Retrieval-Augmented Vision Transformer Boosts Mammogram Classification Accuracy on Two Datasets

ViT-MultiRAGNet, a retrieval-augmented Vision Transformer for mammogram classification, achieved AUC 0.998 (±0.009) on the RTM dataset, outperforming unimodal benchmarks (p<0.01).

• Primary result: ViT-MultiRAGNet achieved AUC 0.998 and accuracy 0.978 on the RTM dataset, with AUC 0.989 and accuracy 0.961 on CBIS-DDSM (five-fold cross-validation, p<0.01).
• The retrieval-augmented generation module improved fusion over simple concatenation, and segmentation Dice scores were 0.882 (RTM) and 0.795 (CBIS-DDSM).
• Inference speed was 0.31 seconds per image, supporting potential clinical use.

Automated summary

RadPigeon summaries are original and for information only. They are not clinical advice.