SSL-OHE: A self-supervised ensemble approach for early diagnosis of biliary atresia from sonographic images

Accurate and early recognition of biliary atresia from sonographic gallbladder images is crucial for early intervention and optimal outcomes. Employing an effective early screening technique can significantly enhance therapeutic outcomes for children with biliary atresia, especially in remote rural areas. To achieve this goal, leveraging deep learning in medical imaging holds the potential to reduce diagnostic errors, alleviate the workload on radiologists, and expedite the diagnosis process. However, training such deep learning models necessitates extensive and precise annotated datasets, which might be scarce, highly imbalanced, and limited in size due to restricted access and the rarity of the biliary atresia disease. This paper introduces a novel approach, SSL-OHE (Self-Supervised Representation Learning combined with Optimized Heterogeneous Ensemble), for biliary atresia diagnosis. Specifically, an online clustering-based self-supervised method is employed to learn unsupervised image representations, utilizing lightweight base learners initialized from a large general vision dataset. The pre-trained base learners serve as a representation model used in a specific classification task by fine-tuning all parameters in supervised training for the downstream task of diagnosing biliary atresia (BA). Subsequently, a numerical optimization method is employed to determine weights for the various base learners. These models are combined using a heterogeneous ensemble approach to ensure accurate BA detection and mitigate the impact of highly unbalanced data. Extensive experimental results demonstrate the effectiveness of our proposed self-supervised ensemble pretraining scheme, leading to significant improvements in diagnostic performance. This approach outperforms both individual and ensemble deep learning counterparts, even surpassing the expertise of human specialists.