FiAt-Net: Detecting Fibroatheroma Plaque Cap in 3D Intravascular OCT Images

TL;DR

A new deep learning based approach, called FiAt-Net, for detecting angular extent of fibroatheroma (FA) and segmenting its cap in 3D intravascular optical coherence tomography (IVOCT) images is reported, demonstrating its effectiveness in accurately detecting FA cap in IVOCT images.

Abstract

The key manifestation of coronary artery disease (CAD) is development of fibroatheromatous plaque, the cap of which may rupture and subsequently lead to coronary artery blocking and heart attack. As such, quantitative analysis of coronary plaque, its plaque cap, and consequently the cap's likelihood to rupture are of critical importance when assessing a risk of cardiovascular events. This paper reports a new deep learning based approach, called FiAt-Net, for detecting angular extent of fibroatheroma (FA) and segmenting its cap in 3D intravascular optical coherence tomography (IVOCT) images. IVOCT 2D image frames are first associated with distinct clusters and data from each cluster are used for model training. As plaque is typically focal and thus unevenly distributed, a binary partitioning method is employed to identify FA plaque areas to focus on to mitigate the data imbalance issue. Additional image representations (called auxiliary images) are generated to capture IVOCT intensity changes to help distinguish FA and non-FA areas on the coronary wall. Information in varying scales is derived from the original IVOCT and auxiliary images, and a multi-head self-attention mechanism is employed to fuse such information. Our FiAt-Net achieved high performance on a 3D IVOCT coronary image dataset, demonstrating its effectiveness in accurately detecting FA cap in IVOCT images.

Figures (12)

• Figure 1: Illustrating the angular coverage of a fibro-atheromatous plaque cap (FA-cap angle) in a 2D IVOCT frame. The red curve marks the inner border between the lumen and vessel wall; the blue curve marks part of the outer border between the vessel wall and periadventitial tissues; the green angular range shows the FA radial segment.
• Figure 2: Illustrating our pre-processing. (a) An original OCT frame in the Cartesian domain; (b) the detected luminal (red line) surface; (c) the resulted frame of the pre-processing. We first employ the pre-processing step in zahnd2017contour to detect the lumen border (the red curve in panel (b)). The boundary between the RoI area and background area (all zero) is detected following the Cartesian-Polar conversion. Each row is subsequently shifted so that the luminal border forms a straight vertical line (the left boundary) in the polar-coordinate frame (panel (c)). This step removes catheter artifacts and blood remnants in the lumen area. Next, we use the row that has the longest lumen-background distance (red-to-blue) as the base and right-pad (with 0's) the other rows whose lumen-background distances are shorter than the longest distance. Finally, we resize each image to the size of $360\times 128$, ensuring that each image has the same size and the original pixel density is not affected.
• Figure 3: Illustrating the frame clustering process. An auto-encoder is used to extract latent features of each frame, and these latent features are used for frame clustering.
• Figure 4: Illustrating FA annotation in (a) the Cartesian domain and (b) the polar domain. The green areas represent FA ranges.
• Figure 5: Illustrating the FiAt-Net. (a) The overall process. For an input frame and each of its three auxiliary images, we extract features at different levels (in this example, marked by red, orange, and blue, respectively) of its BPT (binary partition tree). Next, we aggregate features of the same level from all the four BPTs. Finally, the aggregated features of each level are used to generate the output and compute the loss. (b) The process on one BPT. For one input frame or an auxiliary image, a root-to-leaf path is randomly selected; let its sequence of frame and sub-regions be, e.g., $S=(I_0, I_1, I_2)$. Our model takes $S$ as input, uses a multi-head self-attention mechanism to integrate their feature maps $f_0, f_1$, and $f_2$ at different levels, and outputs refined feature maps $f'_0, f'_1$, and $f'_2$. We only illustrate a path of three levels for an original frame for simplicity.