Endometrial Pipelle Biopsy Computer-Aided Diagnosis: A Feasibility Study

Area of Science:

• Pathology
• Oncology
• Medical Artificial Intelligence

Background:

• Endometrial biopsies are crucial for diagnosing abnormal uterine bleeding and hereditary endometrial cancer risk.
• Approximately 10% of biopsies reveal malignancy, posing a significant workload for pathologists evaluating mostly benign cases.

Purpose of the Study:

• To assess the feasibility of an AI-assisted diagnostic tool for endometrial biopsies.
• To evaluate AI performance on whole-slide images from daily practice, not just selected cases.

Main Methods:

• An AI algorithm was trained on 2819 endometrial biopsies.
• The AI's classification of 91 biopsies was compared against 15 pathologists' assessments.
• Biopsies were categorized into 6 clinical groups: nonrepresentative, normal, nonneoplastic, hyperplasia without atypia, hyperplasia with atypia, and malignant.

Main Results:

• Pathologist agreement was moderate (Cohen's kappa=0.51) for 6 categories and substantial (Cohen's kappa=0.66) for binary benign vs. (pre)malignant classification.
• The AI achieved moderate agreement (Cohen's kappa=0.43) for 6 categories and substantial agreement (Cohen's kappa=0.65) for binary classification.
• AI performance was comparable to pathologists in the binary benign vs. (pre)malignant distinction.

Conclusions:

• AI-assisted diagnosis is feasible for differentiating benign from (pre)malignant endometrial tissues using unselected cases.
• Endometrial premalignancies present challenges for both AI and human pathologists.
• Further refinement is necessary for a comprehensive AI diagnostic solution for endometrial biopsies.