TL;DR
This paper introduces DCL-Net, a novel dual contrastive learning framework for semi-supervised multi-organ segmentation in medical images. By combining global and local contrastive learning, DCL-Net effectively leverages unlabeled data to improve segmentation accuracy, addressing the challenge of limited annotated medical datasets.
What Problem Does This Paper Address?
Accurate segmentation of multiple organs in medical images is crucial for computer-aided diagnosis, treatment planning, and survival prediction. However, training deep learning models for this task requires vast amounts of meticulously annotated data, which are expensive and time-consuming to acquire. Semi-supervised learning (SSL) offers a solution by utilizing both labeled and unlabeled data, but existing SSL methods often overlook the inherent relationships between different images and organ categories, limiting their effectiveness. This paper aims to bridge this gap by developing a more sophisticated SSL approach that explicitly models these relationships.
Methods
The proposed DCL-Net operates in two stages:
- Stage I: Similarity-guided Global Contrastive Learning: This stage pre-trains an encoder by exploring the continuity and similarity between slices within a medical volume. It uses a similarity metric based on relative slice positions to construct positive and negative pairs for contrastive learning. Crucially, it re-weights the loss based on similarity, reducing the impact of semantically false positive pairs.
- Stage II: Organ-aware Local Contrastive Learning: This stage builds upon a mean teacher framework. The pre-trained encoder from Stage I initializes both the student and teacher models. An organ-aware local contrastive learning strategy is employed to pull together feature representations of the same organ category (from both labeled and unlabeled data) and push apart those of different categories. To manage computational complexity, a "mask center computation" algorithm is introduced to compress category representations into a single vector per class.
Datasets
The DCL-Net was evaluated on two datasets:
- ACDC (Automated Cardiac Diagnosis Challenge) dataset: A public MRI dataset containing segmentation labels for three organs: left ventricle, right ventricle, and myocardium.
- RC-OARs dataset: An in-house CT dataset with segmentation labels for four organs: bladder, small intestine, right femoral head, and left femoral head.
For semi-supervised training, the datasets were split into labeled and unlabeled subsets, with varying ratios (e.g., 5/75, 10/70, 20/60).
Key Insights and Ideas
- Leveraging Inter-slice Similarity: The paper highlights that the inherent continuity and similarity between adjacent slices in medical volumes can be effectively exploited for contrastive learning, providing a global context.
- Organ-Level Contrastive Learning: Explicitly modeling relationships between organ categories, rather than just individual pixels, is crucial for learning discriminative local features in multi-organ segmentation.
- Computational Efficiency for Local Contrast: The "mask center computation" is a clever way to reduce the quadratic complexity of pixel-wise local contrastive learning, making it more practical for dense prediction tasks.
- Two-Stage Approach Synergy: The combination of global pre-training and local fine-tuning within a mean teacher framework demonstrates a synergistic effect, leading to improved performance, especially in low-data regimes.
Key Results
The DCL-Net demonstrated superior performance on both the ACDC and RC-OARs datasets across various semi-supervised settings.
- ACDC Dataset: In the extreme case of only 5 labeled samples (5/75), DCL-Net significantly outperformed the baseline UNet by 20.3% Dice and 23.32% JI. It consistently achieved the best performance compared to other state-of-the-art methods, with improvements of up to 6.33% Dice when n=5.
- RC-OARs Dataset: DCL-Net achieved the best Dice scores (70.21% and 74.65% for n=5 and n=10, respectively), outperforming the second-best method (RMT-VAT) by 1.27% and 1.96%.
- Ablation Studies: The results confirmed the effectiveness of both the global contrastive learning (Lgcl) and local contrastive learning (Lici) components, with the full DCL-Net model achieving the highest accuracy.
Limitations and Future Work
- Computational Burden: While the mask center computation alleviates some burden, the overall process still involves significant computation, especially for Stage II.
- Hyperparameter Sensitivity: The paper mentions that weighting coefficients (α1, α2, α3) and similarity thresholds (t) are empirically set, suggesting potential sensitivity to these parameters.
- Future Work: The paper's contributions lay the groundwork for further exploration of contrastive learning in medical image segmentation. Future work could involve investigating more advanced contrastive strategies, exploring different architectural designs, or applying the DCL-Net to other medical imaging modalities and segmentation tasks.
Summary updated