Semi-supervised Medical Image Classification Made Easier with Causality-Driven Learning.

Abstract

Semi-supervised learning (SSL) has significantly improved medical image analysis, especially in cases with limited labeled data, by leveraging information from unlabeled data. Consistency regularization, a key method in SSL, promotes alignment between output distributions across different perturbations. However, conventional consistency learning indiscriminately aligns sample relationships, which may lead to spurious results when non-causal features dominate the learned representations. The non-causal but discriminative feature may replace the intrinsic semantic information when considering the invariant constraints on different perturbations, which leads to spurious consistency. To address these problems, we introduce a novel causality-driven SSL framework to polish the consistency learning process and reduce the impact of confounders or bias for effective medical image classification performance. Integrating causal inference, our framework emphasizes extracting and leveraging causal features, thus sharpening the generalization and interpretability of the model. The proposed Causality-driven Consistency paradigm can independently be designed or applied to existing consistency regularization-based semi-supervised architectures to enhance their efficiency and accuracy without substantially increasing parameters and computational burden. Extensive experiments are conducted to demonstrate the superiority of the proposed approach.