performance of conventional supervised neural networks and specialized histopathology foundation models. The study demonstrates that representation learning can capture intricate pathological changes across muscle phenotypes while alleviating the heavy dependency on labor-intensive manual labeling in computational pathology.

Microscopic examination of tissue morphology remains the definitive diagnostic standard in clinical pathology, requiring pathologists to evaluate cell dimensions, spatial distributions, and structural irregularities within designated regions of interest. However, manual annotation is time-consuming, labor-intensive, and inherently subject to inter-observer and intra-observer variability. Pathologists frequently perform multiple concurrent analytical tasks, increasing the likelihood of diagnostic omissions or misinterpretations that could delay clinical interventions. While deep learning methods have demonstrated utility across clinical classification tasks, traditional supervised algorithms depend on massive datasets with detailed manual annotations. Furthermore, standard models typically rely on weights pretrained on natural images from ImageNet, which exhibit fundamental visual disparities from histopathology specimens. Medical datasets also frequently exhibit severe class imbalances where healthy samples vastly outnumber diseased specimens, introducing systemic bias into supervised predictions.

These diagnostic obstacles are particularly pronounced in skeletal muscle pathology, an area largely unexplored by modern digital pathology foundation models. Recent foundation architectures, such as CTransPath, UNI, CHIEF, and Virchow, were developed using extensive repositories of whole-slide images prepared with hematoxylin and eosin staining. Although capable in oncology contexts, these models encounter transfer difficulties when applied to skeletal muscle specimens stained with wheat germ agglutinin. Unlike hematoxylin and eosin, which emphasizes nuclear and cytoplasmic staining, wheat germ agglutinin specifically targets cell membranes, outlining myofiber boundaries and connective architecture. Skeletal muscle histology involves distinct structural geometries, small cohorts, and unique disease-related morphological alterations, making it uncertain whether representations derived from hematoxylin and eosin datasets can adapt to membrane-specific imaging. Consequently, domain-specific representation frameworks are needed to extract features directly from skeletal muscle microscopy without manual labels.

To address these limitations, the authors acquired skeletal muscle samples from two murine disease models alongside healthy controls. Amyotrophic lateral sclerosis was modeled using transgenic G93ASOD1 mice, which exhibit progressive motor neuron pathology, muscle fiber atrophy, centrally placed nuclei, and irregular myofiber geometries. Type I diabetes was represented by C57BL/6-Ins2Akita/J mice, a strain displaying structural abnormalities and metabolic muscle dysfunction. Healthy control specimens were obtained from wild-type mice and non-carrier littermates on a C57BL/6 background, with all strains procured from The Jackson Laboratories. Animal care and experimental procedures were approved by the Institutional Animal Care and Use Committee at LSU Health Shreveport following National Institutes of Health guidelines. Skeletal muscle tissues were harvested bilaterally from five anatomical regions: the quadriceps, gastrocnemius, soleus, extensor digitorum longus, and tibialis anterior.

Following tissue extraction under isoflurane anesthesia, specimens were fixed in formalin, paraffin-embedded, and sliced into five-micrometer sections. Antigen retrieval was performed by boiling sections in ten-millimolar sodium citrate buffer at pH 6.0 at one hundred degrees Celsius. Sections were blocked with bovine serum albumin, cold water fish skin gelatin, and Tween-20 in phosphate-buffered saline before incubation with Alexa Fluor 488-conjugated wheat germ agglutinin at five micrograms per milliliter to label cell membranes. Cell nuclei were counterstained with 4′,6-diamidino-2-phenylindole, and slides were mounted using Vectashield medium. Confocal images were collected at twenty-times magnification using a Nikon A1R microscope with Nikon NIS Elements software. The final annotated dataset included 409 images, comprising 87 amyotrophic lateral sclerosis cases, 66 diabetic cases, and 256 healthy controls (183 wild-type and 73 Akita controls), totaling 153 diseased and 256 healthy samples across twelve mice, alongside 139 unannotated images.

To extract patch-level morphological representations without manual guidance, the investigators utilized a SimCLR contrastive learning framework. The training set comprised 548 image patches derived from 139 unlabeled images and 409 labeled images, with diagnostic annotations withheld entirely during pretraining. Contrastive learning paired stochastically augmented views of each image patch to encourage the model to capture invariant structural patterns. The data augmentation pipeline included conditional Gaussian blurring with a kernel size of 23 at probability 0.5, random horizontal flipping at probability 0.5, random color jittering affecting brightness, contrast, saturation, and hue at probability 0.8, random grayscale conversion at probability 0.2, and random resized cropping to 224 by 224 pixels with scale factors between 0.2 and 1.0. All views were normalized using ImageNet statistical baselines before feeding into the convolutional architecture.

The encoder architecture employed a ResNeXt-101-32x8d backbone pretrained on ImageNet, modified by replacing its final fully-connected layer with an identity layer to yield a 2048-dimensional feature extractor. For pretraining, a multi-layer perceptron projection head mapped these features through a 512-dimensional hidden layer with rectified linear unit activation into a 128-dimensional latent space. Optimization was driven by Normalized Temperature-scaled Cross-Entropy loss with a temperature parameter of 0.5, maximizing cosine similarity between positive pairs while penalizing negative pairs within each batch. Training proceeded across 100 epochs using the AdamW optimizer with a learning rate of 0.003, weight decay of 1e-5, batch size of 128, and automatic mixed-precision acceleration. The model state exhibiting the lowest training loss was preserved, after which the projection head was discarded, leaving the learned 2048-dimensional feature extractor intact for downstream classification.

