Multi-label Lung Lesion Segmentation with Graph Convolutional Network

The diagnosis of many lung diseases in chest CTs requires considering multiple categories of lesions simultaneously. Locating and segmenting all these lesions accurately is difficult and challenging, due to similar appearance, noise, or class imbalance in the CT images. In this paper, we propose a novel two-stage multi-label lung lesion segmentation method, named MedSegGCN, to segment mutiple categories of lung lesions at the same time. It separates the segmentation of all the lesion regions and the classification of the lesion categories into two stages. After predicting all the lesion regions as a binary segmentation in stage-I, in stage-II, we over-segment CT images into supervoxels, and develop a spatial GCN (Spa-GCN) module to construct a graph based on both supervoxels and binary segmentation predictions, and then perform supervoxel-level lesion category classification. With the Spa-GCN, we are able to exploit the spatial consistency and feature consistency in the whole volume, and hence achieve improved segmentation accuracy. Experiments on a self-collected dataset show that our MedSegGCN disperses the burden of locating and classifying the lesions, and it outperforms the state-of-the-art methods by significant margins in multi-label lesion segmentation.