Semi-supervised Medical Image Segmentation Method Based on Cross-pseudo Labeling Leveraging Strong and Weak Data Augmentation Strategies

TL;DR

Medical image segmentation faces high labeling costs and noise in imaging data. The authors propose DFCPS, a FixMatch-inspired semi-supervised framework that uses cross-pseudo supervision and dual strong/weak augmentations to leverage unlabeled data effectively. The model employs two parameter-sharing groups of networks, where pseudo-labels from weak augmentations supervise stronger predictions and cross-group consistency further stabilizes learning, with a combined loss of $L_S$ and $L_{CPS}$. On the Kvasir-SEG dataset, DFCPS outperforms CPC, CPS, ELN, and ACL-Net across labeling ratios, with competitive training and inference times, demonstrating improved robustness and generalization. The work advances semi-supervised medical segmentation by improving pseudo-label quality and leveraging diverse augmentations, and provides publicly available code for reproducibility.

Abstract

Traditional supervised learning methods have historically encountered certain constraints in medical image segmentation due to the challenging collection process, high labeling cost, low signal-to-noise ratio, and complex features characterizing biomedical images. This paper proposes a semi-supervised model, DFCPS, which innovatively incorporates the Fixmatch concept. This significantly enhances the model's performance and generalizability through data augmentation processing, employing varied strategies for unlabeled data. Concurrently, the model design gives appropriate emphasis to the generation, filtration, and refinement processes of pseudo-labels. The novel concept of cross-pseudo-supervision is introduced, integrating consistency learning with self-training. This enables the model to fully leverage pseudo-labels from multiple perspectives, thereby enhancing training diversity. The DFCPS model is compared with both baseline and advanced models using the publicly accessible Kvasir-SEG dataset. Across all four subdivisions containing different proportions of unlabeled data, our model consistently exhibits superior performance. Our source code is available at https://github.com/JustlfC03/DFCPS.

Figures (2)

• Figure 1: The left half is the overall structure of the DFCPS model. This method incorporates a semi-supervised framework design by combining a cross-pseudo-labelling strategy with a strong and weak data enhancement strategy. The right half is the specific structure of $F(\theta_n)$ neural network. The low level feature is the weakly enhanced feature of the unlabeled sample.
• Figure 2: Loss function design of the DFCPS model. The overall architecture of the DFCPS model involves two key loss functions: the supervised loss function $L_S$ and the cross-pseudo-supervised loss function $L_{CPS}$.