Saliency Based Deep Neural Network for Automatic Detection of Gadolinium-Enhancing Multiple Sclerosis Lesions in Brain MRI

The appearance of contrast-enhanced pathologies (e.g. lesion, cancer) is an important marker of disease activity, stage and treatment efficacy in clinical trials. The automatic detection and segmentation of these enhanced pathologies remains a difficult challenge, as they can be very small and visibly similar to other non-pathological enhancements (e.g. blood vessels). In this paper, we propose a deep neural network classifier for the detection and segmentation of Gadolinium enhancing lesions in brain MRI of patients with Multiple Sclerosis (MS). To avoid false positive and false negative assertions, the proposed end-to-end network uses an enhancement-based attention mechanism which assigns saliency based on the differences between the T1-weighted images before and after injection of Gadolinium, and works to first identify candidate lesions and then to remove the false positives. The effect of the saliency map is evaluated on 2293 patient multi-channel MRI scans acquired during two proprietary, multi-center clinical trials for MS treatments. Inclusion of the attention mechanism results in a decrease in false positive lesion voxels over a basic U-Net [2] and DeepMedic [6]. In terms of lesion-level detection, the framework achieves a sensitivity of 82% at a false discovery rate of 0.2, significantly outperforming the other two methods when detecting small lesions. Experiments aimed at predicting the presence of Gad lesion activity in patient scans (i.e. the presence of more than 1 lesion) result in high accuracy showing: (a) significantly improved accuracy over DeepMedic, and (b) a reduction in the errors in predicting the degree of lesion activity (in terms of per scan lesion counts) over a standard U-Net and DeepMedic.