A Deep Learning Architecture for Stenosis Detection in Invasive Coronary Angiography Images

Coronary artery disease (CAD) is an atherosclerotic narrowing of the coronary artery lumen, that leads to angina or acute myocardial infarction. CAD is a leading cause of death and its prevalence has been increasing rapidly, especially in developing countries. It involves the presence of stenosis when the coronary arteries are narrowed or blocked induced by the atheromatous plaques building up inside, reducing oxygen-rich blood flow to the heart muscle and subsequently resulting in an imbalance state between oxygen demand and supply. Invasive Coronary Angiography (ICA) is considered the reference gold standard imaging technique for the assessment of clinically significant CAD, which enables to reveal the initial CAD symptoms by the morphological features of the coronary arteries such as diameter, length, branching angle, and tortuosity. The challenging task for the interpretation of ICA images depends on complex vessel structure, image noise, poor contrast, and non-uniform illumination. In addition, the manual detection of stenosis is subjective and time-consuming, requiring rich clinical experience. Therefore, developing an ICA-based automatic detection algorithm could improve diagnostic efficiency and could provide huge support to clinicians, reducing bias and subjective interpretations. Several approaches have been proposed in the literature for stenosis detection in ICA images, but these computer-based approaches owe their good results to strong pre-processing techniques. To reduce manual burden in the detection and quantification of coronary stenosis, deep-learning (DL) has been introduced. Therefore, an end-to-end stenosis detection convolutional neural network (CNN) capable of automatically detect stenosis in ICA images is proposed. The CNN for object detection chosen to be trained on the provided dataset is the Single-Shot Multibox Detector (SSD), since it could be adequate for the limited dataset and the number of parameters to train reduced. The approach was validated on 5 models obtained through 5-fold cross-validation, to compensate for the limited dataset dimension and performance comparison of two different network versions, SSD300 and SSD7, is performed. This approach has been shown to be quite good with mean values of Intersection over Union and Dice Similarity Coefficient (DSC) of 0.50 +- 0.06 and 0.64 +- 0.06 for SSD300 respectively and 0.30 +- 0.07 and 0.44 +- 0.08 for SSD7 respectively. Results of this work are good, despite they are slightly lower than values obtained in the literature, when performance metrics are comparable. It is necessary to highlight that the results depend on the annotations, subject to variability, since are drawn manually by operators. Moreover, it is possible to accept values with a certain margin of tolerance, since the aim is to identify coronary stenosis, which is correctly done by visual inspection of the stenosis prediction, except for some cases. Finally, this approach does not apply pre-processing on images, differently from reference works. This work is a good starting point to improve the stenosis detection in ICA images but needs some improvements to be used directly in clinical, minimizing the risk of misinterpretation, and accelerating the decision-making regarding the proper treatment strategy.