Gallbladder diseases present a significant clinical challenge due to their diverse manifestations and the difficulty of accurate interpretation in ultrasound imaging. Manual assessment of gallbladder ultrasound images is time consuming, operator dependent and may delay clinical decision making, motivating the development of automated diagnosis approaches. In this study, we propose a customized capsule network architecture, termed GBCapsNet, for multi-class disease classification using ultrasound images. The model incorporates a modified routing mechanism designed to improve feature representation and class discrimination. The proposed architecture was evaluated across multiple training-test splits, demonstrating high classification performance under image level splits (maximum accuracy of 99.91% and AUC of 1.0). However due to the use of image level splitting rather than patient level separation, these results should be interpreted with caution. Further validation using patient level splits and external datasets is required to establish clinical generalizability. To assess the reliability of predicted probabilities, post training calibration was performed using temperature scaling, resulting in reduced Expected Calibration Error (ECE). These results indicate improved alignment between predicted confidence scores and observed outcomes, although broader validation is required to establish generalizability. To the best of our knowledge, this work represents one of the early investigations into the applications of capsule networks for automated gallbladder disease diagnosis from ultrasound images, Overall, the findings suggest that capsule-based architectures are a promising direction for improving automated interpretation of gallbladder ultrasound data, warranting further validation on larger and more diverse clinical datasets.
Golla et al. (Wed,) studied this question.