How to train your robot: Artificial intelligence predicts treatment response in ovarian cancer (069)

Objectives: In patients with high-grade serous ovarian cancer (HGSOC), differences in gross morphology are often noted despite a singular histologic diagnosis. The purpose of this investigation was to determine if artificial intelligence (AI) can detect distinct morphologic patterns in HGSOC at the time of laparoscopy and correlate these patterns with clinical outcomes. Methods: We reviewed pre-treatment laparoscopic surgical videos of 113 patients with pathologically proven HGSOC at our institution between 2013 and 2019. A total of 435 representative still-frame images from these videos were used to train an AI model utilizing Deep Learning and neural networks (DenseNet 201) to extract morphologic patterns of disease. Images were divided into training (70%), validation (10%), and test (20%) sets for the development of the model. We evaluated the ability of the model to apply trends in morphologic patterns in two patient populations: those with an excellent response (ER) to standard treatment, defined as progression-free survival (PFS) ≥ 12 months, and those with a poor response (PR) to therapy, defined as PFS ≤ 6 months. Results: Artificial intelligence (AI) analysis of 435 surgical images from four anatomic locations (diaphragm, omentum, peritoneum, and pelvis) could discriminate between predefined dichotomous clinical outcomes of ER and PR. Of the 113 patients in this cohort, more than half (n=61, 53%) had a durable response to therapy (ER). Our trained Deep Learning model successfully identified these patients against those with a short interval to recurrence (PR) with a high degree of accuracy (93%). Utilizing still-frame images alone, the model reached 100% sensitivity in detecting patients with ER to therapy, while it misclassified some patients with PR to therapy (specificity: 63%). The observed lack of specificity may be related to the lower number of images corresponding to patients with PR to therapy (n=167) compared to patients with ER (n=268), as this impacts the development of neural networks utilized for machine learning in AI. Conclusions: This study represents a novel use of AI to predict patient outcomes at the time of laparoscopic assessment for advanced HGSOC. By using still-frame images from surgical videos, our AI model accurately classified patients with a durable response to therapy with a high degree of sensitivity. Clinical application of this technology could have implications for treatment planning for women with HGSOC. Objectives: In patients with high-grade serous ovarian cancer (HGSOC), differences in gross morphology are often noted despite a singular histologic diagnosis. The purpose of this investigation was to determine if artificial intelligence (AI) can detect distinct morphologic patterns in HGSOC at the time of laparoscopy and correlate these patterns with clinical outcomes. Methods: We reviewed pre-treatment laparoscopic surgical videos of 113 patients with pathologically proven HGSOC at our institution between 2013 and 2019. A total of 435 representative still-frame images from these videos were used to train an AI model utilizing Deep Learning and neural networks (DenseNet 201) to extract morphologic patterns of disease. Images were divided into training (70%), validation (10%), and test (20%) sets for the development of the model. We evaluated the ability of the model to apply trends in morphologic patterns in two patient populations: those with an excellent response (ER) to standard treatment, defined as progression-free survival (PFS) ≥ 12 months, and those with a poor response (PR) to therapy, defined as PFS ≤ 6 months. Results: Artificial intelligence (AI) analysis of 435 surgical images from four anatomic locations (diaphragm, omentum, peritoneum, and pelvis) could discriminate between predefined dichotomous clinical outcomes of ER and PR. Of the 113 patients in this cohort, more than half (n=61, 53%) had a durable response to therapy (ER). Our trained Deep Learning model successfully identified these patients against those with a short interval to recurrence (PR) with a high degree of accuracy (93%). Utilizing still-frame images alone, the model reached 100% sensitivity in detecting patients with ER to therapy, while it misclassified some patients with PR to therapy (specificity: 63%). The observed lack of specificity may be related to the lower number of images corresponding to patients with PR to therapy (n=167) compared to patients with ER (n=268), as this impacts the development of neural networks utilized for machine learning in AI. Conclusions: This study represents a novel use of AI to predict patient outcomes at the time of laparoscopic assessment for advanced HGSOC. By using still-frame images from surgical videos, our AI model accurately classified patients with a durable response to therapy with a high degree of sensitivity. Clinical application of this technology could have implications for treatment planning for women with HGSOC.