Training Set Augmentation and Harmonization Enables Radiomic Models to Detect Early Onset of Lung Cancer

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Oncology

Background:

• Radiomics-based machine learning shows promise for early lung cancer detection using CT scans.
• Challenges include low malignancy rates in early pulmonary nodules and variable image acquisition.
• Existing models struggle with clinical applicability due to these limitations.

Purpose of the Study:

• To develop clinically applicable radiomics models for early lung cancer detection.
• To address challenges of low malignancy rates and image acquisition variability.
• To investigate the impact of data harmonization and augmentation strategies.

Main Methods:

• Trained machine learning models on radiomic features from early-development pulmonary nodules (PNs).
• Augmented training data with later-development PNs and harmonized for acquisition effects using ComBat.
• Compared harmonization strategies: without biological distinction, with a covariate, and separate harmonization.

Main Results:

• Initial models trained on early-development PNs showed near-chance performance.
• Augmented models harmonized with a covariate achieved an ROC-AUC of 0.72 (0.67-0.76).
• Augmented models harmonized separately achieved an ROC-AUC of 0.69 (0.63-0.74), both significantly improved.

Conclusions:

• Data harmonization and augmentation are crucial for effective radiomics models in early lung cancer detection.
• Accounting for biological differences during harmonization significantly improves model performance.
• Findings support the development of radiomics tools for lung cancer early detection using routine screening CT imaging.