Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Task-oriented Uncertainty Collaborative Learning for Label-Efficient Brain Tumor Segmentation

Zhenxuan Zhang, Hongjie Wu, Jiahao Huang, Baihong Xie, Zhifan Gao, Junxian Du, Pete Lally, Guang Yang

TL;DR

The paper addresses label-efficient brain tumor segmentation from multi-contrast MRI by tackling multi-level specificity across contrasts. It introduces Task-oriented Prompt Attention (TPA) to model cross-contrast and cross-task interactions and Dual-path Uncertainty Refinement (DUR) to iteratively calibrate segmentation centers and boundaries, with a cycle-consistent prompt learning mechanism. The approach achieves state-of-the-art performance on BRATS2021 with limited annotations (e.g., Dice up to $88.2\%$ and HD95 down to $10.853$ mm for 30% labeled data) and shows robust performance even with missing modalities, supported by uncertainty visualizations and improved tumor-volume estimates. These results demonstrate enhanced label efficiency and robustness for clinical brain tumor segmentation and point to potential extensions to other diseases and imaging modalities such as CT and PET.

Abstract

Multi-contrast magnetic resonance imaging (MRI) plays a vital role in brain tumor segmentation and diagnosis by leveraging complementary information from different contrasts. Each contrast highlights specific tumor characteristics, enabling a comprehensive understanding of tumor morphology, edema, and pathological heterogeneity. However, existing methods still face the challenges of multi-level specificity perception across different contrasts, especially with limited annotations. These challenges include data heterogeneity, granularity differences, and interference from redundant information. To address these limitations, we propose a Task-oriented Uncertainty Collaborative Learning (TUCL) framework for multi-contrast MRI segmentation. TUCL introduces a task-oriented prompt attention (TPA) module with intra-prompt and cross-prompt attention mechanisms to dynamically model feature interactions across contrasts and tasks. Additionally, a cyclic process is designed to map the predictions back to the prompt to ensure that the prompts are effectively utilized. In the decoding stage, the TUCL framework proposes a dual-path uncertainty refinement (DUR) strategy which ensures robust segmentation by refining predictions iteratively. Extensive experimental results on limited labeled data demonstrate that TUCL significantly improves segmentation accuracy (88.2\% in Dice and 10.853 mm in HD95). It shows that TUCL has the potential to extract multi-contrast information and reduce the reliance on extensive annotations. The code is available at: https://github.com/Zhenxuan-Zhang/TUCL_BrainSeg.

Task-oriented Uncertainty Collaborative Learning for Label-Efficient Brain Tumor Segmentation

TL;DR

The paper addresses label-efficient brain tumor segmentation from multi-contrast MRI by tackling multi-level specificity across contrasts. It introduces Task-oriented Prompt Attention (TPA) to model cross-contrast and cross-task interactions and Dual-path Uncertainty Refinement (DUR) to iteratively calibrate segmentation centers and boundaries, with a cycle-consistent prompt learning mechanism. The approach achieves state-of-the-art performance on BRATS2021 with limited annotations (e.g., Dice up to and HD95 down to mm for 30% labeled data) and shows robust performance even with missing modalities, supported by uncertainty visualizations and improved tumor-volume estimates. These results demonstrate enhanced label efficiency and robustness for clinical brain tumor segmentation and point to potential extensions to other diseases and imaging modalities such as CT and PET.

Abstract

Multi-contrast magnetic resonance imaging (MRI) plays a vital role in brain tumor segmentation and diagnosis by leveraging complementary information from different contrasts. Each contrast highlights specific tumor characteristics, enabling a comprehensive understanding of tumor morphology, edema, and pathological heterogeneity. However, existing methods still face the challenges of multi-level specificity perception across different contrasts, especially with limited annotations. These challenges include data heterogeneity, granularity differences, and interference from redundant information. To address these limitations, we propose a Task-oriented Uncertainty Collaborative Learning (TUCL) framework for multi-contrast MRI segmentation. TUCL introduces a task-oriented prompt attention (TPA) module with intra-prompt and cross-prompt attention mechanisms to dynamically model feature interactions across contrasts and tasks. Additionally, a cyclic process is designed to map the predictions back to the prompt to ensure that the prompts are effectively utilized. In the decoding stage, the TUCL framework proposes a dual-path uncertainty refinement (DUR) strategy which ensures robust segmentation by refining predictions iteratively. Extensive experimental results on limited labeled data demonstrate that TUCL significantly improves segmentation accuracy (88.2\% in Dice and 10.853 mm in HD95). It shows that TUCL has the potential to extract multi-contrast information and reduce the reliance on extensive annotations. The code is available at: https://github.com/Zhenxuan-Zhang/TUCL_BrainSeg.

Paper Structure

This paper contains 7 sections, 7 equations, 5 figures, 2 tables.

Table of Contents

  1. Introduction
  2. Methods
  3. Task-oriented Prompt Attention Module (TPA)
  4. Dual-path Uncertainty Refinement
  5. Overall Loss Function
  6. Experiment
  7. Conclusion

Figures (5)

  • Figure 1: Motivation of our TUCL framework. (a) Task: Multi-Contrast MRI Analysis. It aims to leverage complementary information from T1, T1ce, T2, and FLAIR contrasts for segmentation and further quantification. (b) Challenge: Multi-Level Specificity. It aims to address the challege of data heterogeneity, granularity differences, and redundant interference to improve analysis accuracy.
  • Figure 2: Workflow of the proposed Task-oriented Uncertainty Collaborative Learning (TUCL) framework. (a) The Task Prompt Attention (TPA) module integrates intra-prompt and cross-prompt attention mechanisms to capture multi-region and multi-contrast features. (b) The model leverages dual-path uncertainty refinement (DUR) to enhance segmentation accuracy. It produces consistent and refined predictions compared to ground truth.
  • Figure 3: Comparison of brain tumor segmentation results with (a) 10% and (b) 30% labeled data. The first five columns show multi-contrast MRI inputs and ground truth, while the rest display segmentation results from different methods.
  • Figure 4: Uncertainty visualization of multi-contrast MRI segmentation. (a) and (b) show the results with 10% and 30% labeled data, respectively. The uncertainty maps highlight regions with higher prediction variability across different contrast removals.
  • Figure 5: Bland-Altman and correlation plots of predicted versus ground truth tumor volumes. (a) The x-axis shows the mean of the predicted and true volumes, while the y-axis shows their difference; dashed lines indicate the mean bias and 95% limits of agreement. (b) The x-axis represents the true volumes, and the y-axis represents the predicted volumes.