GallNet: An Explainable Deep Learning Model for Multi-Class Gallbladder Disease Classification Using Ultrasound Imaging

Accurate and interpretable gallbladder disease categorization from ultrasound images is of prime importance for timely intervention as well as appropriate clinical decision-making in hepatobiliary practice. Current deep learning methods typically fail to cope with low class diversity, morphological heterogeneity, and opaqueness. We introduce GallNet, a thoughtfully constructed, memory-conscious convolutional neural network (CNN) for multi-class classification of gallbladder disease into nine clinically relevant classes with membranous/gangrenous cholecystitis, gallbladder perforation, and carcinoma as its milder and more severe manifestations. The hierarchical four-block architecture of GallNet (filters 32–256) learns dense hierarchies of features, while adaptive batch normalization and dropout aid in generalizability. To mitigate extreme class imbalance, we incorporate a class-weighted sparse categorical cross-entropy loss function. To the best of our knowledge, this is the first study to integrate both Grad-CAM and SHAP for explainable multi-class classification of gallbladder diseases from ultrasound images, offering complementary spatial and feature-level insights. We also point out reproducibility and stability, training on five independent runs with variant random seeds and reporting mean ± standard deviation of all the major performance metrics. Testing on a real-world diverse ultrasound dataset provides a test accuracy of 98.50 ± 0.05%, per-class precision, recall, and F1-scores > 96%, and class-wise AUCs ≥ 0.995. GallNet outperforms current state-of-the-art CNN models and demonstrates real-world viability as an efficient, transparent, and computationally smart gallbladder disease diagnostic aiding system.