How Many Private Data Are Needed for Deep Learning in Lung Nodule Detection on CT Scans? A Retrospective Multicenter...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Oncology

Background:

• Early lung nodule detection is vital for lung cancer prevention but faces challenges due to limited radiologist availability and time constraints.
• Automated lung nodule detection using deep learning is being explored to address these limitations.
• Deep learning models require substantial datasets, posing a challenge for initial research phases.

Purpose of the Study:

• To investigate optimized data collection strategies for deep learning-based lung nodule detection.
• To evaluate the performance of deep learning models trained on smaller, curated datasets compared to larger public datasets.
• To demonstrate the utility of transfer learning in data-scarce scenarios for lung nodule detection.

Main Methods:

• Collected chest CT scans from 515 patients with lung nodules from three hospitals, including radiologist-verified annotations.
• Utilized the YOLOX object detection model for lung nodule detection experiments.
• Compared model performance using the collected dataset against the publicly available LUNA16 dataset.

Main Results:

• The YOLOX model trained on the collected dataset achieved similar or better performance than when trained on the larger LUNA16 dataset.
• Weight transfer learning from pre-trained open data significantly improved performance in data-limited conditions.
• Effective performance was observed with datasets from over 100 patients.

Conclusions:

• Optimized data collection and curation are effective for deep learning-based lung nodule detection.
• Transfer learning is a valuable technique when large datasets are difficult to acquire.
• This study provides guidance for efficient data collection in future lung nodule detection research.