RECIST-Induced Reliable Learning: Geometry-Driven Label Propagation for Universal Lesion Segmentation.
IEEE Transactions on Medical Imaging
|July 12, 2023
Summary
This study introduces a weakly-supervised framework for automatic universal lesion segmentation (ULS) in CT images, leveraging existing RECIST annotations. The novel approach improves segmentation accuracy by using RECIST-induced geometric labeling and soft label propagation.
Area of Science:
- Medical Imaging
- Artificial Intelligence
- Computer Vision
Background:
- Accurate lesion segmentation in Computed Tomography (CT) is crucial for cancer assessment, but current methods like Response Evaluation Criteria In Solid Tumors (RECIST) are limited.
- The lack of large-scale, pixel-wise annotated datasets hinders the development of automatic universal lesion segmentation (ULS) models.
- Existing weakly-supervised methods often rely on shallow interactive segmentation, which can introduce noise and limit generalization.
Purpose of the Study:
- To develop a weakly-supervised learning framework for automatic universal lesion segmentation (ULS) in CT images.
- To utilize existing lesion databases with RECIST annotations for ULS training, overcoming data scarcity.
- To improve segmentation accuracy and generalization compared to current RECIST-based methods.
Main Methods:
- Proposed a unified RECIST-induced reliable learning (RiRL) framework for weakly-supervised ULS.
- Introduced RECIST-induced geometric labeling to generate preliminary labels based on clinical RECIST characteristics, creating a trimap (foreground, background, unclear regions).
- Implemented an on-the-fly soft label propagation strategy using a topological knowledge-driven graph to optimize segmentation boundaries and avoid noisy training.
Main Results:
- The RiRL framework demonstrated superior performance on a public benchmark dataset compared to state-of-the-art RECIST-based ULS methods.
- Achieved significant improvements in Dice scores across various backbones (ResNet101, ResNet50, HRNet, ResNest50), surpassing existing approaches by 1.4% to 2.0%.
- The proposed label generation and propagation strategies effectively mitigated issues of noisy training and poor generalization.
Conclusions:
- The developed weakly-supervised framework effectively leverages RECIST annotations for robust and accurate universal lesion segmentation in CT images.
- The RiRL framework offers a promising solution for ULS, addressing the challenge of limited pixel-wise annotated data.
- This approach has the potential to enhance radiological assessment and streamline cancer treatment evaluation.
