Universal Lesion Detector: Deep Learning for Analysing Medical Scans

Accurate, automated lesion detection is an important yet challenging task due to the large variation of lesion types, sizes, locations and appearances. Most of the recent work only focuses on lesion detection in a constrained setting-detecting lesions in a specific organ. We tackle the general problem of detecting lesions across the whole body and propose approaches which are not limited to a particular dataset. This is done by redesigning RetinaNet to be more applicable to medical imaging, using a general approach for optimising anchor configurations and by generating additional weak labels from the provided ground truth. We evaluate our approach on two different datasets-a large public Computed Tomography (CT) dataset called DeepLesion, consisting of 32,735 lesions, and a much smaller whole-body Magnetic Resonance Imaging (MRI) dataset made up of only 213 scans. We show that our approach achieves state-of-the-art results, significantly outperforming the best reported methods by over 5% on the DeepLesion benchmark, while being generalisable enough to achieve excellent results on the much smaller whole-body MRI dataset.