Deep learning prediction model based on multi-modal transvaginal ultrasound scan images for endometrial cancer.

Abstract

BACKGROUND: Ultrasound-based deep learning (DL) models for the precise diagnosis of endometrial cancer are insufficient. Our aim is to develop and validate an automatic multi-modal ultrasound DL prediction model for the accurate identification of benign versus malignant endometrial diseases. METHODS: The retrospective dataset of patients with endometrial diseases from two hospitals was segregated into an internal set (n = 696) and an external set (n = 78). All patients underwent grayscale, color Doppler, and 3D reconstructed ultrasound scans. We established ResNet-18 DL models using individual sequences (DLgray, DLCDFI, DL3D) and multi-modal sequences (DLfusion), with comparative performance evaluation through area under the receiver operating characteristic curve (AUC) analysis. We compared the best-performing model with 12 radiologists and assessed the diagnostic performance of radiologists with model assistance. RESULTS: The DLfusion model exhibits precise discrimination capabilities for benign and malignant endometrial diseases, achieving an AUC of 0.92. This performance significantly surpasses that of DLgray (0.78, p < 0.01), DLCDFI (0.79, p = 0.01), and assessments by radiologists (0.64, p < 0.001). The DLfusion model is more accurate than radiologists (0.86 vs. 0.78, p < 0.001) at detecting endometrial cancer. With the assistance of the DLfusion model, the average diagnostic accuracy of twelve radiologists significantly improves from 0.77 to 0.85 (p < 0.001). CONCLUSIONS: The DLfusion model based on multi-modal ultrasonic images exhibits a better capacity in diagnosing endometrial cancer compared with radiologists and DL models based on single modal, and significantly improves the diagnostic accuracy of radiologists, offering valuable support for clinical decision-making.