A self-supervised classification model for endometrial diseases.

Area of Science:

• Medical Imaging
• Artificial Intelligence in Healthcare
• Gynecological Oncology

Background:

• Ultrasound is key for early endometrial disease diagnosis due to its non-invasive nature and low cost.
• Accurate interpretation of ultrasound images for endometrial lesions heavily relies on radiologist expertise.
• Objective computer-aided diagnostic models are needed to support radiologists in endometrial disease diagnosis.

Purpose of the Study:

• To develop and evaluate a stable and objective computer-aided diagnostic model for endometrial disease classification.
• To assist radiologists in the accurate and efficient diagnosis of endometrial lesions using ultrasound images.

Main Methods:

• A dataset of 1875 transvaginal ultrasound images from 734 patients with endometrial polyps, hyperplasia, and cancer was utilized.
• A self-supervised endometrial disease classification model (BSEM) was proposed, incorporating joint raw and self-supervised tasks.
• The BSEM model employed self-distillation techniques and ensemble strategies for enhanced diagnostic performance.

Main Results:

• The BSEM model achieved satisfactory performance with 75.1% accuracy, 87.3% AUC, 76.5% precision, 73.4% recall, and 74.1% F1 score via fivefold cross-validation.
• Compared to baseline models (ResNet, DenseNet, VGGNet, ConvNeXt, VIT, CMT), BSEM demonstrated superior performance, enhancing key metrics by 3.1-9.0%.
• The model showed significant improvements in accuracy, AUC, precision, recall, and F1 score over established deep learning architectures.

Conclusions:

• The BSEM model serves as a valuable auxiliary diagnostic tool for the early detection of endometrial diseases via ultrasound.
• It enhances the accuracy and efficiency of radiologists in screening for precancerous endometrial lesions.
• The proposed model offers a promising approach to improve diagnostic outcomes in gynecological imaging.