CycleMix: A Holistic Strategy for Medical Image Segmentation from Scribble Supervision (2203.01475v2)

Abstract: Curating a large set of fully annotated training data can be costly, especially for the tasks of medical image segmentation. Scribble, a weaker form of annotation, is more obtainable in practice, but training segmentation models from limited supervision of scribbles is still challenging. To address the difficulties, we propose a new framework for scribble learning-based medical image segmentation, which is composed of mix augmentation and cycle consistency and thus is referred to as CycleMix. For augmentation of supervision, CycleMix adopts the mixup strategy with a dedicated design of random occlusion, to perform increments and decrements of scribbles. For regularization of supervision, CycleMix intensifies the training objective with consistency losses to penalize inconsistent segmentation, which results in significant improvement of segmentation performance. Results on two open datasets, i.e., ACDC and MSCMRseg, showed that the proposed method achieved exhilarating performance, demonstrating comparable or even better accuracy than the fully-supervised methods. The code and expert-made scribble annotations for MSCMRseg are publicly available at https://github.com/BWGZK/CycleMix.

Analyzing CycleMix: A Holistic Strategy for Medical Image Segmentation

The paper "CycleMix: A Holistic Strategy for Medical Image Segmentation from Scribble Supervision" addresses the intrinsic challenge in medical imaging tasks: creating large fully annotated datasets is both costly and time-consuming due to the necessity for input from specialized clinical experts. Recently, there has been a significant interest in methods exploiting weakly supervised learning (WSL) and semi-supervised learning (SSL) to leverage partial annotations or unlabeled data for model training. This paper contributes to the body of research on WSL by focusing on a specific form that utilizes scribble annotations, which are simpler for experts to generate than full pixel-wise annotations.

Proposed Framework: CycleMix

The authors propose a novel framework named CycleMix designed for medical image segmentation using only scribble annotations. CycleMix leverages a combination of mix-based data augmentation termed as "mix augmentation" and cycle consistency regularization to achieve commendable performance. The approach is novel in integrating mix-up strategies for exploiting weak labels and emphasizes preserving the intrinsic shape prior knowledge crucial for accurate medical image segmentation.

1. Mix Augmentation of Supervision: CycleMix extends traditional mixup techniques to two-stage augmentation—increment and decrement of scribbles. First, it employs Puzzle Mix to increase scribbles by maximizing the saliency of the mixed images. Secondly, random occlusion of regions within images decreases scribbles to further diversify the annotations, enabling the augmentation of supervision with varying levels of scribble retention.
2. Cycle Consistency Regularization: On top of mix augmentation, CycleMix introduces consistency regularization across global and local levels. Global consistency maintains the mix-invariant property by ensuring segmentation consistency across mixed and unmixed image pairs. Local consistency penalizes the segmentation model when producing disallowed disconnected structures, enforcing anatomical accuracy.

Evaluation and Results

CycleMix was evaluated against two prominent datasets, the Automated Cardiac Diagnosis Challenge (ACDC) and the MSCMRseg dataset. The evaluation establishes CycleMix's ability to outperform existing state-of-the-art WSL approaches. Notably, the results illustrate that CycleMix matches or surpasses fully supervised learning models in which full manual annotations were available.

• ACDC Dataset: On this dataset, CycleMix significantly outperforms several baseline and weakly supervised methods, achieving average Dice Scores of 84.8%, which is competitive even against fully supervised networks. The approach is proven superior by a substantial margin, particularly in structures with high anatomical variability such as the right ventricle.
• MSCMRseg Dataset: CycleMix achieves an average Dice Score of 80.0% on the challenging MSCMRseg dataset, which is substantially higher than the baseline methods. This signifies CycleMix's robustness against variabilities introduced by late gadolinium enhancement—a challenge for traditional segmentation techniques.

Implications and Future Work

The implications of CycleMix are substantial both in practice and theory. Practically, it can significantly reduce the burden of annotating medical datasets by clinical experts, democratizing the scope of medical image analysis. Theoretically, CycleMix showcases the potential of employing mixed-up augmentation coupled with consistency regularization in a weakly supervised setting, opening avenues for applying these principles to other forms of noisy or partial annotations.

This paper leaves room for exploring potential extensions of CycleMix, such as integrating other forms of irregular annotations, employing more diverse consistency penalties, or expanding to 3D medical data. Future work could also involve advancing the CycleMix framework for semi-supervised learning tasks, possibly hinging on self-training or self-supervision paradigms to further maximize its versatility and performance across different domains.

In summary, CycleMix demonstrates a commendable strategy for scribble-supervised medical image segmentation, exhibiting capabilities that not only rival fully supervised methods but also promise broader applicability in diverse medical settings.