AFTG-Net: A Deep Attention-based Fusion Framework of Topological and Gradient Features for Pathological Image Analysis.

Skeletal muscle pathology is observed by structural disruptions in sarcomeres, increased central nuclei, and changes in myofiber cross-sectional area. In order to classify amyotrophic lateral sclerosis (ALS), diabetes, and healthy controls, pathologists examine the changes in myofiber size using Wheat Germ Agglutinin (WGA) stained histopathological images of various skeletal muscles (quadriceps, gastrocnemius, tibialis anterior, extensor digitorum longus, and soleus). Histological image analysis of skeletal muscle pathology is laborious and subject to inter- and intra-user variability, which can affect diagnosis accuracy and consistency. Conventional techniques like ImageJ-based tools are time-consuming and produce varying outcomes due to their manual cell counting, segmentation, and thresholding. This study introduces AFTG-Net, an attention-based machine learning framework that classifies skeletal muscle histopathological images using complementary geometric and topological descriptors. The model uses globally structural information from Topological Data Analysis (TDA) based on persistent homology and local edge and texture patterns from the Histogram of Oriented Gradients. We suggest a cross-weighted fusion approach that uses cosine similarity to adaptively balance the contributions of these heterogeneous features in order to improve their discriminative power. This integration enables the model to effectively distinguish pathological changes associated with amyotrophic lateral sclerosis (ALS) and Type I diabetes from healthy muscle tissue. We conducted comprehensive comparisons with various state-of-the-art and baseline methods, such as traditional feature-based and deep learning models. We assessed all models by analyzing WGA-stained skeletal muscle images from wild-type and disease models (G93A*SOD1 for ALS and Akita for type 1 diabetes). AFTG-Net outperformed all other models by achieving 92% classification accuracy in distinguishing healthy and diseased muscle fibers. By reducing human intervention, subjectivity, and analysis time, AFTG-Net improves scalability and diagnostic consistency, making it a valuable tool for both biomedical research and clinical practice.