OCT Medical Image Recognition Based on UNet++

In this paper, we propose an improved UNet++ medical image segmentation method for coronary vascular OCT medical image detection to address the problem of large recognition errors arising from small targets and the limitation of arithmetic power in practical use when identifying stents and guidewire. Two layers of VGG feature extraction layer are used to extract the original image and input it into UNet++ model for cooperative training of multi-layer UNet model. Cross entropy loss and dice loss were used for evaluation, mixing the weight of the output of the multi-layer network model by a given strategy. The multi-objective semantic segmentation with high accuracy can be achieved while reducing the arithmetic power consumption of the model, Meanwhile, according to the requirements of different arithmetic power and recognition speed under the deployment of the model, the UNet model embedded in the model can be removed by pruning operation to reduce the arithmetic power consumption.