A frozen ViT with 0.17% trainable LoRA parameters and a unified 2D/3D embedding layer matches or slightly beats fully fine-tuned ViTs across 12 medical diagnosis datasets.
, author Mintun, E
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.CV 1years
2025 1verdicts
CONDITIONAL 1representative citing papers
citing papers explorer
-
UniCAD: Efficient and Extendable Architecture for Multi-Task Computer-Aided Diagnosis System
A frozen ViT with 0.17% trainable LoRA parameters and a unified 2D/3D embedding layer matches or slightly beats fully fine-tuned ViTs across 12 medical diagnosis datasets.