Annular Prior Prompt Learning for Medical Images Segmentation

Area of Science:

• Medical Image Analysis
• Computer Vision
• Artificial Intelligence

Background:

• Medical image segmentation algorithms struggle with annular objects due to high intra-class variability and inter-class similarity.
• This often leads to regional confusion, particularly in medical imaging where annular regions are common.

Purpose of the Study:

• To develop a novel method, Annular Prior Prompt Learning (APPL), to enhance segmentation of annular objects in medical images.
• To address challenges posed by high intra-class variability and inter-class similarity in these regions.

Main Methods:

• Proposed an Annular Prior Prompt Encoder (APPE) utilizing an annular constraint to standardize prompt features.
• Introduced a Region Connectivity Enhanced Image Encoder (RCEIE) incorporating morphology attention to reduce feature noise and improve region connectivity.
• Developed a mask decoder that leverages collaborative annular features to avoid inter-class confusion and handle weak boundaries.

Main Results:

• Achieved 90.07% mIoU and 94.71% DSC on an in-house dataset.
• Demonstrated strong generalization on public heart segmentation datasets with high end-diastolic and end-systolic DSC scores.
• Consistently superior quantitative and qualitative performance compared to state-of-the-art methods.

Conclusions:

• The APPL method effectively improves medical image segmentation for annular objects.
• The proposed APPE and RCEIE components successfully mitigate challenges from feature variability and similarity.
• APPL offers a robust solution for accurate segmentation of complex anatomical structures.