Multi-Object Segmentation of Coronary Artery OCT Images Based on Multi-Scale Deeping Learning Method

Optical Coherence Tomography (OCT) is one of the most promising new laminar imaging techniques that has been developed rapidly in recent years. To achieve multi-objective segmentation of coronary arteries, a multi-scale and multi-object segmentation method based on the Deeplabv3+ model is proposed. The difficulty of this task is the small size of the targets as well as the size differences and shape similarities between the targets. This paper adds a multi-scale analysis module to the original deeplabv3+ base structure. In the model training, we obtain the pre-training model by training a small number of guidewires with obvious features, and later use the prior knowledge in it to train a larger number of objects with irregularly shaped stent. After experiments, it is shown that the model can achieve better segmentation in coronary OCT images with different object sizes. The accuracy of the model can reach 83.52% and the dice coefficient can reach 0.82. The method in this paper has obvious segmentation advantages for stents and guidewires in coronary arteries with multiple sizes, small objects, and blurred boundaries.