Ovarian Cancer Identification on CT Using Deep Learning

Ovarian cancer is the deadliest gynecologic malignancy, with 20,890 new cases and 12,730 deaths expected in the U.S. in 2025. Despite accounting for only 2.1% of female cancers, it causes 4.3% of cancer-related deaths. The lifetime risk of developing ovarian cancer is 1 in 78, with a mortality risk of 1 in 108. Optimal treatment involves complete surgical resection followed by chemotherapy; however, survival rates have remained largely unchanged over the past two decades. Early detection significantly improves survival, but only 31% of cases are diagnosed at an early stage due to the disease's asymptomatic nature. No universal screening test exists, and current methods, including CA-125 and transvaginal ultrasound, have limitations in reducing mortality.

Computed tomography (CT) is commonly used for ovarian cancer detection, but its effectiveness is limited by nonspecific symptoms. AI-driven CT screening has gained interest, with deep learning showing promise in cancer detection. However, challenges remain in ensuring model generalizability and optimizing technical parameters. Effective screening must minimize unnecessary surgeries, as previous trials reported high false-positive rates and surgical complications. To address this, the study developed CAT-OV, an AI-based tool for ovarian cancer detection using CT imaging. The system integrates a Body Part Regression (BPR) model for pelvic localization and a Multiple Instance Learning (MIL) ensemble classifier with five convolutional neural networks (CNNs) to predict cancer presence. CAT-OV was evaluated on three test sets: an internal dataset, an international dataset from the U.S., and a nationwide multi-institutional dataset from Taiwan.

This retrospective study was approved by the institutional review board, and informed consent was waived. The dataset was constructed from CT scans of patients aged ≥20 who underwent ovarian surgery between 2010 and 2020 at Chang Gung Memorial Hospital. Malignant cases included various histopathological subtypes, while controls consisted of benign ovarian tumors and an enriched dataset of cancer-free individuals. The final dataset comprised 5,680 cases, split into a training/validation set (n=4,554) and an internal test set (n=1,126), including 173 cancer and 953 control cases. The international dataset included 40 cancer and 47 control cases from Brigham and Women's Hospital. The nationwide dataset consisted of 447 ovarian cancer cases and 1,131 controls from Taiwan's National Health Insurance database.

The BPR model, modified from ResNet50, localized pelvic regions on CT scans through unsupervised learning. Training involved preprocessing, augmentation, and regression-based subvolume selection. The MIL classification model treated each 3D subvolume as a "bag" of 2D slices, using EfficientNetV2-S as the backbone and an attention-based aggregation module for final prediction. Training involved preprocessing, augmentation, and a five-fold cross-validation strategy. The final ensemble model determined classification based on averaged logits and optimized thresholds. Visualization was performed using the Per-Sample Bottleneck technique to enhance interpretability.

Surgical histopathology served as the reference standard, reviewed by an experienced pathologist using the 2020 WHO classification. Immunohistochemical analysis was conducted to distinguish primary ovarian cancer from metastases. Tumors were staged according to the 8th edition FIGO system. This study aims to improve ovarian cancer detection and screening efficacy through AI-driven CT analysis.

Inclusion criteria

• Age ≥ 20 years old.
• Female
• undergone a CT scan
• undergone a CT scan within 180 days prior to ovarian surgery for histopathological evaluation.

Exclusion criteria

• Age < 20 years old.
• Non-female
• Non-CT imaging
• Incorrect image orientation
• Number of slices < 10
• Slice thickness >10 mm or < 1 mm
• Unsuccessful DICM-to-NIfTI
• Pelvic subvolume extraction failed
• Non-contrast CT scans
• Metallic artifacts
• Inconclusive cases