Segmentation of Coronary Artery Stenosis in X-ray Angiography using Mamba Models

TL;DR

This work tackles automated segmentation of coronary artery stenosis in X-ray angiography, a challenging problem due to noisy, blurred vessels and imbalanced data. It systematically evaluates five Mamba-based U-Net variants and a Swin Transformer-based model for stenosis localization on the ARCADE/XCA dataset, including encoders that place Mamba blocks in various network regions. The U-Mamba BOT variant achieves the highest F1 score of 0.6879 (68.79%), beating a semi-supervised baseline by 11.8%, with LightM-UNet offering the smallest parameter count. The findings highlight the efficacy of selective state-space models, particularly SS2D Visual State Space blocks, for efficient 2D context modeling in medical image segmentation and suggest practical utility in CAD assessment workflows.

Abstract

Coronary artery disease stands as one of the primary contributors to global mortality rates. The automated identification of coronary artery stenosis from X-ray images plays a critical role in the diagnostic process for coronary heart disease. This task is challenging due to the complex structure of coronary arteries, intrinsic noise in X-ray images, and the fact that stenotic coronary arteries appear narrow and blurred in X-ray angiographies. This study employs five different variants of the Mamba-based model and one variant of the Swin Transformer-based model, primarily based on the U-Net architecture, for the localization of stenosis in Coronary artery disease. Our best results showed an F1 score of 68.79% for the U-Mamba BOT model, representing an 11.8% improvement over the semi-supervised approach.

Figures (5)

• Figure 1: The Mamba architecturegu2024mambalineartimesequencemodeling
• Figure 2: The structure of Mamba and VSS blocksliu2024vmamba
• Figure 3: Methodology used in this study. The input images and corresponding masks are fed into various model architectures, and the predictions for each model are displayed in front of the respective model. (a) The U-Mamba architecture ma2024u, (b) The Swin UNetR architecture hatamizadeh2021swin, (c) The LightM-UNet architecture liao2024lightm, (d) The Swin U-Mamba architecture liu2024swin. (a_1) U-Mamba BOT prediction, (a_2) U-Mamba ENC prediction, (d_1) Swin U-Mamba prediction, (d_2) Swin U-Mamba D prediction.
• Figure 4: Visualization of each model's performance on the ARCADE dataset popov2024dataset. The Y-axis of this chart shows their performance you have to check the central point of each circle for it and circle diameters represent the number of parameters used in each of these models.
• Figure 5: Prediction of different models on test dataset.