Small Intestine Tumor CT Image Segmentation Method Based on Self-Supervised Learning

In order to solve the problem of small amount of labeled data and low segmentation accuracy of CT images of small intestinal tumors, a self-supervised learning-based image segmentation model BT-RC-UNet is proposed. Firstly, depth-separable convolution, residual structure, and CBAM attention mechanism are introduced into U-Net to better retain the integrity of the input information and learn the feature representations of small intestinal tumors; secondly, the encoder network is pretrained based on the redundancy-reducing of the Barlow Twins method, the encoder network of the improved U-N et model is pre-trained using the feature representation of CT images of small bowel tumors in an unsupervised manner; finally, the improved U-Net model is fine-tuned using the limited labeled data in the actual segmentation task, so as to extract the more effective features of the original image and refine the image segmentation effect. The experimental results show that the Dice coefficient of the model reaches 87.01%, the Jaccard similarity coefficient and the Hausdorff distance reach 67.60% and 11.2170, respectively, and the indexes such as recall are improved relative to those of the U-Net model, which can effectively realize the accurate segmentation of the lesion region of the CT image of the small intestinal tumor.