Improving Colorectal Cancer Screening and Risk Assessment through Predictive Modeling on Medical Images and Records.

Area of Science:

• Oncology
• Digital Pathology
• Machine Learning

Background:

• Colonoscopy screening is vital for colorectal cancer (CRC) prevention, but current surveillance relies on histopathology, overlooking other risk factors.
• Pathologist variability in polyp characterization complicates consistent follow-up decisions.
• Digital pathology and deep learning offer new avenues for integrating diverse data to improve CRC risk prediction.

Purpose of the Study:

• To develop and evaluate a deep learning model for predicting 5-year CRC progression risk.
• To explore multi-modal fusion strategies combining histopathology images and clinical records.
• To enhance risk stratification for improved patient surveillance.

Main Methods:

• A transformer-based deep learning model was adapted for histopathology image analysis to predict CRC progression.
• Longitudinal data from the New Hampshire Colonoscopy Registry were utilized.
• Multi-modal fusion strategies integrated deep learning image features with clinical data.

Main Results:

• Predicting intermediate clinical variables improved 5-year progression risk prediction (AUC, 0.630) over direct prediction (AUC, 0.615).
• Integrating whole-slide imaging predictions with nonimaging features yielded superior performance (AUC, 0.672) compared to nonimaging data alone (AUC, 0.666).
• The combined approach significantly outperformed models using only nonimaging clinical data.

Conclusions:

• Integrating diverse data modalities, including digital pathology and clinical records, significantly enhances CRC progression risk stratification.
• Computational methods combined with multi-modal data offer a powerful approach to personalize cancer surveillance.
• This strategy holds promise for more accurate and consistent colorectal cancer risk assessment.