Low-Rank Adaptation of Pre-Trained Large Vision Models for Improved Lung Nodule Malignancy Classification.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Large vision models (LVMs) pretrained with self-supervised learning (SSL) show promise for medical image analysis.
• Adapting these models for specific tasks like lung nodule malignancy classification is crucial but challenging.
• Traditional fine-tuning methods can be computationally expensive and require large datasets.

Purpose of the Study:

• To investigate the efficacy of Low-Rank Adaptation (LoRA) for adapting SSL-pretrained LVMs for lung nodule malignancy classification.
• To evaluate LoRA's impact on classification performance, parameter efficiency, and training speed compared to traditional fine-tuning.
• To explore the potential of LoRA in improving lung cancer diagnostic accuracy.

Main Methods:

• Utilized two large lung nodule datasets (NLSTx and LIDC) with CT scans.
• Applied LoRA technique to adapt SSL-pretrained LVMs.
• Compared LoRA-adapted models against traditional fine-tuning approaches.

Main Results:

• The best LoRA-adapted model achieved a 3% increase in ROC AUC compared to the state-of-the-art.
• LoRA utilized 89.9% fewer parameters than traditional fine-tuning.
• Training time was reduced by 36.5% using LoRA.

Conclusions:

• LoRA is a highly effective method for adapting out-of-domain pretrained LVMs for lung nodule malignancy classification.
• LoRA offers significant improvements in performance, parameter efficiency, and training speed.
• This approach presents a promising direction for advancing lung cancer diagnostics through AI.