An Intestinal Tumors Detection Model Based on Feature Distillation With Self-Correction Mechanism and PathGAN

Small intestine tumors are gastrointestinal tumors with unclear clinical manifestations, and the current diagnosis relies on expert analysis of abdominal Computed Tomography (CT) images, resulting in low efficiency. The use of artificial intelligence technology for automatic detection can effectively improve diagnostic efficiency. However, the detection of intestinal tumors is characterized by the difficulty of tumor identification and fuzzy boundary detection. Therefore, we introduce knowledge distillation technologies and propose an intestinal tumor detection model based on feature distillation, which includes a self-correction mechanism and PathGAN. Specifically, a threshold constraint value is set for the loss function of the teacher network during feature distillation. If the loss values exceed the set threshold, then the offset matrix between the output and the labels is calculated to perform deformable convolution on the feature map of the teacher network, to correct the errors; otherwise, PathGAN is used to calculate the distillation loss for achieving regional feature learning in the student network. The proposed feature distillation strategy effectively prevents the error of the teacher network from being transmitted to the student network and assigns different weights to each region, guiding the student network to better learn the features of the foreground regions. Experimental results from our collected dataset on small intestinal stromal tumors indicate that, compared to existing knowledge-based methods, our proposed approach integrating a self-correction mechanism with a regional weight loss function through joint optimization significantly improves detection accuracy while reducing the number of parameters required.