[Application of three-dimensional U-shaped residual coordinated attention network in early detection of small...

Objective: To establish a three-dimensional U-shaped residual coordinated attention network (URCA-Net) based on enhanced CT images for small bowel polyp detection and analyze its application effectiveness in intelligent detection of small bowel polyps. Methods: Abdominal CT data of patients with small bowel polyps were collected from the Air Force Medical Center between June 2019 and July 2023. All patients underwent bowel preparation followed by thin-slice spiral CT scanning to obtain enhanced CT arterial phase images. The data were randomly divided into training, validation and test sets in an 8∶1∶1 ratio. The URCA-Net deep learning model was used for small bowel polyp segmentation. The training set was used for model parameter training, the validation set for hyperparameter adjustment and monitoring of model generalization performance and the test set for final unbiased evaluation of the model. An early intelligent detection model for small bowel polyps was constructed, and its performance was evaluated. Evaluation metrics included pixel-level metrics for the segmentation task [Dice Similarity Coefficient (DSC)], as well as sensitivity and precision for polyp detection. A two-stage segmentation strategy was adopted: the first stage segmented the small bowel region to remove external interference, and the second stage performed polyp segmentation within the small bowel region. Results: A total of 78 subjects were included in the study, with an average age of (54±7) years. A total of 23 400 scan images were extracted, including 136 hyperplastic polyps, 298 hamartomatous polyps, 14 adenomatous polyps, and 4 cancerous polyps. On the test set, the average DSC for the first stage (small bowel segmentation) and the second stage (polyp segmentation) was 0.790 and 0.314, respectively. In the second stage task (polyp segmentation based on small bowel region), the polyp segmentation DSC increased to 0.701, with a precision of 0.836 (95%CI: 0.700-0.972) and a sensitivity of 0.759 (95%CI: 0.631-0.888) for polyp detection. Conclusion: The URCA-Net deep learning technique demonstrates good auxiliary diagnostic effectiveness in small bowel polyp detection and can provide a reference for screening and detection of small bowel polyps. The model is capable of generating high-quality segmentation results, which could facilitate evaluating polyp lesion morphology and provide support for downstream tasks such as preoperative navigation and risk prediction.