Logo image
Back
Identification of ovarian cancer with AI analysis of CT scans
Abstract   Peer reviewed

Identification of ovarian cancer with AI analysis of CT scans

Katherine Sawaya, Vincent Wagner, Samantha Metzger, Megan McDonald, Michael Goodheart, David Bender and Jesus Gonzalez Bosquet
Gynecologic oncology, Vol.208(Supplement), pp.S39-S40
05/2026
DOI: 10.1016/j.ygyno.2026.01.060

View Online

Abstract

Objectives There is an increasing number of benign cases Gynecologic Oncologists perform each year. There are many helpful tools for triaging benign and malignant masses, but there is no universal triaging method to date. In prelimary studies, we designed prediction models to discriminate benign from malignant pelvic masses in abdominal CTs. This study aims to improve epithelial ovarian cancer (EOC) detection using abdominal-pelvic CT images processed with semi-supervised AI (deep learning) analytics. Methods This was a retrospective case-control study. We included patients with benign pelvic masses (controls, N = 210) and patients with EOC (cases, N = 348) that had abdominal-pelvic CT studies before diagnosis. CT images were digitized, normalized, formatted and processed with two main analytic architectures: Convoluted Neural Network (CNN) and Vision Transformer (ViT) architecture. We used several models based on these architectures. Additionally, we applied novel mixed architectures. Datasets were divided into 60–10–30 ratio for training, validation and testing respectively. Classifcation performance of the resulting models into benign mass or EOC was measured by accuracy. Results Prediction models using novel deep learning semi-supervised methods had an overall high accuracy of predicting EOC in CT images. Methods with CNN-based architecture tend to have a higher accuracy, specifically the 3D CNN, with an accuracy up to 0.94. The performance models based on ViT was lower. For example, ViT for smaller data sets had an accuracy of 0.88. The overall highest performance was observed with the mixed architecture (combination of CNN and ViT) with an accuracy of 0.97. Conclusions Overall, deep learning accuracy of cancer detection in CT is improving with decreased computation time and cost. To date, the number of AI-based CT analysis methods is increasing. There are many models to choose from, but there is no ideal method to be used as the gold standard. Although CNN-based methods were overall more accurate, ViT are faster and have the potential for superior performance if they can take advantage of pre-trained models with large datasets.

Details

Metrics

2 Record Views
Logo image