Finding Holes: Pathologist-Level Performance Using AI for Cribriform Morphology Detection in Prostate Cancer

Area of Science:

• Digital pathology
• Artificial intelligence in oncology
• Histopathology image analysis

Background:

• Cribriform morphology in prostate cancer signifies poor prognosis but is underreported.
• Significant interobserver variability exists among pathologists for its detection.

Purpose of the Study:

• To develop and validate an artificial intelligence (AI) system for improved cribriform pattern detection in prostate cancer.
• Enhance diagnostic reliability and standardize reporting of cribriform morphology.

Main Methods:

• A deep learning model (EfficientNetV2-S encoder with multiple instance learning) was developed for whole-slide classification.
• The model was trained on 640 digitized prostate core needle biopsies and validated internally and externally on independent cohorts.
• Performance was assessed using AUC and Cohen's κ, with comparisons against nine expert uropathologists.

Main Results:

• Strong internal validation (AUC: 0.97, Cohen's κ: 0.81) and robust external validation (AUC: 0.90, Cohen's κ: 0.55) were achieved.
• The AI model demonstrated superior agreement (Cohen's κ: 0.66) compared to individual pathologists (κ: 0.35–0.62).
• Limitations include retrospective design and potential overestimation of performance in certain analyses.

Conclusions:

• The AI model achieved pathologist-level performance in detecting cribriform morphology.
• This AI approach can enhance diagnostic reliability, standardize reporting, and improve treatment decisions for prostate cancer patients.