Consistency-based Semi-supervised Evidential Active Learning Framework for Robust Classification of Radiology Images

Deep learning offers high performance for radiology image classification, but relies on large, expert annotated datasets. Semi-supervised learning and active learning approaches can leverage unlabelled samples and mitigate the annotation burden. Combining these techniques via semi-supervised active learning (SSAL) can compound their benefits, yet the effectiveness of this approach may be hindered by unreliable uncertainty estimation and consistency enforcement. To address these challenges, we propose Consistency-based Semi-supervised Evidential Active Learning (CSEAL). CSEAL is a principled SSAL framework leveraging evidential learning for reliable estimation of predictive uncertainty for consistency enforcement and prioritised sampling. First, we develop evidential counterparts of leading semi-supervised methods with different consistency enforcement mechanisms: Pseudo-labelling, Virtual Adversarial Training, Mean Teacher, and NoTeacher, and demonstrate customisability of CSEAL. Second, we introduce Noise Robust-evidential NoTeacher, an enhancement over evidential NoTeacher, that uses consensus principles and a small-loss inclusion mechanism to learn with noisyannotations. In extensive experiments on many X-ray, CT, and MRI datasets, we demon strate that CSEAL offers substantial performance gains over competitive SSAL baselines and enhances robustness in noisy annotation scenarios. Third, we translate CSEAL into an annotation platform and demonstrate its value as an aid for real-world radiology image annotation. Specifically, our CSEAL-assisted platform performs close to fully supervised learning with very small labelled datasets, en ables accurate auto-labelling, and maintains performance even when only noisy labels from junior annotators are available for training. Our work offers new opportunities to enhance the efficiency of clinical image annotation and model development workflows.