A pioneering artificial intelligence tool to predict treatment outcomes in ovarian cancer via diagnostic laparoscopy.

Ovarian cancer is associated with high rates of patient mortality and morbidity. Laparoscopic assessment of tumor localization can be used for treatment planning in newly diagnosed high-grade serous ovarian carcinoma (HGSOC). While spread to multiple intra-abdominal areas is correlated with worse outcomes, whether other morphological tumor differences are also associated with patient outcomes is unknown. Given the large volume of visual information in laparoscopic videos, we investigated whether deep-learning models can capture implicit features and predict treatment outcomes. We developed a novel deep-learning framework using pre-treatment laparoscopic images to assess clinical outcomes following upfront standard treatment, defined as short progression-free survival (PFS) (< 8 months) or long PFS (> 12 months). The deep-learning framework consisted of contrastive pre-training to capture morphological features of images and a location-aware transformer to predict patient-level treatment outcomes. We trained and extensively evaluated the model using cross-validation and analyzed the extracted features via UMAP visualizations and Grad-CAM saliency maps. The model reached an AUROC of 0.819 (± 0.119) on fivefold cross-validation and an out-of-fold AUROC of 0.807 on the whole dataset, successfully discriminating between patients with short PFS and long PFS using only laparoscopic images. Our approach demonstrates the potential of deep learning to simplify HGSOC triage and improve early treatment planning by accurately stratifying the patients based on minimally invasive laparoscopy at the diagnostic stage.