A learnable transformer decoder for blurred near-infrared blood vessel segmentation using domain adaptation with limited data

Le résumé fourni par la source

Transillumination imaging using near-infrared (NIR) light is an effective method for visualizing subcutaneous blood vessels. However, as the vessel depth increases, images become severely blurred due to strong light scattering in body tissue. Although deep learning has shown great promise in clear vessel segmentation, many models face challenges achieving accurate segmentation on blurred NIR blood vessels, due to limited training data and the non-learnability of bilinear upsampling. To address these challenges, in this study, a new decoder model named the DeMerge Transformer (DMTrans) is introduced. This model uses a new learnable upsampling approach that can adaptively enlarge useful information from the encoder pretrained in related fields, maintaining accurate results with limited data during fine-tuning. The DMTrans model includes both a DeMerging (DM) operation and a channel transformer mechanism. The DM operation also leverages channel information, which becomes more sufficient as the model layers deepen, to initialize spatial features. The transformer reconstructs the DM features to fuse global information and long-range dependencies, facilitating a dense, learnable upsampling process. The proposed method was evaluated on five blood vessel datasets (DRIVE, STARE, CHUAC, CHASE_DB1, and HRF) and an NIR transillumination image dataset (HV_NIR). The proposed method achieved an average 1.89% improvement in the Dice score across these benchmarks. The results demonstrate that the DMTrans model not only offers stronger capabilities for transferring prior knowledge but also proves to be an effective approach for datasets that are small in scale but require high precision.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.