Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

A Weakly Supervised and Globally Explainable Learning Framework for Brain Tumor Segmentation

Ruitao Xie, Limai Jiang, Xiaoxi He, Yi Pan, Yunpeng Cai

TL;DR

The paper tackles brain tumor segmentation under weak supervision, addressing the high cost of pixel-level annotations and the demand for explainable AI in medical imaging. It introduces a Class Association Embedding (CAE) framework that combines a symmetrical cyclic GAN with paired random shuffle training to disentangle class-related from identity features, and uses topological data analysis to build a globally explainable class-related manifold. Counterfactual generation along shortest-paths on this manifold enables segmentation without pixel-level labels by comparing abnormal samples to meaningful counterfactual normal references. Empirical results on BraTS2020/2021 show superior IOU and DICE metrics compared with weakly supervised baselines, demonstrating both improved accuracy and interpretability for MRI brain tumor segmentation.

Abstract

Machine-based brain tumor segmentation can help doctors make better diagnoses. However, the complex structure of brain tumors and expensive pixel-level annotations present challenges for automatic tumor segmentation. In this paper, we propose a counterfactual generation framework that not only achieves exceptional brain tumor segmentation performance without the need for pixel-level annotations, but also provides explainability. Our framework effectively separates class-related features from class-unrelated features of the samples, and generate new samples that preserve identity features while altering class attributes by embedding different class-related features. We perform topological data analysis on the extracted class-related features and obtain a globally explainable manifold, and for each abnormal sample to be segmented, a meaningful normal sample could be effectively generated with the guidance of the rule-based paths designed within the manifold for comparison for identifying the tumor regions. We evaluate our proposed method on two datasets, which demonstrates superior performance of brain tumor segmentation. The code is available at https://github.com/xrt11/tumor-segmentation.

A Weakly Supervised and Globally Explainable Learning Framework for Brain Tumor Segmentation

TL;DR

The paper tackles brain tumor segmentation under weak supervision, addressing the high cost of pixel-level annotations and the demand for explainable AI in medical imaging. It introduces a Class Association Embedding (CAE) framework that combines a symmetrical cyclic GAN with paired random shuffle training to disentangle class-related from identity features, and uses topological data analysis to build a globally explainable class-related manifold. Counterfactual generation along shortest-paths on this manifold enables segmentation without pixel-level labels by comparing abnormal samples to meaningful counterfactual normal references. Empirical results on BraTS2020/2021 show superior IOU and DICE metrics compared with weakly supervised baselines, demonstrating both improved accuracy and interpretability for MRI brain tumor segmentation.

Abstract

Machine-based brain tumor segmentation can help doctors make better diagnoses. However, the complex structure of brain tumors and expensive pixel-level annotations present challenges for automatic tumor segmentation. In this paper, we propose a counterfactual generation framework that not only achieves exceptional brain tumor segmentation performance without the need for pixel-level annotations, but also provides explainability. Our framework effectively separates class-related features from class-unrelated features of the samples, and generate new samples that preserve identity features while altering class attributes by embedding different class-related features. We perform topological data analysis on the extracted class-related features and obtain a globally explainable manifold, and for each abnormal sample to be segmented, a meaningful normal sample could be effectively generated with the guidance of the rule-based paths designed within the manifold for comparison for identifying the tumor regions. We evaluate our proposed method on two datasets, which demonstrates superior performance of brain tumor segmentation. The code is available at https://github.com/xrt11/tumor-segmentation.
Paper Structure (12 sections, 6 equations, 4 figures, 2 tables)

This paper contains 12 sections, 6 equations, 4 figures, 2 tables.

Table of Contents

  1. Introduction
  2. Methods
  3. Overall Framework for Brain Tumor Segmentation
  4. Class Association Embedding for Counterfactual Generation
  5. Topology Analysis of CS Codes for Exploring and Explaining Global Rule
  6. Designing Rule-based Paths on Topology Graph for Brain Tumor Segmentation
  7. Experiment
  8. Datasets and Implementation Details
  9. Evaluation Metrics
  10. Results
  11. Ablation Studies
  12. Conclusion

Figures (4)

  • Figure 1: Overall framework of brain tumor segmentation.
  • Figure 2: Symmetric cyclic adversarial network with paired random shuffle training (PRST) scheme, where $X$ represents the sample, while $C$ and $S$ refer to class-style and individual-style codes respectively.
  • Figure 3: Segmentation cases using different algorithms. The regions surrounded by the green lines are the groundtruth, while the regions surrounded by the red lines are the predicted results.
  • Figure 4: The subgraph above is the results of topology analysis (topology graphs) of learned class-related manifolds on BraTS2020 and BraTS2021 datasets, while the subgraph below is some synthetic cases obtained based on the defined paths within the manifolds. In the topology graphs, values with red font inside the nodes refer to the ratios of the abnormal cases involved in these nodes. In the synthetic cases, the IS codes of the EXAMPLE are extracted for combinations with the center vectors of the CS codes involved in each node along the defined paths. The predicted classes by the external classifiers are presented above the synthetic cases.