Intravascular optical coherence tomography (IVOCT) is a crucial micro-resolution imaging modality used to assess the internal structure of blood vessels. Lumen segmentation in IVOCT images is vital for measuring the location and the extent of vessel blockages and for guiding percutaneous coronary intervention. Obtaining such information in real-time is essential, necessitating the use of fast automated algorithms. In this paper, we proposed an innovative polynomial-regression convolutional neural network (CNN) for fast and automated IVOCT lumen segmentation.

The polynomial-regression CNN architecture was uniquely crafted to enable single-pass extraction of lumen borders via IVOCT image regression, ensuring real-time processing efficiency without compromising accuracy. The architecture designed convolution for regression while omitting fully connected layers, leading to the spatial output of lumen representation as polynomial coefficients, thus enabling the formation of interconnected lumen points. The approach equipped the network to comprehend the intricate and continuous geometries and curvatures intrinsic to blood vessels in transverse and longitudinal dimensions. The network was trained on a dataset of 16,165 images and evaluated using 7,016 images.

The predicted segmentations exhibited a distance error of less than 2 pixels (26.40 μm), Dice's coefficient of 0.982, Jaccard Index of 0.966, sensitivity of 0.980, specificity of 0.999, and a prediction time of 4 s (for a pullback containing 360 images). This technique demonstrated significantly improved performance in both accuracy and speed compared to published techniques.

The strong segmentation performance, fast speed, and robustness to image variations highlight the practical clinical utility of the proposed polynomial-regression network.