Boundary-guided VMUNet for atrophic gastritis lesion segmentation in magnifying endoscopy with narrow-band imaging

Atrophic gastritis (AG) is a gastric precancerous condition, and objective delineation of lesion extent may support risk assessment, targeted biopsy, and standardized endoscopic documentation. Magnifying endoscopy with narrow-band imaging (ME-NBI) highlights mucosal and microvascular patterns, but automated segmentation remains difficult because AG lesions often show weak, gradual, and heterogeneous boundaries.

We conducted a single-center retrospective study including 137 ME-NBI images of AG and 87 images of intestinal metaplasia (IM) from 25 subjects. Expert consensus masks were used as reference standards. We developed boundary-guided VMUNet (BG-VMUNet), a weak-boundary-aware segmentation framework combining a five-channel chromatic-textural input representation, visual state-space modeling, coordinate attention, and boundary-guided feature modulation. Model performance was evaluated using five-fold cross-validation and compared with representative segmentation baselines. Dice coefficient, intersection over union (IoU), precision, recall, and 95th percentile Hausdorff distance (HD95) were used to assess region overlap and boundary localization. An extended validation experiment was performed on IM segmentation.

In AG segmentation, BG-VMUNet achieved a Dice coefficient of 0.7228 ± 0.0504, IoU of 0.6079 ± 0.0592, precision of 0.7111 ± 0.0658, recall of 0.8266 ± 0.0265, and HD95 of 94.4665 ± 17.8695 pixels. The five-channel chromatic-textural representation and the boundary-guided architecture contributed separately: the representation provided a general, transferable benefit across backbones, whereas the boundary-guided components added further improvement on top of the representation. In a subject-level paired analysis after aggregating AG validation images within each subject, BG-VMUNet showed statistically significant improvements over the RGB VMUNet baseline in Dice, IoU, and precision after Holm correction, whereas the differences in recall and HD95 were favorable in direction but did not reach significance. In the IM extension task, BG-VMUNet showed numerically higher Dice, IoU, and precision and a lower HD95 than the baseline, with a slightly lower recall.

BG-VMUNet provides a feasible weak-boundary-aware approach for automated AG lesion segmentation in ME-NBI images. By improving region-level delineation under challenging mucosal patterns, this framework may provide a technical basis for future quantitative assessment of gastric precancerous conditions, subject to multicenter external validation and explicit linkage to clinical endpoints. These findings are preliminary and exploratory.