Radlora: a smart low-rank adaptive approach for radiological image classification

Download PDF
Open Webpage

Yu Zheng, Yuze Gao, Jingren Liu, Ning Yao, Zhong Ji

Published: 2025, Last Modified: 04 Nov 2025Multim. Syst. 2025EveryoneRevisionsBibTeXCC BY-SA 4.0
Abstract: Deep learning has made substantial inroads across numerous domains, particularly in radiological medicine. Parameter-Efficient Fine-Tuning (PEFT) addresses the traditional dependency on deepening models and the limitations of knowledge transfer. Among PEFT techniques, Low-Rank Adaptation (LoRA) (Hu et al. in arXiv:2106.09685, 2021) stands out for its negligible inference latency and high performance. However, LoRA’s uniform rank optimization can result in inefficient resource allocation within the context of medical radiology image analysis, thereby degrading the model’s overall performance. To mitigate these limitations, we introduce a method that integrates architectural enhancements with strategic regularization. By incorporating Gating Units and a novel initialization technique, our model selectively optimizes based on weight importance, achieving precise and efficient resource allocation. Additionally, we employ proximal gradient descent to enhance optimization efficiency, thereby improving training effectiveness within a fixed architectural framework. Notably, to enhance the model’s safety and interpretability, we also conduct a visualization validation, providing the attention maps of the final layer to make the model’s reasoning transparent. Experimental evaluations conducted on six diverse medical imaging datasets demonstrate that RadLoRA achieves significant performance enhancements. Specifically, RadLoRA improves the average accuracy by 2.4% and increases the average AUC by 0.45% compared to LoRA. Moreover, RadLoRA has also achieved certain improvements compared to other methods. In terms of average accuracy, RadLoRA is 7.1% higher than Fine-Tuning, 2.4% higher than Adaptative Low-Rank Adaptation (AdaLoRA) (Zhang et al. in arXiv:2303.10512, 2023), 2.55% higher than Prefix-Tuning (Li and Liang in arXiv:2101.00190, 2021), and 0.63% higher than BitFit (Zaken et al. in arXiv:2106.10199, 2021). These advancements render RadLoRA more suitable for clinical applications, enhancing both efficiency and predictive accuracy.