Detection of thin‐cap fibroatheroma in IVOCT images based on weakly supervised learning and domain knowledge. Issue 5 (2nd February 2023)

• Record Type: Journal Article
• Authors: Shi, Peiwen; Xin, Jingmin; Wu, Jiayi; Deng, Yangyang; Cai, Zhuotong; Du, Shaoyi; Zheng, Nanning
• Is Part Of: Journal of biophotonics. Volume 16:Issue 5(2023)
• Page Start: n/a
• Page End: n/a
• Publication Date: 2023-02-02
• Subjects: deep learning -- optical coherence tomography -- the thin‐cap fibroatheroma -- weakly supervised learning; Photonics -- Periodicals; Optical materials -- Periodicals; Optics -- Periodicals; Medical instruments and apparatus -- Periodicals; 621.3605
• DOI: 10.1002/jbio.202200343
• Languages: English
• ISSNs: 1864-063X

Abstract

Abstract: Automatic detection of thin‐cap fibroatheroma (TCFA) on intravascular optical coherence tomography images is essential for the prevention of acute coronary syndrome. However, existing methods need to mark the exact location of TCFAs on each frame as supervision, which is extremely time‐consuming and expensive. Hence, a new weakly supervised framework is proposed to detect TCFAs using only image‐level tags as supervision. The framework comprises cut, feature extraction, relation, and detection modules. First, based on prior knowledge, a cut module was designed to generate a small number of specific region proposals. Then, to learn global information, a relation module was designed to learn the spatial adjacency and order relationships at the feature level, and an attention‐based strategy was introduced in the detection module to effectively aggregate the classification results of region proposals as the image‐level predicted score. The results demonstrate that the proposed method surpassed the state‐of‐the‐art weakly supervised detection methods. Abstract : Intravascular optical coherence tomography is the only imaging technique that allows the identification of thin‐cap fibroatheroma (TCFA). In this study, a weakly supervised method was proposed to detect TCFAs with only image‐level tags, thus significantly reducing the burden of annotating data by physicians. Furthermore, it may enable physicians to label more weakly annotated images, thereby leveraging deep learning capabilities to build more accurate models, which is useful for clinical and research applications.