Pulmonary nodule detection in low dose computed tomography using a medical-to-medical transfer learning approach

Area of Science:

• Medical Imaging
• Artificial Intelligence in Healthcare
• Oncology

Background:

• Lung cancer is a leading global cause of death, necessitating effective early detection methods.
• Low-dose computed tomography (LDCT) is the standard for lung cancer screening.
• Existing computer-aided detection (CAD) methods for CT scans are often unsuitable for LDCT due to image quality differences.

Purpose of the Study:

• To develop a semi-automated, transfer learning-based approach for detecting lung nodules in LDCT images.
• To adapt existing CAD models, originally designed for high-dose CT (HDCT), for use with LDCT scans.
• To improve the accuracy and efficiency of lung nodule detection in clinical screening workflows.

Main Methods:

• Utilized the You Only Look Once (YOLO) object detection model for nodule identification.
• Employed a medical-to-medical transfer learning strategy, initializing the YOLO model with weights pre-trained on HDCT data.
• Retrained the model on a dataset of LDCT scans from lung cancer patients and compared its performance against a model initialized with weights from the COCO dataset.

Main Results:

• The model initialized with HDCT pre-trained weights achieved superior performance: precision (0.982 vs. 0.93), specificity (0.923 vs. 0.849), and F1-score (0.924 vs. 0.903) compared to the COCO-initialized model.
• Statistically significant improvements were observed in precision (p=0.0054) and specificity (p=0.00034).
• The HDCT-initialized model demonstrated high performance in detecting progressing nodules, with precision of 1, specificity of 0.92, and sensitivity of 0.930.

Conclusions:

• A semi-automated transfer learning method effectively detects lung nodules in LDCT scans by leveraging HDCT pre-trained weights.
• This approach shows potential to enhance early lung nodule detection in screening programs, reducing overdiagnosis and improving patient outcomes.
• The developed method can aid in timely treatment initiation and potentially lower lung cancer mortality rates.