Deep learning model for classifying endometrial lesions.

Area of Science:

• Gynecologic Oncology
• Medical Imaging
• Artificial Intelligence

Background:

• Hysteroscopy is vital for diagnosing endometrial lesions, but subjective clinician judgments necessitate objective diagnostic aids.
• Accurate diagnosis is crucial as different endometrial lesions require distinct treatments.

Purpose of the Study:

• To develop and evaluate an automated convolutional neural network (CNN) model for classifying endometrial lesions using hysteroscopic images.
• To provide objective diagnostic evidence to assist clinicians in hysteroscopy.

Main Methods:

• A VGGNet-16 model was trained on 6478 preprocessed hysteroscopic images of various endometrial lesions.
• The model was tested on 250 images and its performance compared against diagnoses made by gynecologists.

Main Results:

• The VGGNet-16 model achieved an overall accuracy of 80.8% in classifying five types of endometrial lesions.
• For differentiating benign from premalignant/malignant lesions, the model demonstrated 90.8% accuracy, with 83.0% sensitivity and 96.0% specificity.
• The model's diagnostic performance slightly surpassed that of three experienced gynecologists.

Conclusions:

• The VGGNet-16 model effectively classifies endometrial lesions from hysteroscopic images.
• This AI tool offers objective diagnostic support, potentially enhancing clinical decision-making and improving diagnostic accuracy for hysteroscopists.