Stenosis Detection of X-Ray Coronary Angiographic Image Sequence

Automatic lesion detection from coronary X-ray angiography images is important for the auxiliary diagnosis of coronary heart diseases. However, the current methods are inefficient, and the detection accuracy cannot meet the criteria of doctors. This paper proposes a two-step method including video phase partition and video stenosis detection to automatically identify coronary stenosis from the complete X-ray angiography (XRA) video. First, convolutional neural network and long short-term memory based spatial-temporal network are used to automatically extract a continuous video segment that is full of contrast agent. Second, a detection network for attention video object is used to accurately and efficiently discern coronary stenosis on the continuous video segment. In the experiment, 166 video data were used for training and testing. The accuracy of video phase partition network can reach 0.838, and the precision and F1 of video stenosis detection results are 0.8 and 0.76 respectively. This performance is the best among all comparison methods. Therefore, we have implemented a complete process for detecting stenoses from coronary XRA sequences.