Epistemic and aleatory uncertainty guided transparency masking for medical image segmentation

초록

Three-dimensional Medical image segmentation is essential for clinical tasks such as surgical planning and radiotherapy, but is limited by scarce annotation. To address the challenge, self-supervised Masked Image Modeling methods have gained attention. Nevertheless, the prevalent use of random masking tends to ignore rare but clinically significant anatomical structures, and existing MIM approaches, still fail to explicitly quantify or leverage the underlying sources of uncertainty in medical data, remaining limited in transparency about what information is masked and why. We present EATMASK, a unified and transparency-driven MIM framework that leverages uncertainty as a principled signal for explainable masking. EATMASK first masks high-epistemic regions for ambiguous boundaries, then masks high-aleatoric areas to avoid overfitting to noise. A dedicated structural guidance learning enforces anatomical coherence by aligning student and teacher features. Extensive evaluations on dataset with in-domain and out-of-domain scenarios show that EATMASK substantially elevates performance over leading baselines, evidencing its capacity for safer, more robust, and effective representation learning in medical imaging.

키워드