Downstream disease detection was executed by coupling the pretrained feature encoder to a classification network featuring a multi-head attention mechanism. During classifier training, Gaussian noise with a standard deviation of 0.05 was added to the 2048-dimensional input vector to enhance generalization. The vector was reshaped into eight 256-dimensional segments and processed by an eight-head multi-head attention layer with a 0.1 dropout rate to capture interdependencies among feature segments. The attended features were flattened and passed through a four-block multi-layer perceptron comprising dense layers of 1024, 512, 256, and 1024 units with Gaussian Error Linear Unit activations and dropout rates between 0.2 and 0.4. This bottleneck structure encouraged compact representation before concluding in a two-unit dense layer with softmax activation for binary healthy-versus-diseased classification.

To guarantee rigorous validation, the authors implemented stratified twenty-fold cross-validation, preserving healthy-to-diseased class proportions in every fold. The classification dataset consisted of 408 labeled images, split into 337 training images (211 healthy and 126 diseased) and 71 held-out test images (44 healthy and 27 diseased). SSL-HistoNet was compared against supervised convolutional networks (InceptionV3, ResNet50, EfficientNet variants B0, B1, and B2), transformer architectures (ViT-B/16 and Swin-V2-B), and histopathology foundation models (RetCCL, CTransPath, UNI, and Virchow). For fine-tuning baseline models, base networks were frozen and modified with global average pooling, batch normalization, a 256-unit dense layer with rectified linear activation, a 0.2 dropout layer, and a softmax output layer. Models were trained using the Adam optimizer with a learning rate of 0.01, categorical cross-entropy loss, and classification accuracy as the monitoring metric.

In quantitative evaluations, SSL-HistoNet attained the highest overall performance, achieving a precision of 0.98, recall of 0.98, F1-score of 0.98, accuracy of 0.978 (± 0.03), area under the curve of 0.98, Matthews correlation coefficient of 0.953, and Cohen’s Kappa of 0.953. Among comparative baselines, RetCCL delivered the strongest results with a precision of 0.98, recall of 0.98, F1-score of 0.98, accuracy of 0.977 (± 0.04), area under the curve of 0.98, Matthews correlation coefficient of 0.953, and Cohen’s Kappa of 0.952. InceptionV3 led supervised convolutional models with an F1-score of 0.96, accuracy of 0.96 (± 0.11), and area under the curve of 0.95. EfficientNetB0 and EfficientNetB1 achieved F1-scores of 0.93 with areas under the curve of 0.95 and 0.94, while ResNet50 and EfficientNetB2 attained F1-scores near 0.91 with areas under the curve of 0.94.

Normalized confusion matrices demonstrated that SSL-HistoNet achieved balanced classification, correctly identifying 98 percent of healthy and 97 percent of diseased samples, yielding low error rates. RetCCL exhibited similar discrimination, correctly classifying 99 percent of both classes. Conversely, transformer architectures and foundation models showed notable limitations. ViT-B/16 recorded a 33 percent false negative rate, identifying 92 percent of healthy but only 67 percent of diseased samples (F1-score 0.80, accuracy 0.82, AUC 0.87). Swin-V2-B exhibited a 43 percent false negative rate, classifying 96 percent of healthy but only 57 percent of diseased samples (F1-score 0.78, accuracy 0.81, AUC 0.83). Foundation models suffered severe domain shifts: UNI and Virchow achieved F1-scores of 0.40 and AUC of 0.58, detecting only 25 percent and 16 percent of diseased tissues. CTransPath achieved an F1-score of 0.68, accuracy of 0.78, AUC of 0.68, and 53 percent diseased sample accuracy.

Exploratory feature analyses showed that self-supervised representations effectively encoded distributed disease-specific signatures. Standardized feature profiles and class-wise mean curves revealed distinct divergence between healthy and pathological tissues. While individual feature dimensions exhibited partial overlap, multivariate patterns captured morphological irregularities, such as centralized nuclei and myofiber alterations in amyotrophic lateral sclerosis, contrasting with polygonal healthy myofibers having peripheral nuclei. Radar plots confirmed class-dependent shifts across encoding blocks A through D. Crucially, the authors emphasized key methodological limitations: images originated from twelve mice, and stratified twenty-fold cross-validation was conducted at the image level rather than the animal level, meaning intra-animal correlation cannot be fully excluded. Consequently, results represent image-level classification rather than proven biological generalization. Future validation requires larger cohorts with mouse-level cross-validation, multi-class frameworks, cross-tissue generalization, and multi-modal integration with genomic data.

Subject of Research: Self-supervised deep learning for computational pathology of skeletal muscle tissue

Article Title: From unlabeled to labeled: self-supervised deep learning in computational pathology

Article References: Akan, T., Aishwarya, R., Bhuiyan, M. S., Conrad, S. A., Vanchiere, J. A., & Bhuiyan, M. A. N. (2026). From unlabeled to labeled: self-supervised deep learning in computational pathology. BioData Mining, 19(1), Article 81. https://doi.org/10.1186/s13040-026-00586-6

Image Credits: AI Generated

DOI: 10.1186/s13040-026-00586-6

Keywords: computational pathology, self-supervised learning, skeletal muscle, amyotrophic lateral sclerosis, type 1 diabetes, deep learning, image classification, contrastive learning, unlabeled, labeled, self-supervised, deep