Pathologist Validation of a Machine Learning-Derived Feature for Colon Cancer Risk Stratification

Area of Science:

• Oncology
• Computational Pathology
• Digital Pathology

Background:

• Identifying novel prognostic markers in colon cancer is crucial for refining histopathologic review and enhancing patient management.
• While prognostic artificial intelligence (AI) systems show promise in cancer risk stratification, the interpretability and clinical usability of their derived features by pathologists remain underexplored.

Purpose of the Study:

• To assess if pathologist scoring of a machine learning-identified histopathologic feature, tumor adipose feature (TAF), correlates with survival outcomes in colon cancer patients.
• To validate the prognostic significance of TAF as a feature learnable and scorable by human experts.

Main Methods:

• A prognostic study involving 258 colon adenocarcinoma cases from 2013-2015.
• Two pathologists performed semiquantitative scoring of tumor adipose feature (TAF) on histologic slides.
• TAF was previously identified using a prognostic deep learning model on an independent cohort.

Main Results:

• Tumor adipose feature (TAF) was identified in 120 of 258 cases.
• TAF demonstrated independent prognostic value for overall and disease-specific survival (HR for presence: 1.55, P=.02; HR for widespread TAF: 1.87, P=.004).
• High interpathologist agreement (90%, κ=0.69) was achieved for TAF scoring.

Conclusions:

• Pathologists can learn to reproducibly score TAF, a machine learning-derived feature, enabling significant risk stratification in colon cancer.
• This study represents the first validation of human expert learning from machine learning in pathology.
• TAF holds potential for integration into routine pathology practice as a human-identifiable prognostic marker.