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.
Introduction
Deep learning has revolutionized medical image segmentation, yet the reliance on pixel-perfect "gold standard" masks remains a significant bottleneck. Creating these annotations for organs like the heart or liver requires hours of labor from clinical experts. Scribble supervision—where experts draw sparse lines to indicate tissue types—offers a faster, more cost-effective alternative. However, the sparsity of scribbles makes it difficult for models to learn precise boundaries and anatomical shape priors. Existing methods to expand these scribbles often introduce noise, leading to unstable training.
To address this, we introduce CycleMix, a framework evaluated on two major cardiac MRI benchmarks: the ACDC dataset (100 patients, cine-MRI) and the MSCMRseg dataset (45 patients, LGE-MRI). Notably, we have released a new set of expert scribble annotations for MSCMRseg to facilitate further research in this domain.
Methodology
CycleMix is built upon two primary pillars: specialized mix augmentation and dual-level cycle consistency regularization.
Two-Step Mix Augmentation
Standard Mixup strategies are often insufficient for the sparse nature of scribbles. CycleMix utilizes a specialized version called Puzzle Mix in two distinct phases:
- Increments: This step intelligently combines two images to maximize saliency, effectively increasing the density of annotated pixels the model processes in a single step.
- Decrements: By applying random occlusion (hiding parts of the image), the model is forced to localize objects even when portions of the scribble are missing, enhancing the segmentor's robustness.
Dual-Level Cycle Consistency
To ensure the model learns realistic anatomical shapes without full masks, CycleMix enforces consistency at two levels:
- Global Consistency: The model is penalized if its prediction for a "mixed" image does not match the mathematical mix of its individual predictions for the source images.
- Local Consistency: This acts as an anatomical prior. Since organs are typically single, connected structures, CycleMix identifies the largest connected component in a prediction and trains the model to minimize the distance between its raw output and this "cleaned" version. This encourages the generation of solid, interconnected shapes rather than fragmented pixels.
Experimental Results
CycleMix achieved state-of-the-art (SOTA) performance for scribble-based learning across both benchmarks. On the ACDC dataset, it reached an average Dice score of 84.8%, outperforming the standard UNet baseline (76.6%) and even slightly exceeding a fully supervised UNet (84.0%). On the more challenging MSCMRseg dataset, CycleMix achieved a Dice score of 80.0%, a significant improvement over the 37.8% baseline.
Key insights from our results include:
- Saliency as Supervision: Prioritizing high-saliency regions ensures sparse scribble information is utilized where it most impacts gradient flow.
- Anatomy as a Loss Function: The local consistency loss effectively enforces structural rules inherent to human anatomy.
- Regularization Power: The heavy regularization provided by mixup and consistency losses acts as a powerful denoiser, occasionally allowing the model to outperform fully supervised counterparts by preventing overfitting to image artifacts.
Conclusion
CycleMix demonstrates that high-quality medical image segmentation is achievable with minimal expert supervision. While currently 2D-based, future work will explore extending these anatomical priors to 3D volumes. The success of CycleMix suggests that structural regularization and intelligent augmentation can effectively bridge the gap between sparse and full supervision in clinical settings.
Summary updated