Quantifying the incremental value of deep learning: Application to lung nodule detection.

Area of Science:

• Oncology
• Artificial Intelligence
• Medical Imaging

Background:

• Machine learning (ML) models show promise in prediction tasks but their implementation and value assessment are in early development.
• Existing methods often compare ML models directly to traditional approaches, lacking a framework to quantify added value.
• This study introduces an incremental value framework to evaluate ML model performance in lung cancer research.

Observation:

• A novel ML model was developed to identify and classify lung nodules using data from The Cancer Imaging Archive (TCIA).
• The model's incremental value was assessed by integrating it with existing lung cancer risk prediction models.
• External datasets were utilized for robust validation of the ML model's performance.

Findings:

• The ML image model enhanced existing risk prediction models that relied solely on patient characteristics.
• The model's improvement over prediction models incorporating nodule features was inconclusive.
• Variable performance across different validation datasets highlighted challenges in achieving population generalization with ML models.

Implications:

• This work provides a framework for evaluating the true incremental value of ML models in clinical research.
• The findings suggest that while ML can augment traditional models, careful validation is needed for broad applicability.
• Further research is warranted to address generalization challenges and optimize ML model performance in diverse populations for lung cancer detection.