SO011/#834  Development of deep learning-based auto-segmentation algorithms for peritoneal metastases using computed tomography image analysis of ovarian cancer

• Published in: Focused Plenary 02: Surgery Vol. 33; no. Suppl 4; p. A18
• Main Authors: Kim, Se Ik, Lee, Maria, Kim, Jae-Weon, Chung, Hyun Hoon
• Format: Journal Article
• Language: English
• Published: Kidlington BMJ Publishing Group Ltd 01.11.2023; Elsevier Inc; Elsevier Limited
• Subjects: Clinical medicine; Datasets; Deep learning; Focused Plenary 02: Surgery; Metastasis; Ovarian cancer; Tomography
• ISSN: 1048-891X; 1525-1438
• DOI: 10.1136/ijgc-2023-IGCS.25

IntroductionTo facilitate image-guided surgery in ovarian cancer, pre-treatment diagnosis of peritoneal metastases (PM) is essential. However, manual labeling and quantifying the whole PM lesions are impractical in clinical practice. Thus, we aimed to develop a deep learning-based auto-segmentation algorithm for PM using computed tomography (CT) scan images of newly diagnosed epithelial ovarian cancer.MethodsWe retrospectively collected pre-treatment CT scan images from patients with epithelial ovarian cancer who were treated at our institutional hospital. Patients were randomly assigned to training, development, and test sets with 8:1:1 ratio, and underwent 5-fold cross validation. The whole PM lesions in the abdominal-pelvic cavity of the training dataset were manually drawn by one radiologist. They also referred to surgical records and descriptions of PM lesions. 3D nnU-Net was selected as the deep-learning architecture. One radiologist manually drew the whole PM lesions in the abdominal-pelvic cavity in the test dataset twice and submitted them as references for validation.ResultsMean age at initial diagnosis was 58.2 years, and 95.5% of the study population had FIGO stage IIIB-IVB diseases. Complete resection was achieved in 57.5% of the patients. The final model was validated using corresponding test dataset, and yielded the average Dice similarity coefficient (DSC), sensitivity, and precision as 83.1%, 83.1%, and 83.9%, respectively, across all folds.Abstract SO011/#834 Figure 1Conclusion/ImplicationsWe 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.