OC05.03: *Deep learning‐enabled ovarian cancer detection with ADNEX‐AI: a prospective, multicentre study

Résumé fourni par la source

Traditional ultrasound (US)-based tools to diagnose ovarian cancer, including ADNEX, require manual identification and measurement of tumour features. We introduce a deep learning (DL)-based ADNEX-AI framework for the automated detection of ADNEX variables in US scans, assess its performance and compare it with current clinical practice. Interim analysis of IOTA7-AI, an international multicentre prospective study, including patients with adnexal masses from 2019 to 2023. Eligible patients were examined by transvaginal US and selected for surgery. 2D US variables and images were collected at inclusion, and histological outcome after removal of the mass. Two cohorts were created: a training cohort for developing a multi-task convolutional neural network to segment four US ADNEX features (lesion, locules, solid tissue, and papillations), and a validation cohort to assess the two-stage cascade combining the segmentation model and ADNEX without CA-125. For training, expert sonographers (level 3 EFSUMB), blinded-to-histology, provided pixelwise feature delineations serving as ground-truth. We evaluated the discrimination between benign and malignant masses using ROC curves and AUCs. The training and validation cohorts consist of 861 US images from 385 patients and 5074 US images from 655 patients, respectively. ADNEX-AI achieves an AUC of 0.91 (95% CI: 0.87-0.94), compared to 0.93 (0.90-0.96) achieved by ADNEX applied manually during the US exam. The AUC using pixelwise delineations on the training cohort was 0.90 (0.85-0.95). Our DL pipeline effectively distinguishes benign from malignant ovarian tumours on US scans. The AUC difference between ADNEX-AI and traditional ADNEX may arise from the clinician's ability to examine the tumour from various angles, versus ADNEX-AI's reliance on a limited set of 2D slices. As such, when compared to ADNEX using pixelwise delineations under similar constraints of tumour visibility, ADNEX-AI demonstrated comparable performance.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.