Abstract
Semi-supervised learning (SSL) uses unlabeled data to compensate for the scarcity of annotated images and the lack of method generalization to unseen domains, two usual problems in medical segmentation tasks. In this work, we propose POPCORN, a novel method combining consistency regularization and pseudo-labeling designed for image segmentation. The proposed framework uses high-level regularization to constrain our segmentation model to use similar latent features for images with similar segmentations. POPCORN estimates a proximity graph to select data from easiest ones to more difficult ones, in order to ensure accurate pseudo-labeling and to limit confirmation bias. Applied to multiple sclerosis lesion segmentation, our method demonstrates competitive results compared to other state-of-the-art SSL strategies.
Introduction
In medical imaging, the development of robust deep learning models is often hindered by a significant bottleneck: the scarcity of expert-labeled data. While raw MRI scans are abundant, high-quality annotations for tasks like Multiple Sclerosis (MS) lesion segmentation require significant time and expertise from neuroradiologists. Furthermore, models trained on limited datasets often suffer from poor generalization when faced with domain shifts—variations in scanner manufacturers, field strengths, or imaging protocols.
Semi-supervised learning (SSL) offers a solution by leveraging large volumes of unlabeled data. However, traditional SSL approaches like Pseudo-Labeling (PL) often fall victim to "confirmation bias," where the model reinforces its own incorrect predictions. Consistency Regularization (CR) methods, while effective, often focus on low-level image perturbations that may not capture the complex anatomical features necessary for precise segmentation. This paper introduces POPCORN, a framework designed to mitigate these issues through a progressive, curriculum-based approach.
Methodology
The POPCORN framework (Progressive Pseudo-labeling with COnsistency Regularization and Neighboring) integrates three primary innovations to improve the reliability of semi-supervised segmentation.
1. Bottleneck Consistency Regularization
Unlike standard consistency methods that operate on the final output, POPCORN applies a consistency loss at the network's bottleneck—the most compressed latent representation in a U-Net architecture. This high-level regularization forces the model to map images with similar segmentations to similar latent feature representations, ensuring that the internal logic of the model remains consistent across different augmentations.
2. Proximity Graph Selection
To combat confirmation bias, POPCORN employs a "neighboring" strategy. The model calculates Euclidean distances between the latent representations of labeled and unlabeled samples. By constructing a proximity graph, the framework identifies unlabeled images that are most similar to the known labeled distribution. These "easy" samples are prioritized for labeling, as the model's predictions on them are statistically more likely to be accurate.
3. Progressive Pseudo-labeling
POPCORN adopts a curriculum learning strategy. Rather than labeling the entire unlabeled dataset at once, the model iteratively selects the $K$ most similar neighbors, generates pseudo-labels, adds them to the training set, and retrains. This progressive integration allows the model to build a solid foundation on certain data before attempting to segment more ambiguous or difficult cases.
Experimental Results
The framework was evaluated using a small "seed" set of 21 longitudinal images from the ISBI dataset and a massive unlabeled pool of 2,901 FLAIR MRI scans. Testing was conducted on a diverse dataset of 43 subjects to assess generalization across unseen scanners.
Quantitative Performance
POPCORN achieved a state-of-the-art Dice score of 73.09%, significantly outperforming the supervised baseline (64.41%) and other established SSL methods such as Uncertainty-guided Pseudo-labeling and Multi-task SSL. Ablation studies confirmed that both the proximity graph and the consistency regularization were vital; removing the proximity graph led to a 5% drop in Dice performance.
Qualitative Analysis
Visually, the method demonstrated a superior balance between sensitivity and precision. While competing methods often suffered from over-segmentation (high false positives) or missed small lesions entirely (low sensitivity) when encountering domain shifts, POPCORN maintained robust segmentation boundaries across varying image qualities and scanner types.
Conclusion
POPCORN represents a significant step forward in semi-supervised medical image segmentation. By combining latent-space consistency with a progressive proximity-based selection process, the framework effectively mitigates confirmation bias and improves generalization. While the iterative nature of the method increases computational costs, the gains in accuracy and robustness make it a compelling choice for clinical applications where labeled data is scarce. Future work will focus on automating hyperparameter selection and testing the framework's efficacy on other medical imaging modalities.
Summary updated