238O - Deep learning magnetic resonance imaging radiomic predict platinum-sensitive in patients with epithelial ovarian cancer

Background

Platinum-sensitive is an important basis for the treatment of recurrent epithelial ovarian cancer (EOC) without effective methods to predict. We aimed to develop and validate the EOC deep learning system to predict the platinum-sensitive of EOC patients through analysis of enhanced magnetic resonance imaging (MRI) images before initial treatment.

Methods

Ninety-three EOC patients received platinum-based chemotherapy (>= 4 cycles) and debulking surgery from Sun Yat-sen Memorial Hospital in China from January 2011 to January 2020 were enrolled. We defined platinum-resistant and platinum-refractory patients as platinum-resistant group, and patients who relapsed 6 months or more after initial platinum-base chemotherapy as platinum-sensitive group. Patients were collected and randomly assigned (2:1) to the training and validation cohorts. A deep learning model-Med3D (Resnet 10 version) was first applied to two MRI sequences (T1+C, T2WI) to automatically extract 1024 features of each patient, then established signatures to predict platinum-resistance.

Results

The area under curve (AUC) of the whole MRI volume signature yielded was 0.97, 0.98 for the training and validation cohorts, respectively, which was better than that with the primary tumor signature (AUC 0.78 and 0.85 in training and validation cohorts, respectively). The whole MRI volume signature sensitivity was 0.96 in identifying platinum-sensitive in training cohort, and validated in 0.96(95%CI 0.88-1.0) in the validation cohort. The whole MRI volume signature was superior sensitivity than with MRI primary tumor signature (0.86 and 0.84 [95%CI 0.70-0.98] in training and validation cohort, respectively). The whole MRI volume signature’s specificity was 0.92 and 1 (95%CI 1.0-1.0) in the training and validation cohorts. The primary tumor MRI signature’s specificity was 0.77 and 0.66 (95%CI 0.28-1.0) in the training and validation cohorts.

Conclusions

This deep-learning EOC signature achieved a high predictive power for platinum-sensitive, and the signature based on MRI whole volume is better than that on primary tumor area only.

Clinical trial identification

Editorial acknowledgement