Hybrid Computer Vision Model to Predict Lung Cancer in Diverse Populations

Area of Science:

• Medical Imaging
• Artificial Intelligence in Healthcare
• Public Health

Background:

• Lung cancer incidence is disproportionately high in Black populations.
• Current lung cancer screening criteria may not adequately serve Black individuals due to elevated risk.
• Individualized risk prediction models integrating clinical and imaging data offer a potential solution to mitigate these disparities.

Purpose of the Study:

• To develop and evaluate deep and machine learning models for lung cancer risk prediction.
• To compare the performance of clinical models versus hybrid computer vision models (clinical + CT imaging features).
• To assess the impact of data inclusivity on model generalizability and racial disparities.

Main Methods:

• A cross-sectional study using data from the National Lung Screening Trial (NLST) and University of Illinois Health (UIH) cohorts.
• Inclusion criteria based on age and tobacco use for individuals at risk of lung cancer.
• Development and testing of deep learning and machine learning models, including hybrid computer vision approaches.

Main Results:

• Optimized clinical models showed moderate predictive performance (ROC-AUC 0.60-0.67).
• Hybrid computer vision models significantly improved performance in the NLST cohort (ROC-AUC 0.78-0.91).
• Performance of hybrid models deteriorated in the UIH cohort, particularly for Black participants (ROC-AUC 0.63-0.72), but improved upon retraining with inclusive data (ROC-AUC 0.70-0.87).

Conclusions:

• Hybrid computer vision models enhance lung cancer risk prediction accuracy compared to clinical models alone.
• Biases in image training data can limit the generalizability of these models in specific populations, such as Black individuals.
• Inclusive training and validation datasets are crucial for minimizing racial disparities and improving the clinical utility of AI-driven risk prediction models for lung cancer.