AI-Driven Pathomics for Predicting Chemotherapy Response in Metastatic Colorectal Cancer: A Transfer Learning...

Area of Science:

• Oncology
• Digital Pathology
• Artificial Intelligence

Background:

• Predicting chemotherapy response in metastatic colorectal cancer (mCRC) is difficult due to a lack of reliable biomarkers.
• Current methods lack the precision needed for effective personalized treatment strategies.

Purpose of the Study:

• To develop and validate an AI-driven pathomics approach for predicting chemotherapy response in mCRC.
• To enhance the interpretability of predictive models using visual explanations.

Main Methods:

• Utilized transfer learning and attention-based multiple instance learning (MIL) on whole slide images (WSIs).
• Pretrained a deep learning algorithm on The Cancer Genome Atlas (TCGA) and fine-tuned it on a multicenter mCRC cohort.
• Integrated multicenter data to improve model stability and reduce variability.

Main Results:

• Achieved a significant improvement in predictive accuracy, increasing the area under the curve (AUC) from 0.54 to 0.68.
• Attention-weighted visualizations highlighted tumor microenvironment features associated with chemotherapy resistance.
• Demonstrated improved model stability and reduced variability compared to a baseline model.

Conclusions:

• AI-driven digital pathology, enhanced by transfer learning, shows feasibility for improving chemotherapy response prediction in mCRC.
• The approach provides valuable insights for clinical decision-making and biological research.
• This framework supports the integration of multi-omics data for AI-powered personalized oncology.