Abstract 62P
Background
Peritoneal carcinomatosis (PC) is frequently observed in newly diagnosed ovarian cancer and is associated with a poor prognosis. An accurate diagnosis of PC and indication of all lesions on pre-treatment imaging studies is very necessary. Thus, we aimed to develop deep learning-based auto-segmentation algorithms to diagnose and indicate PC lesions in ovarian cancer using computed tomography (CT) scan images.
Methods
We retrospectively collected pre-treatment CT scan images from patients with epithelial ovarian cancer who received primary treatment, consisting of cytoreductive surgery and postoperative adjuvant chemotherapy, between 2010 and 1019 at a tertiary institutional hospital. Patients were randomly assigned to training and test sets with 9:1 ratio. The tumors and PC lesions in the abdominal-pelvic cavity of the training set were manually drawn by two radiologists, who were experts in gynecologic malignancies, in the delayed phase. 3D nnU-Net was selected as the deep-learning architecture. Using the validation set, one radiologist manually drew lesions of interest twice, and the developed deep learning-based auto-segmentation algorithm was performed.
Results
The training and validation sets included 180 and 20 patients, respectively. The median age of patients in the validation set was 58.6 years; 86.1% and 85.0% of them had ascites and FIGO stage IIIC-IVB diseases, respectively, and 56.1% achieved complete cytoreduction. The model was validated using a validation set. Its predictive performance yielded a Dice similarity coefficient, sensitivity, and precision of 93.2%, 92.4%, and 94.1%, respectively.
Conclusions
We successfully developed deep learning-based auto-segmentation algorithms to identify and indicate PC lesions in ovarian cancer. This model can aid radiologists’ reading and facilitate image-guided surgery for advanced-stage ovarian cancer in clinical practice.
Legal entity responsible for the study
S.I. Kim.
Funding
Has not received any funding.
Disclosure
All authors have declared no conflicts of interest.