Artificial intelligence assisted bone lesion detection and classification in computed tomography scans of prostate cancer patients.

e17567 Background: Patients diagnosed with prostate cancer undergo computed tomography (CT) for pretreatment staging to rule out bone metastases. However, detection and classification of bone lesions on CT is challenging and subject to inter-reader variability. We present a cascaded deep learning algorithm for automatic detection and classification of bone lesions on staging CT in patients diagnosed with prostate cancer. Methods: CT scans from 56 patients with histopathologically proven prostate cancer were included. An expert radiologist annotated the extent of individual bone lesions (N = 4217) and labelled all regions as either benign or malignant. All scans were anonymized and normalized at the patient-level prior to training. Our method can be described as a two-stage framework, 1) A detection algorithm: Inspired by Yolo-v3 detection method, we designed a network with a backbone of darknet-53 pretrained on Coco dataset and four final scaling blocks to compensate for wide range of lesion diameters, 2) A classification algorithm: we formed a binary classifier based on ResNet-50 pretrained using the ImageNet dataset. We used a train/validation split equal to 90%/10% for this study. To facilitate the learning process, horizontal flipping, relative zooming and mean weighted averaging were used for data augmentation in stage 1. Instead, the classification algorithm took advantage of synthesized patches generated by Deep Convolutional Generative Adversarial Network (DC-GAN) for augmentation. Results: We could achieve a real-time (~120ms per slice) performance on our validation set with a median penalty of 0.3(0.02-0.78) false positives per true positive within each patient. Overall performance of our detection algorithm was 81% sensitivity and 86% positive predictive value. In stage 2, we obtained an accuracy of 89% for correct classification of benign from malignant bone lesions with no augmentation which was improved to 91% when we incorporated the augmented data for training. Conclusions: Our 2-stage algorithm sequentially detects and classifies bone lesions on CT of prostate cancer patients with a significant performance. To further improve our results and for generalizability we are accruing more data from different centers. Eventually, with greater dataset, both algorithms will be cascaded and trained as a whole unit to become one single tool for fully automatic detection and classification which serves as an aid for radiologists who read the staging CTs.