To evaluate the performance of deep learning (DL) classifiers in discriminating normal and abnormal 18F-FACBC (fluciclovine, Axumin®) PET scans based on the presence of tumor recurrence and/or metastases in patients with prostate cancer (PC) and biochemical recurrence (BCR). A total of 251 consecutive 18F-fluciclovine PET scans were acquired between September 2017 and June 2019 in 233 PC patients with BCR (18 patients had 2 scans). PET images were labeled as normal or abnormal using clinical reports as the ground truth. Convolutional neural network (CNN) models were trained using two different architectures, a 2D-CNN (ResNet-50) using single slices (slice-based approach) and the same 2D-CNN and a 3D-CNN (ResNet-14) using a hundred slices per PET image (case-based approach). Models' performances were evaluated on independent test datasets. For the 2D-CNN slice-based approach, 6800 and 536 slices were used for training and test datasets, respectively. The sensitivity and specificity of this model were 90.7% and 95.1%, and the area under the curve (AUC) of receiver operating characteristic curve was 0.971 (p < 0.001). For the case-based approaches using both 2D-CNN and 3D-CNN architectures, a training dataset of 100 images and a test dataset of 28 images were randomly allocated. The sensitivity, specificity, and AUC to discriminate abnormal images by the 2D-CNN and 3D-CNN case-based approaches were 85.7%, 71.4%, and 0.750 (p = 0.013) and 71.4%, 71.4%, and 0.699 (p = 0.053), respectively. DL accurately classifies abnormal 18F-fluciclovine PET images of the pelvis in patients with BCR of PC. A DL classifier using single slice prediction had superior performance over case-based prediction.
Read full abstract