PolarFormer: A Transformer-Based Method for Multi-Lesion Segmentation in Intravascular OCT.

Area of Science:

• Cardiovascular Imaging
• Artificial Intelligence in Medicine
• Medical Image Analysis

Background:

• Deep learning methods exist for single-class vulnerable plaque segmentation in intravascular optical coherence tomography (OCT).
• Current research is hindered by a lack of large-scale, multi-class annotated OCT datasets.
• Segmentation challenges include irregular plaque distribution, unique shapes, and fuzzy boundaries, with existing methods neglecting geometric and spatial prior information.

Purpose of the Study:

• To address the limitations of existing methods and datasets for multi-class vulnerable plaque segmentation in OCT images.
• To develop a novel deep learning model incorporating spatial prior knowledge for improved segmentation accuracy.
• To introduce a new, publicly available dataset for advancing research in this field.

Main Methods:

• Collected a new dataset comprising 70 pullback intravascular OCT data with multi-class annotations.
• Developed PolarFormer, a deep learning model integrating spatial distribution prior knowledge of vulnerable plaques.
• Introduced Polar Attention as a key module to model spatial relationships in the radial direction.

Main Results:

• The proposed PolarFormer model demonstrated superior performance compared to existing baseline methods on the new dataset.
• The model effectively addresses challenges related to irregular plaque geometry and fuzzy boundaries.
• The introduced dataset and model provide a valuable resource for future research in vulnerable plaque segmentation.

Conclusions:

• PolarFormer offers a significant advancement in multi-class vulnerable plaque segmentation from OCT images.
• Incorporating spatial prior knowledge, particularly through Polar Attention, enhances segmentation accuracy.
• The availability of the new dataset and code facilitates further development and validation of deep learning models in cardiovascular imaging.