Towards radiologist-level cancer risk assessment in CT lung screening using deep learning

Area of Science:

• Medical Imaging and Diagnostics
• Artificial Intelligence in Healthcare
• Oncology and Cancer Research

Background:

• Lung cancer is the leading cause of cancer mortality in the US.
• Low-dose computed tomography (CT) screening significantly reduces lung cancer mortality.
• Deep learning (DL) models show promise for lung cancer risk assessment but require robust validation on large, diverse datasets for clinical adoption.

Purpose of the Study:

• To investigate a deep learning framework for lung cancer risk assessment on large, heterogeneous datasets.
• To ensure strong model generalization and stability for clinical adoption.
• To compare the DL framework's performance against state-of-the-art models and existing risk assessment tools.

Main Methods:

• Utilized three low-dose CT lung cancer screening datasets: National Lung Screening Trial (NLST), Lahey Hospital and Medical Center (LHMC), and Kaggle competition data.
• Employed a two-stage framework: a nodule detector followed by a neural network using nodule image context and features to estimate malignancy risk.
• Trained the algorithm on a portion of the NLST dataset and validated it on independent datasets, ensuring no patient overlap between training and validation sets.

Main Results:

• The DL model demonstrated strong generalization across datasets, achieving Area Under the Curve (AUC) scores between 86% and 94%.
• Outperformed the PanCan Risk Model by 6-9% AUC and surpassed state-of-the-art models from the Kaggle Data Science Bowl 2017.
• Achieved performance comparable to radiologists in patient-level cancer risk estimation.

Conclusions:

• The developed deep learning framework shows robust performance and generalization capabilities for lung cancer risk assessment using low-dose CT.
• The model's effectiveness on large, external datasets supports its potential for clinical adoption in lung cancer screening programs.
• This approach offers a promising advancement over existing risk models and approaches human-level performance in identifying cancer risk.