PathViT Model for Automated Disease Classification from Skeletal Muscle Histopathology

Abstract

Analyzing skeletal muscle pathology from histological images is labor intensive (requiring manual cell counting, segmentation, and thresholding), time consuming, and prone to inter- and intrauser variability, influencing the accuracy and consistency of diagnoses. To address these difficulties, PathViT, a transformer-based deep-learning model, was designed to automatically distinguish between healthy and diseased muscle fibers, with the aims of reducing human intervention, minimizing subjectivity and variability, and significantly decreasing analysis time compared to conventional manual methods. Skeletal muscle pathology is characterized by changes in myofiber cross-sectional area, increased central nuclei, and structural disruptions in sarcomeres. To investigate these changes in myofiber size, wheat germ agglutinin staining and digital histopathology of skeletal muscle (quadriceps, gastrocnemius, tibialis anterior, extensor digitorum longus, and soleus) was utilized to classify diseased tissue [amyotrophic lateral sclerosis (SOD1∗G93A) and type 1 diabetes (Akita)] versus nondiseased controls. The performance of PathViT in distinguishing diseased versus nondiseased muscle fibers was compared with that of state-of-the-art deep-learning models. PathViT classified healthy and diseased muscle fibers with 96% accuracy, outperforming the other models. This approach enhanced scalability and diagnostic accuracy and decreased variability, making PathViT a potentially powerful biomedical research and clinical tool.