Learning From Target-Level Incomplete Annotation: A Novel Perspective for Weakly-Supervised Multi-Lesion...

Area of Science:

• Medical Imaging
• Computer Vision
• Machine Learning

Background:

• Accurate segmentation of lesions in whole-body computed tomography (CT) scans is vital for automated diagnosis and treatment planning.
• Training automated segmentation models requires extensive pixel-wise labeled data, which is costly and time-consuming.
• Current weakly-supervised methods often fail to accurately segment lesion boundaries.

Purpose of the Study:

• To propose a novel annotation strategy, target-level incomplete annotation (TIA), to minimize manual annotation effort.
• To develop a multi-lesion segmentation framework that effectively utilizes incomplete annotations for medical image segmentation.
• To achieve state-of-the-art performance in medical image segmentation using the proposed TIA and framework.

Main Methods:

• Introduced target-level incomplete annotation (TIA), annotating only one complete target region per slice.
• Developed a multi-lesion segmentation framework incorporating a medical cut-paste branch and a prior-assisted target localization branch.
• Employed a graph neural network (GNN) for noisy label correction and reliable pixel propagation.

Main Results:

• The proposed TIA strategy significantly reduces annotation effort while maintaining boundary accuracy.
• The multi-lesion segmentation framework, utilizing TIA, achieved state-of-the-art performance.
• Validation on the Crohn's dataset demonstrated the framework's effectiveness in medical image segmentation.

Conclusions:

• Target-level incomplete annotation (TIA) offers an efficient alternative for medical image annotation.
• The proposed weakly-supervised segmentation framework effectively leverages TIA for accurate lesion segmentation.
• This approach holds significant promise for improving automated diagnosis and treatment planning in medical imaging.