Non-iterative scribble-supervised learning with pacing pseudo-masks for medical image segmentation

Scribble-supervised medical image segmentation tackles the limitation of sparse masks. Conventional ap-proaches alternate between: labeling pseudo-masks and optimizing network parameters. However, such iterative two-stage paradigm is unwieldy and could be trapped in poor local optima since the networks undesirably regress to the erroneous pseudo-masks. To address these issues, we propose a non-iterative method where a stream of varying (pacing) pseudo-masks teach a network via consistency training, named PacingPseudo. Our contributions are summarized as follows. First, we design a non-iterative process. This process is achieved gracefully by a siamese architecture that comprises two weight-sharing networks. The siamese architecture naturally allows a stream of pseudo-masks to assimilate a stream of predicted-masks during training. Second, we make the consistency training effective with two necessary designs: (i) entropy regularization to obtain high-confidence pseudo-masks for effective teaching; and (ii) distorted augmentations to create discrepancy between the pseudo-mask and predicted-mask streams for consistency regularization. Third, we devise a new memory bank mechanism that provides an extra source of ensemble features to complement scarce labeled pixels. We evaluate the proposed PacingPseudo on public abdominal organ, cardiac structure, and myocardium datasets, named CHAOS T1&T2, ACDC, and LVSC. Evaluation metrics include the Dice similarity coefficient (DSC) and the 95th percentile of Hausdorff distance (HD95). Experimental results show that PacingPseudo achieves a 68.0% DSC and 14.1 mm HD95 on CHAOS T1, 73.7% DSC and 12.2 mm HD95 on CHAOS T2, 82.9% DSC and 4.3 mm HD95 on ACDC, and 61.4% DSC and 11.9 mm HD95 on LVSC. These results improve the baseline method by >= 3.1% in DSC and >= 14.2 mm in HD95. These results also outcompete previous methods. The fully-supervised method attains a 67.0% DSC and 16.7 mm HD95 on CHAOS T1, 71.2% DSC and 12.6 mm HD95 on CHAOS T2, 84.0% DSC and 3.9 mm HD95 on ACDC, and 72.9% DSC and 7.6 mm HD95 on LVSC. PacingPseudo's performance is comparable to the fully -supervised method on CHAOS T1&T2 and ACDC. Overall, the above results demonstrate the feasibility of PacingPseudo for the challenging scribble-supervised segmentation tasks. The source code is publicly available at https://github.com/zefanyang/pacingpseudo.