Deep Learning for Grading Endometrial Cancer.

Area of Science:

• Computational pathology
• Artificial intelligence in oncology
• Digital pathology

Background:

• Endometrial cancer is a prevalent malignancy in women, necessitating precise histologic and molecular classification for optimal treatment.
• Accurate grading of endometrial cancer is crucial for patient management and treatment selection.

Purpose of the Study:

• To introduce EndoNet, a novel deep learning model for classifying endometrial cancer whole-slide images.
• To evaluate EndoNet's performance in distinguishing between high- and low-grade endometrial cancer cases.

Main Methods:

• EndoNet employs convolutional neural networks for feature extraction and a vision transformer for slide classification.
• The model was trained on 929 digitized hematoxylin and eosin-stained endometrial cancer whole-slide images.
• Performance was assessed on internal (110 patients) and external (100 patients) test sets.

Main Results:

• EndoNet achieved a weighted average F1 score of 0.91 and an AUC of 0.95 on the internal test set.
• On the external test set, the model demonstrated an F1 score of 0.86 and an AUC of 0.86.
• The AI model effectively classified slides into low-grade (grades 1-2) and high-grade (grade 3, serous carcinoma, carcinosarcoma) categories.

Conclusions:

• EndoNet shows significant potential as an AI-powered tool to assist pathologists in classifying endometrial cancer grades.
• The model's ability to classify tumors without manual annotations could streamline the diagnostic workflow.
• Further validation is required, but EndoNet may support grading of gynecologic pathology tumors.