Enhanced Malignancy Prediction of Small Lung Nodules in Different Populations Using Transfer Learning on Low-Dose...

Area of Science:

• Radiology
• Artificial Intelligence
• Medical Imaging

Background:

• Predicting malignancy in small lung nodules (SLNs) is difficult due to demographic and clinical variations.
• Low-dose computed tomography (LDCT) is crucial for lung nodule detection.

Purpose of the Study:

• To investigate transfer learning (TL) for improving SLN malignancy prediction across international datasets.
• To assess TL's effectiveness in overcoming population variations in LDCT data.

Main Methods:

• Collected Asian (CGH) and American (NLST) LDCT datasets.
• Trained initial U-Net models on each dataset.
• Applied TL to transfer model parameters between datasets.
• Evaluated performance using accuracy, sensitivity, specificity, and AUC.

Main Results:

• Significant demographic differences between datasets were confirmed (p < 0.001).
• Initial models showed performance declines (15.2%–97.9%) when applied cross-dataset.
• TL significantly enhanced cross-dataset performance (21.1%–159.5%, p < 0.001), achieving high accuracy (0.86–0.91) and AUC (0.90–0.97).

Conclusions:

• Transfer learning effectively addresses population variations in SLN malignancy prediction.
• TL enables robust application of AI models across diverse international LDCT datasets.
• This approach enhances the reliability of AI-driven lung nodule malignancy assessment.