From Non-Expert to Expert: Recurrent Refined Learning for Medical Image Segmentation

Precise lesion segmentation is essential in computer-aided diagnosis and treatment. Prevalent deep-learning-based approaches need high-quality annotations to achieve satisfying performance. However, blurring effects and infiltration hamper the accurate delineation of lesions. A common solution is to divide the annotation process into initial annotation by non-specialized personnel and subsequent modification by expert physicians, where the latent correction patterns and non-expert labels are seldom utilized effectively. To explore their helpfulness, we propose ReReNet, a recurrent refined network for lesion segmentation that learns from non-expert to expert. It achieves progressively refined segmentation results through multiple iterations with tailored discrepancy-aware supervision. During training, as the iterative process perceives discrepancies between refining and expert labels, the model gradually grasps the knowledge of turning the barely correct into the clinically accurate. We validate ReReNet’s capability on three medical image segmentation (MIS) datasets, including magnetic resonance (MR) and computed tomography (CT) modalities. Comparison results indicate that the proposed approach achieves superior performance by introducing the designed recurrent mechanism and outperforms mainstream methods, demonstrating the effectiveness of mining hidden correction patterns by utilizing non-expert information.