Edge-Aware and Wavelet-Enhanced Detection Transformer for Esophageal Cancer Screening in Barium Meal Images.
Authors
Affiliations (3)
Affiliations (3)
- College of Computer Science, South-Central Minzu University, Key Laboratory of Cyber-Physical Fusion Intelligent Computing, State Ethnic Affairs Commission, Wuhan, 430074, China.
- Renmin Hospital of Wuhan University, Wuhan, 430060, China.
- College of Computer Science, South-Central Minzu University, Key Laboratory of Cyber-Physical Fusion Intelligent Computing, State Ethnic Affairs Commission, Wuhan, 430074, China. [email protected].
Abstract
Early detection of esophageal cancer through non-invasive barium meal imaging is critical for improving patient outcomes, yet the weak lesion boundaries, subtle textures, and high background noise in these images pose fundamental challenges for automated detection. Current methods face two technical bottlenecks: backbone networks lose fragile edge information during downsampling, while feature pyramids introduce detail loss and spatial misalignment during upsampling, together causing high miss rates for superficial lesions and imprecise boundary localization. Here, we present EWE-DETR, a detection framework that addresses these limitations through two complementary innovations. First, an Edge-Aware Partial-Attention backbone (EAP) employs parallel edge and semantic branches to simultaneously capture local boundary details and global anatomical context, preserving lesion edges from the earliest layers. Second, a Wavelet-Enhanced Edge-Semantic Feature Pyramid Network (WES-FPN) uses wavelet transforms for frequency-preserving upsampling and cross-layer edge injection to eliminate traditional upsampling artifacts. Across 10 matched random seeds on the EC-BSID dataset, EWE-DETR achieved an mAP50 of 86.06% ± 0.19% and an F1 score of 83.20% ± 0.60%. Compared with JS-DETR, the closest-performing comparator in terms of mAP50, EWE-DETR yielded higher point estimates across all core detection metrics, including a 7.04-percentage-point improvement in precision. These findings support further evaluation of EWE-DETR as a candidate computer-aided screening tool; however, given the single-center retrospective design of this study, they do not establish its direct deployability in primary-care hospitals or across different imaging devices.