Improving Crohn's disease lesion detection in capsule endoscopy with an advanced feature pyramid network

Capsule endoscopy (CE) is an essential tool for detecting various lesions in the small bowel and plays a crucial role in assessing Crohn's disease (CD). However, manual interpretation of CE images is time-consuming and susceptible to missed detections, primarily due to multi-scale lesions, complex backgrounds, and ambiguous boundaries.

To address these challenges, we propose ISFPE-Net, a novel feature pyramid network for robust multi-scale lesion detection in CE images. The network incorporates two key modules: the Cross-Dual Attention Fusion (CDAF) module, which adaptively integrates multi-scale features to emphasize lesion-relevant regions while suppressing background interference; and the Multi-Scale Convolutional Modulation Partial (MSCMP) module, which captures both global dependencies and local details through progressive fusion and modulation to improve boundary delineation.

Extensive experiments on internal CE datasets demonstrate that ISFPE-Net achieves state-of-the-art performance. Compared to the strong baseline Sparse R-CNN + FPN, it improves average precision (AP) by 1.6% on the test set. It particularly enhances recall for subtle, small, and boundary-blurred lesions in complex backgrounds. Furthermore, the model shows excellent generalization on an external multicenter validation set acquired from different devices, delivering an additional 2.3 percentage point improvement in AP.

ISFPE-Net effectively reduces missed detections in CE-based Crohn's disease lesion detection and has strong potential to alleviate physicians' reading burden. The model's superior performance and generalization ability highlight its promise for broader clinical applicability in small bowel lesion detection.