A first explainable-AI-based workflow integrating forward-forward and backpropagation-trained networks of label-free multiphoton microscopy images to assess human biopsies of rare neuromuscular disease.

BACKGROUND AND OBJECTIVE: Diagnosis of rare neuromuscular diseases often relies on muscle biopsy analysis, which varies based on the evaluator's experience. Advances in deep learning show promise in improving diagnostic accuracy by identifying standardized features and phenotypic expressions in biopsy images. Explainable artificial intelligence extracts these features from the neural network's "black box," ensuring compliance with European ethical standards for the use of clinical data in real-world applications. This study proposes a clinic-friendly workflow, based on explainable artificial intelligence. It combines forward-forward and backpropagation-trained convolutional networks to identify complementary features of Duchenne Muscular Dystrophy. Our proposal sets the forward-forward training, applied here for the first time on biomedical images, as a potential new standard for interpretable deep learning applications in clinics.

METHODS: We analyzed a multiphoton microscopy dataset of 1600 images from 16 human muscle biopsies obtained during elective spinal surgery, combining autofluorescence, second and third harmonic generation signals. Class Activation Maps unveiled the dual decision-making process of the convolutional network, independently trained with both standard backpropagation and forward-forward algorithms. We evaluated the significance of the discovered features by using the Mann Whitney method. Entire biopsies were analyzed by providing an attention metric, computed as the weighted mean of all significant parameters.

RESULTS: Backpropagation, gold standard for 35 years, and forward-forward achieved over 90 % accuracy in distinguishing healthy and diseased patients tissue. Class activation maps revealed that, when trained independently with both algorithms, the same network identifies Duchenne Muscular Dystrophy tissue by focusing on different features. Both methods identified intramuscular collagen as a key feature. Backpropagation also highlighted collagen waviness and fatty tissue. Forward-forward emphasized collagen density. We integrated these complementary insights, validated by significance analysis, into a standardized attention metric, allowing a multi-structural, quantitative assessment of tissue changes and highlighting areas for further clinical analysis.

CONCLUSIONS: Our workflow, using clinically routine biopsies and transparent diagnostic features, demonstrates the unique potential of forward-forward in providing novel, reliable, interpretable results from biomedical images. This paves the way for its dual use with backpropagation, shifting benchmarks by enabling the discovery of features potentially overlooked by backpropagation across biomedical datasets.