RECISTSup: Weakly-Supervised Lesion Volume Segmentation Using RECIST Measurement.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computational Pathology

Background:

• Manual lesion volume segmentation is time-consuming and requires expertise.
• Current methods like RECIST offer a coarse but potentially useful surrogate for segmentation.
• Automated segmentation can improve efficiency and consistency in lesion assessment.

Purpose of the Study:

• To develop a novel weakly-supervised method for automatic lesion volume segmentation using RECIST measurements.
• To investigate the feasibility of leveraging RECIST data for precise segmentation.
• To reduce the annotation cost associated with medical image segmentation.

Main Methods:

• Proposed RECISTSup, a weakly-supervised method utilizing RECIST measurements.
• Developed a RECIST measurement propagation algorithm to generate pseudo masks.
• Introduced novel loss functions incorporating spatial prior knowledge from RECIST.
• Implemented iterative model training using automatically segmented results.

Main Results:

• RECISTSup achieved state-of-the-art performance among weakly-supervised methods.
• Demonstrated comparable performance to voxel-level annotation with significantly reduced annotation cost.
• Experiments on three datasets validated the method's effectiveness.

Conclusions:

• RECIST measurements can effectively guide automated lesion volume segmentation.
• The proposed RECISTSup method offers a cost-effective alternative to manual segmentation.
• Weakly-supervised learning with RECIST holds significant promise for medical image analysis.