LoRA-fine-tuned Large Vision Models for Automated Assessment of Post-SBRT Lung Injury

Area of Science:

• Medical Imaging and Artificial Intelligence
• Radiotherapy and Oncology
• Computational Efficiency in Deep Learning

Background:

• Stereotactic Body Radiation Therapy (SBRT) is a common lung cancer treatment.
• Radiation-Induced Lung Injury (RILI) is a significant side effect requiring accurate diagnosis.
• Large Vision Models (LVMs) show promise for medical image analysis but require efficient fine-tuning.

Purpose of the Study:

• To evaluate the efficacy of Low-Rank Adaptation (LoRA) for fine-tuning LVMs (DinoV2, SwinV2) for RILI diagnosis.
• To compare LoRA's performance against full fine-tuning and inference-only methods.
• To assess the impact of spatial context (image cropping, 2D to 3D adaptation) on diagnostic accuracy.

Main Methods:

• Fine-tuning of DinoV2 and SwinV2 LVMs using LoRA and full fine-tuning.
• Comparison with inference-only models.
• Utilizing cropped X-ray CT scan images (50 mm³, 75 mm³) centered on the SBRT isocenter.
• Employing various 2D to 3D adaptation techniques for LVMs.

Main Results:

• LoRA achieved performance comparable or superior to traditional full fine-tuning.
• LoRA significantly reduced computational costs and training time due to fewer trainable parameters.
• Model sensitivity to spatial context was evaluated using different cropping strategies and adaptation methods.

Conclusions:

• LoRA presents an efficient and effective method for fine-tuning LVMs for RILI diagnosis.
• This AI-driven approach supports clinical decision-making in lung cancer patients undergoing SBRT.
• The findings highlight the potential of parameter-efficient fine-tuning techniques in medical imaging.