Using Deep Learning to Predict Treatment Response in Patients with Hepatocellular Carcinoma Treated with Y90 Radiation Segmentectomy.

Abstract

Treatment of hepatocellular carcinoma (HCC) with Y90 radioembolization segmentectomy (Y90-RE) demonstrates a tumor dose-response threshold, where dose estimates are highly dependent on accurate SPECT/CT acquisition, registration, and reconstruction. Any error can result in distorted absorbed dose distributions and inaccurate estimates of treatment success. This study improves upon the voxel-based dosimetry model, one ofthe most accurate methods available clinically, by using a deep convolutional network ensemble to account for the spatially variable uptake of Y90 within a treated lesion. A retrospective analysis was conducted in patients with HCC who received Y90-RE at a single institution. Seventy-seven patients with 103 lesions met the inclusion criteria: three or fewer tumors, pre- and post treatment MRI, and no prior Y90-RE. Lesions were labeled as complete (n = 57) or incomplete response (n = 46) based on 3-month post treatment MRI and divided by medical record number into a 20% hold-out test set and 80% training set with 5-fold cross-validation. Slice-wise predictions were made from an average ensemble of models and thresholds from the highest accuracy epochs across all five folds. Lesion predictions were made by thresholding all slice predictions through the lesion. When compared to the voxel-based dosimetry model, our model had a higher F1-score (0.72 vs. 0.2), higher accuracy (0.65 vs. 0.60), and higher sensitivity (1.0 vs. 0.11) at predicting complete treatment response. This algorithm has the potential to identify patients with treatment failure who may benefit from earlier follow-up or additional treatment.