Leveraging Fixed and Dynamic Pseudo-labels for Semi-supervised Medical Image Segmentation

TL;DR

This work tackles the challenge of semi-supervised medical image segmentation with limited annotations. It introduces a co-training framework using two pseudo-labels per unlabeled image: a fixed pseudo-label $\hat{y}^{u_f}$ and a dynamic pseudo-label $\hat{y}^{u_d}$ created by shifting crops and fusing overlapping regions via COM, enhanced by diversity from CutMix. The approach achieves consistent, state-of-the-art results on LA, Pancreas-CT, and Brats-2019 under low-label regimes, with comprehensive ablations validating the contributions of fixed/dynamic losses and diversification. The proposed method reduces reliance on high-confidence single pseudo-labels and improves generalization, offering a practical path toward robust semi-supervised segmentation in clinical settings.

Abstract

Semi-supervised medical image segmentation has gained growing interest due to its ability to utilize unannotated data. The current state-of-the-art methods mostly rely on pseudo-labeling within a co-training framework. These methods depend on a single pseudo-label for training, but these labels are not as accurate as the ground truth of labeled data. Relying solely on one pseudo-label often results in suboptimal results. To this end, we propose a novel approach where multiple pseudo-labels for the same unannotated image are used to learn from the unlabeled data: the conventional fixed pseudo-label and the newly introduced dynamic pseudo-label. By incorporating multiple pseudo-labels for the same unannotated image into the co-training framework, our approach provides a more robust training approach that improves model performance and generalization capabilities. We validate our novel approach on three semi-supervised medical benchmark segmentation datasets, the Left Atrium dataset, the Pancreas-CT dataset, and the Brats-2019 dataset. Our approach significantly outperforms state-of-the-art methods over multiple medical benchmark segmentation datasets with different labeled data ratios. We also present several ablation experiments to demonstrate the effectiveness of various components used in our approach.

Figures (7)

• Figure 1: The process of obtaining shifted cropped images $C = {\{c_{d_1}, c_{d_2}, c_{d_3}, c_{d_4}\}}$. Where $x^{u_f}$ represents the fixed cropped image extracted by random cropping, and $c_{d_1}, c_{d_2}, c_{d_3}, c_{d_4}$ are the cropped images extracted by applying shifting with respect to fixed cropped image $x^{u_f}$.
• Figure 2: An illustration of our fixed and dynamic pseudo-labeling strategy. (a) To learn from labeled data, supervised loss $\mathcal{L}_{sup}^l$ is applied between the predictions of both networks and the ground truth. (b) When subnet A ($\mathcal{SN}_1$) learns from subnet B ($\mathcal{SN}_2$), both subnet receives the original image and $\mathcal{L}_{fix}^u$ is applied between the prediction of $\mathcal{SN}_1$ and fixed pseudo-label of $\mathcal{SN}_2$. Similarly, $\mathcal{L}_{dyn}^u$ is applied between the prediction of $\mathcal{SN}_1$ and dynamic pseudo-label of $\mathcal{SN}_2$. (c) When $\mathcal{SN}_2$ learns from $\mathcal{SN}_1$, $\mathcal{SN}_1$ receives the original image while $\mathcal{SN}_2$ receives the cut-mix image. The same fixed and dynamic loss is then applied between both subnets. (d) $COM(.)$ operation creates a dynamic pseudo-label $\hat{y}^{u_d}$ by taking the non-overlapping region from the fixed pseudo-label $\hat{y}^{u_f}$ and the overlapping region from the temporary pseudo-label $\hat{y}^{u_t}$.
• Figure 3: An illustration of $CUT(.)$ operation. Where $x^{u_{cut}}_p$ and $x^{u_{cut}}_q$ are the images obtained after applying the $CUT(.)$ operation. $x^u_{p}$ and $x^u_{q}$ are two different unlabeled images from a batch.
• Figure 4: Visualization results of different semi-supervised segmentation techniques are depicted, utilizing 20% labeled data alongside ground truth on the pancreas dataset.
• Figure 5: Results obtained by changing $\alpha$ values in the supervised loss function on the pancreas dataset with 20% labeled data.