Learning from the Truth: Fully Automatic Ground Truth Generation for Training of Medical Deep Learning Networks

Automatic medical image analysis has become an invaluable tool in the different treatment stages of diseases. Especially medical image segmentation plays a vital role, since segmentation is often the initial step in an image analysis pipeline. Convolutional neural networks (CNNs) have rapidly become a state of the art method for many medical image analysis tasks, such as segmentation. However, in the medical domain, the use of CNNs is limited by a major bottleneck: the lack of training data sets for supervised learning. Although millions of medical images have been collected in clinical routine, relevant annotations for those images are hard to acquire. Generally, annotations are created (semi-)manually by experts on a slice-by-slice basis, which is time consuming and tedious. Therefore, available annotated data sets are often too small for deep learning techniques. To overcome these problems, we proposed a novel method to automatically generate ground truth annotations by exploiting positron emission tomography (PET) data acquired simultaneously with computed tomography (CT) scans in combined PET/CT systems.