Harnessing the Power of Graph Propagation in Lung Nodule Detection

Automatic detection of lung nodules has been a key element in modern medical research in the past decade. By utilizing accurate nodule detection approaches, lung cancer can be treated in its early stages, reducing its mortality rate. Manual detection of nodules is hindered by structures contained in the computed tomography scan images such as veins, bronchioles and lymphatics. Furthermore, nodules are tiny and sometimes have haphazard boundaries making them easy to miss. In this work, we propose a two-step lung nodule detection method that uses a novel hierarchical superpixel merging (HSPM) module to reduce the number of proposed regions of interest and couples it with a new graph neural network namely, the regional downsampled residual superpixel aggregation network (dRes-SPAN) to accurately detect nodules in the lung parenchyma. Experiments on the Lung Image Database Consortium and Image Database Resource Initiative (LIDC-IDRI) dataset show a high improvement in performance in terms of nodule classification and detection while achieving an overall low false positive rate. A significant reduction in training time and trainable parameters has also been observed along with the relaxation in the requirement for heavy annotation compared to the state-of-the-art object detection architectures.