25 Background: [F18]DCFPyL (PyL) is a PSMA targeted imaging agent for prostate cancer. Independent of manual feature selection, a deep learning algorithm might offer additional insight into the disease biology. We explored the performance of a deep learning algorithm on PyL images of the primary tumor to predict co-existing distant metastases. Methods: 74 veterans with high risk primary prostate cancer tumors were imaged with both PyL PSMA PET/CT and conventional imaging (bone scan and CT or MRI of the abdomen/pelvis). 26% were confirmed with metastatic disease (M1) by conventional imaging. The PyL images of the primary tumor were analyzed with EXINI’s PyL-AI algorithm. Location of the prostate was defined on low dose CT via automatic segmentation using a deep convolutional network. The segmentations were used to map the PyL PET image of the prostate. The image based PyL-AI model was made up of a Conv3D layer of 4 kernels, a Conv3D layer of 8 kernels, a dense layer of 64 nodes followed by a final dense layer with 2 nodes. The model training was performed on the images using 5-fold cross validation with non-overlapping validation sets. The test predictions were compared with ground truth (M1); the area under ROC curve (AUC) was computed to determine the performance of the model in predicting the presence of distant metastases. A logistical regression model from baseline clinicopathologic features of the primary tumor (baseline PSA, biopsy gleason score, percent cores positive, T stage) was created as a comparator. Results: The logistical regression model using clinicopathologic features had an AUC of 0.71, while the PyL-AI model based on intra-prostatic PyL Images alone had an AUC of 0.81 for prediction of metastatic disease as defined by conventional imaging. Adding clinical parameters in the image based PyL-AI model incrementally increased the AUC to 0.82. Conclusions: The image based PyL-AI deep learning model demonstrates a higher predictive accuracy over the logistic model using classical clinicopathologic features. The study is hypothesis generating observation that needs prospective validation in an independent data set.
Development and validation of a deep learning algorithm for improving Gleason scoring of prostate cancer
