Convolutional Neural Networks for Classification of Prostate Cancer Metastases Using Bone Scan Images

Abstract: Convolutional neural networks (CNNs) are used to classify directly on bone scan images in two medical tasks: classifying anterior / posterior pose, and classifying bone scan hotspots as metastatic / non-metastatic in patients with prostate cancer and suspected metastatic disease. The networks trained produce highly accurate results in both tasks and current methods are outperformed for all tested body regions when classifying metastatic / non-metastatic hotspots. For one such dataset current methods obtain an area under receiver operating characteristic (ROC) score of 0.9352. By utilising CNNs and other developments in the field an area under ROC of 0.9739 is obtained for the same test set. We consider this a remarkable result given the exclusion of hand-designed heuristics used in current methods.