Adnexal Lesion Discrimination Using Deep Learning Analysis of Dynamic Contrast-enhanced US Images.

Area of Science:

• Radiology
• Artificial Intelligence
• Oncology

Background:

• Adnexal lesions require accurate classification to determine malignancy risk.
• Current diagnostic tools like O-RADS US and the ADNEX model have limitations in performance.
• Deep learning offers potential for improving diagnostic accuracy in medical imaging.

Purpose of the Study:

• To develop and validate a multimodality deep learning model, Ovarian Cancer Network (OCNet), for classifying adnexal lesions.
• To compare the diagnostic performance of OCNet against established methods (O-RADS US and ADNEX model).
• To assess the impact of OCNet assistance on radiologist diagnostic performance.

Main Methods:

• Development of two deep learning models (OCNet_manual and OCNet_automated) using dynamic contrast-enhanced ultrasound images.
• Retrospective study including 395 female patients across 14 hospitals.
• Comparison of OCNet models with O-RADS US and the ADNEX model using area under the receiver operating characteristic curve (AUC).
• Evaluation of radiologist performance with and without OCNet assistance.

Main Results:

• OCNet_manual achieved an AUC of 0.94 and OCNet_automated achieved an AUC of 0.91.
• Both OCNet models significantly outperformed O-RADS US (AUC 0.79) and the ADNEX model (AUC 0.86).
• OCNet assistance improved junior radiologists' average AUC from 0.86 to 0.94 and specificity from 52% to 73%.

Conclusions:

• The OCNet deep learning model demonstrates superior performance in classifying adnexal lesions compared to O-RADS US and the ADNEX model.
• OCNet has the potential to enhance diagnostic accuracy, particularly for less experienced radiologists.
• This AI-driven approach shows promise for improving the management of adnexal lesions.