VDTAR-Net: A Cooperative Dual-Path Convolutional Neural Network–Transformer Network for Robust Highlight Reflection Segmentation

In medical endoscopic imaging, specular reflection (SR) frequently leads to local overexposure, obscuring essential tissue information and complicating computer-aided diagnosis (CAD). Traditional convolutional neural networks (CNNs) face difficulties in modeling global illumination phenomena due to their biased local receptive fields and the inherent “object assumption.” Conversely, pure transformer models often lose high-frequency boundary details and incur substantial computational costs. To tackle these challenges, this paper introduces VDTAR-Net, a specialized framework adapted to address the unique optical characteristics of specular reflections. Building upon hybrid architectures, our contribution focuses on two core mechanisms: (1) a Cross-architecture Fusion Module (CFM) that enables deep, bidirectional information flow, allowing the Transformer’s global illumination modeling to continuously correct the CNN’s local texture biases; and (2) a Reflective-Aware Module (RAM), which explicitly integrates the physical prior of high-intensity saturation into the attention mechanism. This task-specific design significantly enhances sensitivity to boundary details in overexposed regions. We also created the first large-scale, expert-labeled cervical white light segmentation dataset, Cervix-WL-900. High-quality ground truth labels were generated through rigorous double-blind annotation and arbitration by senior experts. Experimental results show that VDTAR-Net achieves a Dice score of 92.56% and a mean Intersection over Union (mIoU) score of 87.31% on Cervix-WL-900, demonstrating superior performance compared to methods like U-Net, DeepLabv3+, SegFormer, and PSPNet. Ablation studies further confirm the substantial contributions of dual-path collaboration, CFM deep fusion, and RAM task-specific priors. VDTAR-Net provides a robust baseline for precise highlight segmentation, laying a foundation for subsequent image quality assessment, restoration, and feature decoupling in diagnostic models.