Domain-Specific Data Augmentation for Lung Nodule Malignancy Classification.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Oncology

Background:

• Lung cancer is a leading cause of cancer mortality, often diagnosed late.
• Computer-Aided Diagnosis (CAD) systems for lung nodule detection in CT scans can improve early diagnosis.
• Existing deep learning models often lack generalizability to diverse datasets.

Purpose of the Study:

• To address the lack of generalization in deep learning models for lung nodule classification.
• To improve the performance of CAD systems on out-of-domain Computed Tomography (CT) data.
• To develop a robust method for classifying lung nodule malignancy using CT scans.

Main Methods:

• Utilized a ResNet architecture with 2.5D inputs to preserve spatial nodule information.
• Implemented domain-specific data augmentation techniques tailored for CT scans.
• Evaluated model performance on both in-domain (LIDC-IDRI, LNDb) and out-of-domain (LUNGx) datasets.

Main Results:

• The 2.5D input ResNet model combined with data augmentation achieved high Area Under the Curve (AUC) scores.
• Achieved an AUC of 0.914 on internal test data and 0.746 on external test data.
• Significantly improved out-of-domain AUC from 0.576 to 0.695, approaching expert radiologist performance.

Conclusions:

• The proposed method enhances lung nodule malignancy classification accuracy and generalizability across different datasets and hospitals.
• This approach offers a promising automatic tool for lung cancer screening and diagnosis.
• Future work should address uncertainty in nodule annotations based solely on CT observations.