Image Segmentation of Rectal Tumor Based on Improved U-Net Model with Deep Learning

Rectal tumor is a common malignancy in the intestine. The death rate of rectal tumor ranks fourth among the malignant tumors of digestive system, which seriously threaten the life and health of patients. Endoscopic ultrasonography is the most commonly used method to detect rectal tumors. After obtaining CT images, doctors diagnose the condition with the naked eye and experience, which brings a certain workload to both the doctor and the patient. With the development of in-depth learning and the continuous iterative convolution neural network, more and more techniques have been applied in the field of medical image. Therefore, this paper studies and improves an ultrasonic image segmentation U-Net model for rectal tumors based on fuzzy logic attention mechanism. This paper first preprocesses the original image, enhances the details and reduces the image size.Then the image feature map is weighted by fuzzy logic and attention mechanism. In addition, the loop-back residual mechanism is used to optimize the model. At last, the results of several models are analyzed and compared. The results show that, compared with the U-Net model, the optimized model has a nearly 3% increase in image segmentation precision, almost unchanged recall, and both IoU and Dice have increased by about 2%. Overall, the model has good segmentation performance, and the introduction of RoI aware U-Net greatly reduces the use of video memory.