Mutual Consistency Learning for Semi-supervised Medical Image Segmentation (2109.09960v4)

Abstract: In this paper, we propose a novel mutual consistency network (MC-Net+) to effectively exploit the unlabeled data for semi-supervised medical image segmentation. The MC-Net+ model is motivated by the observation that deep models trained with limited annotations are prone to output highly uncertain and easily mis-classified predictions in the ambiguous regions (e.g., adhesive edges or thin branches) for medical image segmentation. Leveraging these challenging samples can make the semi-supervised segmentation model training more effective. Therefore, our proposed MC-Net+ model consists of two new designs. First, the model contains one shared encoder and multiple slightly different decoders (i.e., using different up-sampling strategies). The statistical discrepancy of multiple decoders' outputs is computed to denote the model's uncertainty, which indicates the unlabeled hard regions. Second, we apply a novel mutual consistency constraint between one decoder's probability output and other decoders' soft pseudo labels. In this way, we minimize the discrepancy of multiple outputs (i.e., the model uncertainty) during training and force the model to generate invariant results in such challenging regions, aiming at regularizing the model training. We compared the segmentation results of our MC-Net+ model with five state-of-the-art semi-supervised approaches on three public medical datasets. Extension experiments with two standard semi-supervised settings demonstrate the superior performance of our model over other methods, which sets a new state of the art for semi-supervised medical image segmentation. Our code is released publicly at https://github.com/ycwu1997/MC-Net.

• The paper introduces the MC-Net+ model with a mutual consistency constraint that robustly improves segmentation in ambiguous regions.
• It employs a multiple decoder architecture for effective uncertainty estimation and efficient use of limited annotated data.
• Experimental results demonstrate significant Dice score improvements on LA, Pancreas-CT, and ACDC datasets, nearly matching fully supervised models.

Mutual Consistency Learning for Semi-supervised Medical Image Segmentation

The paper presents a novel approach to semi-supervised medical image segmentation using a Mutual Consistency Network (MC-Net+). This approach addresses the limitations of conventional deep learning models that struggle with limited annotated data, particularly in ambiguous regions of medical images such as adhesive edges or thin branches. The MC-Net+ model introduces new architectural designs and training strategies to effectively leverage unlabeled data, thereby improving segmentation performance.

Model Architecture and Training Strategy

MC-Net+ leverages the insight that deep models often output uncertain predictions in challenging regions due to limited labeled data. To mitigate this, the model introduces two main innovations:

1. Multiple Decoder Architecture: The MC-Net+ model incorporates one shared encoder and multiple slightly different decoders. These decoders differ in their up-sampling strategies, which allows for model uncertainty estimation based on the statistical discrepancy of their outputs. The epistemic uncertainty is effectively captured by comparing these outputs, identifying hard-to-segment regions.
2. Mutual Consistency Constraint: A novel mutual consistency constraint is applied during training. This involves aligning the probability output of one decoder with the soft pseudo-labels generated by other decoders. The goal is to minimize output discrepancies and achieve consistent and invariant results, particularly in the ambiguous regions, thereby regularizing model training.

Experimental Results

The efficacy of the MC-Net+ model is validated through experiments on three public medical datasets: LA, Pancreas-CT, and ACDC. The model outperforms five state-of-the-art semi-supervised segmentation approaches across these datasets.

• LA Dataset: Employing 10% of labeled data, the MC-Net+ achieves a Dice score improvement from 55% to 70%. With 20% labeled data, it almost matches the fully supervised model trained on 100% labeled data, achieving a Dice score of 91.07% compared to 91.62%.
• Pancreas-CT Dataset: The model demonstrates superior performance in terms of Dice score and Jaccard index, indicating enhanced performance in pancreas segmentation.
• ACDC Dataset: The MC-Net+ model, adapted for 2D multi-class segmentation tasks, exhibits significant gains in average Dice scores across multiple anatomical structures, outperforming state-of-the-art methods in both the 10% and 20% labeled data scenarios.

Implications and Future Developments

The MC-Net+ introduces substantial improvements in semi-supervised medical image segmentation by effectively utilizing unlabeled data. The novel architecture and training regimen offer a scalable solution to the challenges posed by limited labeled data, a common issue in medical image analysis due to the high cost and expertise required for annotating medical images.

Practical implications of this research are profound, as better segmentation models could significantly enhance computer-aided diagnosis systems and clinical decision-making. Theoretically, the approach presents new opportunities to explore mutual consistency constraints in other semi-supervised learning domains.

Future research may focus on expanding the model's applicability to other medical imaging modalities and tasks. Additionally, integrating data-level perturbations, in conjunction with model-level diversity, may offer further improvements. As AI in medical imaging continues to evolve, approaches like MC-Net+ pave the way for robust, efficient, and clinically valuable tools that operate effectively with limited labeled data.