Computational pathology for skeletal muscle disease in mouse models using topological signatures and visual word encoding

Résumé fourni par la source

Tissue analysis is considered the gold standard for the diagnosis of a wide spectrum of disorders. However, pathologists perform labor-intensive evaluations to ensure accurate results. Computational pathology has made significant advances in the development of task-specific predictive models. Nevertheless, traditional pixel- or texture-based features often fail to capture both local and global structural patterns together with their spatial organization. This study developed TopoBoW, a computational framework to objectively characterize morphological patterns using local and global microscopy image features. We developed TopoBoW by integrating Topological Data Analysis (TDA) and Bag-of-Visual-Words (BoVW), combining it with an attention-guided multi-layer perceptron (MLP) trained to distinguish between healthy and pathological muscle tissue. TDA captures global structural features, whereas BoVW encodes local textural. We also utilized visualizations to examine the statistical behavior of feature vectors across disease classes and healthy controls, evaluating their discriminative ability. We compare TopoBoW with several baseline and deep learning models, including TDA-based models, histogram of oriented gradients (HOG), XGBoost classifiers, attention-based MLP models, and modern convolutional and transformer architectures. TopoBoW demonstrated state-of-the-art performance on muscle tissue classification within the image-level cross-validation framework applied to the present preclinical dataset, and outperformed all baselines in terms of all classification criteria, including accuracy, F1-score, and AUC. With its interpretable feature-based computational framework, TopoBoW can assist with pathological research, education, and interactive diagnostic workflows by integrating global structural and local textural information from images.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.