SHTA Semantic Hard Token Correction and Center Alignment for Semi-Supervised Medical Image Segmentation

Jul 16, 2026

Upload by Skyler

Semi-supervised medical image segmentation often improves supervision quality via prediction consistency, pseudo-label learning, and hard-region guidance, yet fails to explicitly enforce semantic consistency in learned representations of hard regions. This can leave difficult areas with unstable semantic assignments, restricting further gains. SHTA introduces a lightweight training-time semantic representation branch that corrects intermediate semantic representations through semantic assignment, hard token refinement, and semantic center alignment, without adding inference cost. Integrated into GA-CPS, CPS, URPC, and MagicNet and evaluated on Synapse and AMOS, it improves accuracy, weak-organ recovery, and semantic ambiguity reduction.

LanguageEnglish

CategoryResearch & Report

FormatPDF

Pages9

Reading Time23 min

Introduction

Semi-supervised medical image segmentation and annotation challenges

Limitations of prediction-level hard-region supervision

Proposed solution: SHTA and its motivation

Semantic assignment, hard token refinement, and center alignment

SHTA: Semantic Hard Token Correction and Center Alignment for Semi-Supervised Medical Image Segmentation Zhuoru Zhang2 , Yiheng Zhong 1 , Zimu Zhang2 , Xiaofeng Liu 1,† 1Yale University, United States 2Xi’an Jiaotong-Liverpool University, China Emails: [xiaofeng.liu@yale.edu](xiaofeng.liu@yale.edu) arXiv :2607 .070 19v 1 [ cs .CV] 8 Jul 2026 Abstract—Recent advances in semi-supervised medical image segmentation have achieved remarkable performance through prediction consistency, pseudo-label supervision, and hard-region supervision. However, these methods primarily improve supervision quality rather than explicitly enforcing semantic consistency in the learned representations of hard regions. Consequently, even under increasingly stronger prediction-level supervision, difficult regions exhibiting unstable semantic assignment often fail to establish semantically consistent representations during training, thereby limiting further segmentation improvement. To address this issue, we propose SHTA (Semantic Hard Token Correction and Center Alignment), a lightweight trainingtime semantic representation branch. Instead of introducing additional prediction supervision, SHTA refines intermediate semantic representations through Semantic Assignment, Hard Token Refinement, and Semantic Center Alignment, thereby improving semantic consistency in hard regions while preserving the original prediction pathway and introducing no additional inference cost. We integrate SHTA into representative semisupervised segmentation frameworks, including GA-CPS, CPS, URPC, and MagicNet, and conduct evaluations on the Synapse and AMOS datasets. Experimental results demonstrate that SHTA delivers consistent paired improvements across frameworks, with especially clear gains in segmentation accuracy, weak-organ recovery, and semantic ambiguity reduction, while incurring only training-time overhead. The code is available at [https://anonymous.4open.science/r/release_SHTA-42D5/](https://anonymous.4open.science/r/release_SHTA-42D5/). Index Terms—Semi-Supervised Medical Image Segmentation, Hard Token Correction, Semantic Consistency, Representation Alignment, Semantic Representation Learning I. INTRODUCTION Anatomical organ segmentation is a fundamental task in medical image analysis and plays an important role in diagnosis assistance, treatment planning, and quantitative clinical assessment. In 3D abdominal CT segmentation, models are required to delineate multiple organs from volumetric scans, where accurate voxel-wise predictions are essential for reliable clinical interpretation. Although deep learning has achieved remarkable progress in medical image segmentation [1]–[4], dense voxel-level annotation remains expensive and timeconsuming, especially when multiple anatomical structures must be labeled by trained experts. †Corresponding authors. Fig. 1: Conceptual comparison between prediction-level SSL and SHTA. (a) Conventional SSL selects prediction-level hard evidence. (b) SHTA performs representation-level assignment, correction, and alignment. (c) Corrected hardregion tokens form a more stable semantic space. Semi-supervised medical image segmentation alleviates this annotation burden by learning from a small set of labeled volumes together with abundant unlabeled data. Existing SSL methods mainly improve how unlabeled data are used for supervision. Prediction-consistency methods enforce agreement across teacher–student networks, dual branches, or perturbed views [5]–[8], while reliable-supervision methods refine pseudo labels, mine hard regions, estimate uncertainty, or select reliable samples [9]–[15] . These strategies are effective for deciding which predictions, pseudo labels, samples, or regions should supervise training. However, as summarized in Fig. 1(a), they still mainly operate at the prediction or regionselection level; after a hard region is selected, the semantic organization of its intermediate token representation

FAQs

What problem does SHTA address in semi-supervised medical image segmentation?

SHTA targets post-selection ambiguity where selected hard regions may still have unstable token-to-class assignments and semantically inconsistent representations, limiting segmentation improvement.

How does SHTA improve semantic consistency without adding inference cost?

SHTA adds a lightweight training-time semantic representation branch that refines intermediate semantic representations using Semantic Assignment, Hard Token Refinement, and Semantic Center Alignment while preserving the original prediction pathway.

Which frameworks and datasets were used to evaluate SHTA?

SHTA was integrated into GA-CPS, CPS, URPC, and MagicNet and evaluated on the Synapse and AMOS datasets.

Jul 16, 2026

Upload by Skyler