Rethinking Barely-Supervised Volumetric Medical Image Segmentation from an Unsupervised Domain Adaptation Perspective

TL;DR

This work tackles barely-supervised volumetric medical image segmentation by reframing it as unsupervised domain adaptation. It introduces BvA, a framework combining Noise-Free Labeled Data Construction (NFC) to synthesize volumetric labels from single-slice annotations and a Frequency and Spatial Mix-Up (FSX) to bridge domain gaps between synthesized and original images, all within a mean-teacher training regime. Empirical results on LA and BraTS demonstrate substantial improvements over state-of-the-art registration-based BSS and SSL methods, including robust performance with as little as 5% labeled data. The findings support that NFC+FSX, along with favoring multiple images with single-slice annotations, provides a practical, scalable approach to high-quality volumetric segmentation under extremely limited supervision.

Abstract

This paper investigates an extremely challenging problem: barely-supervised volumetric medical image segmentation (BSS). A BSS training dataset consists of two parts: 1) a barely-annotated labeled set, where each labeled image contains only a single-slice annotation, and 2) an unlabeled set comprising numerous unlabeled volumetric images. State-of-the-art BSS methods employ a registration-based paradigm, which uses inter-slice image registration to propagate single-slice annotations into volumetric pseudo labels, constructing a completely annotated labeled set, to which a semi-supervised segmentation scheme can be applied. However, the paradigm has a critical limitation: the pseudo-labels generated by image registration are unreliable and noisy. Motivated by this, we propose a new perspective: instead of solving BSS within a semi-supervised learning scheme, this work formulates BSS as an unsupervised domain adaptation problem. To this end, we propose a novel BSS framework, \textbf{B}arely-supervised learning \textbf{via} unsupervised domain \textbf{A}daptation (BvA), as an alternative to the dominant registration paradigm. Specifically, we first design a novel noise-free labeled data construction algorithm (NFC) for slice-to-volume labeled data synthesis. Then, we introduce a frequency and spatial Mix-Up strategy (FSX) to mitigate the domain shifts. Extensive experiments demonstrate that our method provides a promising alternative for BSS. Remarkably, the proposed method, trained on the left atrial segmentation dataset with \textbf{only one} barely-labeled image, achieves a Dice score of 81.20%, outperforming the state-of-the-art by 61.71%. The code is available at https://github.com/Senyh/BvA.

Figures (8)

• Figure 1: Illustration of different learning paradigms. (a) fully supervised learning, (b) semi-supervised learning, (c) weakly-supervised learning, and (d) barely-supervised learning.
• Figure 2: Illustration of (a) the registration-based paradigm, (b) our BvA framework, and (c-d) barely-supervised segmentation results on the left atrial and brain tumor segmentation tasks. Note that DeSCO requires two orthogonal labeled slices per image; therefore, it is unavailable in situations where the training set contains only one labeled slice.
• Figure 3: Overview of the proposed Barely-supervised learning via unsupervised domain Adaptation (BvA). BvA consists of 1) a noise-free labeled data construction algorithm (NFC) for generating volumetric labeled data, and 2) a frequency and spatial Mix-Up strategy (FSX) for alleviating domain shifts between the synthesized images and the original images. Note that BvA only requires the student model in the testing stage.
• Figure 4: Schematic diagram of the proposed Noise-Free Labeled Data Construction (NFC). It synthesizes volumetric image-label pairs using only barely-labeled slices, including three steps: 1) Divide the slice into patches with a sliding window strategy, 2) Stack the patches sequentially into a volume along the depth dimension, and 3) Reshape the volume to match the original image's height and width.
• Figure 5: Detailed procedure of the proposed Frequency and Spatial Mix-Up (FSX). It performs frequency and spatial Mix-Up to generate perturbed images. Note that since the spatial Mix-Up operation is not semantic-preserving, it should be conducted on images and segmentation maps simultaneously.