A MobileNetV2-Based Boundary-Aware Improved DeepLabv3+ Network for Gastrointestinal Endoscopic Lesion Segmentation

Abstract Objective: Gastrointestinal endoscopic lesion segmentation is challenging because lesions often exhibit substantial variations in morphology and size, blurred boundaries, small foreground regions, and complex mucosal background interference. This study proposed an improved DeepLabv3+ segmentation model that integrates a lightweight backbone, cascaded multi-scale contextual modeling, and boundary-aware auxiliary supervision to improve automatic lesion segmentation in endoscopic images. Methods: DeepLabv3+ was used as the baseline framework. The original feature extraction backbone was replaced with MobileNetV2 to reduce model complexity while extracting both low-level spatial details and high-level semantic features. The ASPP module was redesigned as a cascaded multi-scale atrous convolution module to enable feature reuse among branches with different dilation rates. In the low-level feature pathway, a boundary attention module was introduced to enhance contour-related responses according to a predicted boundary probability map. During training, boundary targets were automatically generated from ground-truth masks, and the model was optimized using a combination of cross-entropy loss, Dice loss, and boundary BCE loss. The dataset included 1,445 gastrointestinal endoscopic images with pixel-level labels, of which 1,300 were used for training and 145 for testing. F1-score, sensitivity, specificity, accuracy, and IoU were used for evaluation. Results: The complete model achieved an F1-score of 0.878, sensitivity of 0.860, specificity of 0.995, accuracy of 0.990, and IoU of 0.784 on the test set. Compared with the original DeepLabv3+ baseline, IoU increased from 0.755 to 0.784. Compared with SegNet, the best-performing competing model, the proposed model improved IoU by 0.040. Ablation experiments demonstrated that both the cascaded ASPP module and the boundary attention branch contributed to performance improvement. Conclusion: The proposed improved DeepLabv3+ model effectively integrates multi-scale semantic information and low-level boundary details, achieving favorable segmentation performance for gastrointestinal endoscopic lesions. This method has potential value for lesion localization, contour delineation, and quantitative analysis in computer-assisted endoscopy.