P3Net: Progressive and Periodic Perturbation for Semi-Supervised Medical Image Segmentation

TL;DR

Semi-supervised medical image segmentation often suffers from over-perturbation and boundary inaccuracies when leveraging unlabeled data. The authors introduce P$^3$Net, featuring a Progressive and Periodic Perturbation Mechanism (P$^3$M) that gradually and periodically adjusts interpolations between labeled and unlabeled data, along with a boundary-focused loss that concentrates learning on edge regions. P$^3$M uses a dynamic interpolation ratio $\alpha_1(iter_1)$ and a periodic reset $\alpha(iter)$, paired with a mixing rule for inputs and pseudo-labels to guide learning, while the boundary loss assigns higher penalties to hard boundary pixels via per-pixel weights derived from a $5\times5$ neighborhood. Empirical results on two 2D and two 3D datasets show state-of-the-art performance, with notable gains at low labeling fractions and successful transfer of the perturbation mechanism to other methods, underscoring the approach's scalability and practical impact for reducing labeling costs in clinical segmentation tasks.

Abstract

Perturbation with diverse unlabeled data has proven beneficial for semi-supervised medical image segmentation (SSMIS). While many works have successfully used various perturbation techniques, a deeper understanding of learning perturbations is needed. Excessive or inappropriate perturbation can have negative effects, so we aim to address two challenges: how to use perturbation mechanisms to guide the learning of unlabeled data through labeled data, and how to ensure accurate predictions in boundary regions. Inspired by human progressive and periodic learning, we propose a progressive and periodic perturbation mechanism (P3M) and a boundary-focused loss. P3M enables dynamic adjustment of perturbations, allowing the model to gradually learn them. Our boundary-focused loss encourages the model to concentrate on boundary regions, enhancing sensitivity to intricate details and ensuring accurate predictions. Experimental results demonstrate that our method achieves state-of-the-art performance on two 2D and 3D datasets. Moreover, P3M is extendable to other methods, and the proposed loss serves as a universal tool for improving existing methods, highlighting the scalability and applicability of our approach.

Figures (7)

• Figure 1: We assessed the learning progress of unlabeled data every 200 iterations. The Dice coefficient served as our metric to evaluate the amount of information acquired. Fixed interpolation representative: BCP (blue). Random interpolation representative: Cutmix(yellow). Non-interpolation method representative: SSNet (green).
• Figure 2: The whole pipeline of our proposed P$^3$Net (left) and progressive and periodic perturbation mechanism (right) in the second stage.
• Figure 3: Visualization of 3D datasets LA with 5% labeled data and Pancreas-NIH with 6% labeled data and ground truth.(best viewed by zoom-in on screen).
• Figure 4: Visualization of 3D datasets LA with 5% labeled data and Pancreas-NIH with 6% labeled data and ground truth.(best viewed by zoom-in on screen).
• Figure 5: Demonstrate a period of different P$^3$M curve functions. (Corresponding to each row in Tab. \ref{['table:temporal curve']}.)