Deep learning-based Segmentation of Multi-site Disease in Ovarian Cancer

PurposeTo determine if pelvic/ovarian and omental lesions of ovarian cancer can be reliably segmented on computed tomography (CT) using fully automated deep learning-based methods. Materials and MethodsA deep learning model for the two most common disease sites of high grade serous ovarian cancer lesions (pelvis/ovaries and omentum) was developed and compared against the well-established "no-new-Net" (nnU-Net) framework and unrevised trainee radiologist segmentations. A total of 451 pre-treatment and post neoadjuvant chemotherapy (NACT) CT scans collected from four different institutions were used for training (n=276), hyper-parameter tuning (n=104) and testing (n=71) of the methods. The performance was evaluated using the Dice similarity coefficient (DSC) and compared using a Wilcoxon test on paired results ResultsOur model outperforms the nnU-Net framework by a significant margin for both disease (validation: p=1x10-4,1.5x10-6, test: p=0.004, 0.005) and it does not perform significantly different from a trainee radiologist for the pelvic/ovarian lesions (p=0.392). On an independent test set (n=71), the model achieves a performance of 72{+/-}19 mean DSC for the pelvic/ovarian and 64{+/-}24 for the omental lesions. ConclusionAutomated ovarian cancer segmentation on CT using deep neural networks is feasible and achieves performance close to a trainee-level radiologist for pelvic/ovarian lesions. SummaryDeep learning-based models were used to assess whether fully automated segmentation is feasible for the main two disease sites in high grade serous ovarian cancer. Key PointsO_LIFirst automated approach for pelvic/ovarian and omental ovarian cancer lesion segmentation on CT images. C_LIO_LIAutomated segmentation of ovarian cancer lesions can be comparable with manual segmentation of trainee radiologists with three years of experience in oncological and gynecological imaging. C_LIO_LICareful hyper-parameter tuning can provide models significantly outperforming strong state-of-the-art baselines. C_LI