3266 - Estimating Risk of Biochemical Recurrence after RT Utilizing a Convolutional Neural Network

G. Pratt¹, M. Rowan¹, C. Belant², J. Parisi², J. E. Leeman³, L. K. Lee⁴, D. D. Yang⁵, A. V. DAmico⁶, P. L. Nguyen⁶, P. F. Orio III⁷, and M. T. King⁶; ¹Brigham and Womens Hospital/Dana-Farber Cancer Institute, Boston, MA, ²Dana-Farber Cancer Institute/Brigham and Womens Hospital, Boston, MA, ³Department of Radiation Oncology, Dana-Farber Cancer Institute/Brigham and Women’s Hospital, Harvard Medical School, Boston, MA, ⁴Department of Radiology, Brigham and Womens Hospital, Boston, MA, ⁵Harvard Radiation Oncology Program, Boston, MA, ⁶Department of Radiation Oncology, Brigham and Women’s Hospital/Dana-Farber Cancer Institute, Boston, MA, ⁷Department of Radiation Oncology, Dana-Farber Cancer Institute, Boston, MA

Purpose/Objective(s): Previously, we demonstrated that a deep learning segmentation algorithm of magnetic resonance images (MRI) could identify dominant intraprostatic lesions (DILs). We showed that DIL volume, V_AI, was significantly associated with biochemical failure (BF) and metastasis for radiation therapy (RT). However, ~15% of PI-RADS 5 lesions were missed (false negative) by the algorithm. We hypothesize that convolutional neural network approaches could provide prognostic information, particularly for missed lesions. Materials/

Methods: We conducted a single-institutional retrospective study of 373 patients who had at least one PI-RADS 3-5 lesion on MRI and were treated with RT between 2010-2017. The dataset was split into train (246) and test (127) sets. All patients had reference contours delineated by genitourinary radiation oncologists. Reference volumes (V_Ref) were obtained from contours. For each patient, the slice with the largest cross-sectional diameter was identified, and areas =10 mm from the DIL were cropped out. The T2, apparent diffusion weighted (ADC), and high B-value diffusion weighted images (DWI) were assembled into a 3-channel image. A ResNet image classification model, pre-trained on ImageNet-1k, was modified to provide a risk score (scaled from 0 to 100) for BF. The ResNet was parameterized on the training set using 5-fold cross validation, then applied to the test set. We evaluated whether ResNet score was associated with BF utilizing Cox regression, while adjusting for ADT duration, MRI scanner model (2010-2013 vs 2013-2017), and NCCN risk category (low/FIR, UIR, high). We also calculated the 7-year area under the curve (AUC) values for ResNet score, V_Ref, and NCCN risk category.
Results: Of the 127 test set patients, PI-RADS scores were 18 PI-RADS 3, 51 PI-RADS 4, and 58 PI-RADS 5 respectively. Median V_Ref for respective PI-RADS groups were 0.14 [inter-quartile range (IQR) 0.09, 0.38], 0.61 [0.38, 0.91], and 2.72 [1.87, 7.33] mL. Corresponding ResNet scores 32.6 [19.9, 41.4], 42.2 [35.7, 51.0], and 62.5 [53.7, 69.7]. The correlation coefficient between V_Ref and ResNet scores was 0.53. Of note, 28 patients (13 PI-RADS 3, 9 PI-RADS 4, and 6 PI-RADS 5 lesions) had false negative V_AI, while the correlation coefficient between ResNet score and V_Ref was 0.73. There were 19 BF events at a median follow-up of 7.2 years. ResNet score was significantly associated with BF (adjusted hazard ratio: 1.08 (1.03-1.12, p=0.001)). The 7-year AUC values for V_Ref, ResNet, and NCCN risk were 85.7 [77.5, 93.9], 87.1 [77.7, 96.6], and 69.1 [56.4, 81.7], respectively. ResNet had significant improved AUC (18.1; p = 0.002) compared with NCCN, but not V_Ref (p = 0.76).
Conclusion: ResNet is a feasible option for prognosticating patient outcomes in prostate cancer patients treated with RT, particularly for those whose tumors are missed by deep learning segmentation algorithms. Further evaluation in external datasets is warranted.