Enhancing Medical Image Retrieval Using 3D Reconstructions and CrossNorm

Accurate medical image retrieval plays a pivotal role in clinical decision-making, yet it faces two challenges: the scarcity of annotated data and the inherent inconsistency in medical image styles. To address these issues, we propose a novel Siamese network-based retrieval method that leverages 3D model information to enhance the model's performance. The proposed approach incorporates CrossNorm normalization, improving robustness to noise and unseen corruptions. In addition, the proposed method leverages 3D information to facilitate more effective image comparisons, resulting in improved accuracy in similarity measurements. Experiments on the USenhance2023 and ISIC2019 datasets demonstrate that our method consistently outperforms traditional L₂ normalization, with mAP gains typically exceeding 1%.