基于 IVOCT 的动脉粥样硬化斑块识别与风险评估

Objective The rupture of vulnerable plaques caused by atherosclerosis has become one of the most serious threats to human health. Intravascular optical coherence tomography (IVOCT) can accurately identify vulnerable plaque characteristics, such as thin-cap fibroatheroma plaques, owing to its high resolution, and has gradually become the gold standard for the diagnosis of vulnerable plaques. Typically, clinicians must manually mark the location of plaques in an image based on their experience. However, this method is time-consuming and labor-intensive and is susceptible to the subjective assessment of the clinician. Manual interpretation significantly reduces the speed and precision of vulnerable plaque diagnosis. Some studies based on traditional machine learning have been conducted for the detection of vulnerable plaques and have achieved the classification of single-frame images. However, the accuracy of frame-level information is insufficient to assist clinicians in determining treatment strategies. These methods require a second interpretation by clinicians. This study proposes an evaluation algorithm for vulnerable plaque identification in IVOCT images based on an improved Faster R-CNN (regional convolutional neural network) framework. In addition to accurately locating vulnerable plaques, the algorithm can quantitatively assess the risk of plaque rupture, providing diagnostic suggestions to clinicians and assisting in the formulation of treatment plans. The comprehensive nature of this approach is expected to play an important role in improving the efficiency and precision of vulnerable plaque diagnosis. Methods This study is divided into two parts: automatic identification of vulnerable plaques and assessment of vulnerable plaque rupture risk. To identify vulnerable plaques based on the Faster R-CNN, this study proposes an improved strategy for enhanced cyclic shift data, (X, W) encoding BBox, and the introduction of additional semantic segmentation heads according to the characteristics of IVOCT images. The network is generally divided into four parts (feature extraction, region extraction, secondary detection, and A-scan classification), allowing the network to locate vulnerable plaques with higher accuracy. In this study, the angle of accumulation of the lesion, the thickness of the fibrous cap, macrophage infiltration, superficial microcalcification, and vascular stenosis degree of vulnerable plaques are selected as indicators to assess the risk of rupture. The vascular lumen area is used to characterize the degree of vascular stenosis in vulnerable plaques; the smaller the lumen area, the more severe the stenosis. Furthermore, an adaptive threshold method is designed to calculate the thickness of the fibrous cap, which is considered thin when the thickness is less than 65 μm. The risk of plaque rupture is indicated by a lesion accumulation angle greater than 90°, and a polar graph is used to measure the lesion accumulation angle. To identify superficial microcalcifications and macrophage infiltration, features are extracted from the images and reclassified. The application of these methods makes our study more comprehensive and accurate. Results and Discussions The proposed method is trained and tested using the public dataset CCCV2017 IVOCT. This study presents the results of the ablation experiment for Faster R-CNN (Table 2). The improved network performs well in positioning vulnerable plaques, with mAP50 increasing to 0.744 and the Dice value increasing to 0.905. Compared with weakly supervised detection (WSD) and salient-region-based convolutional neural network (SRCNN) methods, the method proposed in this study significantly improves the recall and Dice values (Table 3). The intersection of union (IOU) value of the lumen area is 0.9445, and the prediction result is consistent with actual result of the lumen area [Fig. 7(c)]. The root mean square error R_MSE and the goodness of fit R² are used to verify the feasibility of the calculation of the thickness of the fiber cap, and the test results are 1.17 pixel and 0.62, respectively. After positioning the region accurately, the cumulative angle of the lesion is also accurately assessed [Fig. 7(d)]. To evaluate the performance of the model in predicting superficial plaque microcalcifications and macrophage infiltration, a comprehensive analysis is performed using a confusion matrix [Figs. 7(e) and (f)]. These results demonstrate that the proposed method achieves satisfactory results for multiple evaluation metrics and provides a reliable solution for the identification of vulnerable plaques and rupture risk assessment. Conclusions In this study, the cyclic shift, (X, W), and encoding BBox are added, and additional semantic segmentation heads are introduced to the Faster R-CNN network to improve the detection performance for vulnerable plaques. Compared to the initial network and adding only a single change, the method proposed in this study significantly improves the mAP50 and Dice values of the network. Compared with WSD and SRCNN, our method also achieves significant improvements in the recall rate and Dice value. Furthermore, to obtain accurate location results for the vulnerable plaque region, the angle of the plaque region is used to measure the angle of accumulation of the lesion, and the cumulative pixels of the fiber cap are used to calculate the thickness of the fiber cap. Deep neural network features combined with gradient direction histogram features are used to analyze macrophage infiltration and superficial microcalcification, and the vascular stenosis degree is evaluated in the lesion lumen region. Multiple single-evaluation results are used to measure the risk of rupture of vulnerable plaques. The comprehensive method proposed in this paper achieves a significant breakthrough in vulnerable plaque detection and provides more comprehensive and reliable data support for clinical diagnosis in terms of rupture risk assessment.