Objective: To facilitate image-guided surgery in epithelial ovarian cancer (EOC), pre-treatment diagnosis of peritoneal metastases (PM) is essential. However, manual labelling and quantifying all PM lesions is impractical in clinical practice. This st...
Objective: To facilitate image-guided surgery in epithelial ovarian cancer (EOC), pre-treatment diagnosis of peritoneal metastases (PM) is essential. However, manual labelling and quantifying all PM lesions is impractical in clinical practice. This study aimed to develop a deep learning-based auto-segmentation algorithm for PM using computed tomography (CT) scan images of patients with ovarian cancer.
Methods: This retrospective study included 200 patients with histologically confirmed EOC from Seoul National University Hospital for model development and internal validation, and 100 patients from Korea University Anam Hospital for external validation. Three gynecologic oncologists manually annotated PM lesions on contrast-enhanced CT images, which were used to train a 3D nnU-Net–based segmentation model. Segmentation performance was evaluated using standard metrics (Dice coefficient, sensitivity, and precision), as well as additional metrics, including boundary-based metrics (Hausdorff distance and boundary F1 score), and the weighted Dice score. Comparisons were made between the standard threshold (0.5) and gray-level morphological normalization (GMN) threshold settings.
Results: A total of 200 patients with EOC were included for model development and internal validation. The median age was 58.2 years, and the majority were diagnosed at FIGO stage IIIC (49.5%) or IVB (28.0%), with high-grade serous carcinoma being the most common histological subtype (77.5%). An independent external validation cohort comprised 100 patients, including 70 with EOC, 20 with borderline ovarian tumors, and 10 with no malignancy. Among EOC cases, high-grade serous carcinoma accounted for 42.0%, and nearly half were diagnosed at FIGO stage IIIC-IVB. The model achieved higher performance in the external validation cohort compared to the internal cohort. At the GMN threshold, the model achieved improved segmentation performance across both internal and external cohorts. In external validation, the Dice coefficient was 86.96% ± 4.65%, the sensitivity was 92.68% ± 3.16%, and the precision was 82.92% ± 5.51%. Weighted Dice scores remained stable across thresholds, confirming robustness in detecting small-volume lesions. Hausedorff distance (95%) increased slightly with GMN threshold, reflecting broader boundary inclusion while maintaining acceptable contour accuracy.
Conclusions: We successfully developed a deep learning-based auto-segmentation algorithm to identify and indicate PM lesions in ovarian cancer. This model will aid radiologists’ reading and facilitate image-guided surgery for advanced-stage ovarian cancer in clinical practice.