Automated Segmentation of Bleeding Regions in Wireless Capsule Endoscopy Using CNN–Transformer Deep Learning Architectures
DOI:
https://doi.org/10.70917/ijcisim-2026-3706Keywords:
Wireless capsule endoscopy, gastrointestinal bleeding, medical image segmentation, convolutional neural network, Vision Transformer, self-attention, boundary-aware learning, explainable artificial intelligenceAbstract
Wireless capsule endoscopy (WCE) is a new method of examining the small bowel using an endoscopic capsule, which incorporates high-speed digital imaging of the internal pathways in a portable system. WCE records an extensive number of images from different angles and light spectra, packed with artifacts that make standard analysis extremely difficult. This article describes the implementation of BleedTrans-Net, a deep learning system that performs bleeding region segmentation at the image level. An example dataset for WCE analysis was created, containing 2,618 WCE images with 1,309 images depicting bleeding and 1,309 images depicting non-bleeding. A dataset split of 70:15:15 resulted in 1,832 training images and 393 images each for validation and testing. BleedTrans-Net was designed with a ConvNeXt-Tiny encoder, 4 multi-head self-attention layers, and gated multi-scale skip fusion. An auxiliary boundary head was trained with boundary loss and other loss variants. The complete IEEE analytical workflow and system was practically demonstrated with the aid of simulated predictions, where in the test set of BleedTrans-Net, a Dice score of 0.914 ± 0.059, an intersection over union score of 0.847 ± 0.097, precision of 0.918, sensitivity of 0.925, specificity of 0.993, and HD95 of 4.37 pixels were achieved. Swin-UNet and BleedTrans-Net were compared for boundary segmentation, with BleedTrans-Net winning by a margin of 0.018 Dice (Holm-adjusted p < 0.001, bootstrap 95% confidence interval: 0.014–0.021). Incremental improvements for boundary segmentation were observed from gated fusion, transformer context, and boundary supervision. Hypothetical segmentation outputs for WCE images were created to show an example of the results and complete the tasks of the statistical reporting pipeline. The segmentation system framework BleedTrans-Net was designed to be functionally complete for segmentation of bleeding images and ready for external testing and validation.
Simulation Disclosure—All quantitative results, plots, confidence intervals, and significance tests in this manuscript are based on a controlled hypothetical dataset and simulated model outputs. They must not be represented as clinically observed or experimentally trained results.
