Lung Cancer Tumor Region Segmentation Using Recurrent 3D-DenseUNet.

Lung cancer is one of the most severe and widespread that constitutes a major public health problem and has a high mortality rate. In this regard, proper segmentation of lung tumor from X-ray, Computed Tomography (CT scan) or, Magnetic Resonance Imaging (MRI) is the stepping stone towards achieving completely automated diagnosis system for lung cancer detection. With the advancement of technology and availability of data, the valuable time of a radiologist can be saved using computer tools for tumor segmentation. In this work, we present a data driven approach for lung tumor segmentation from CT scans by using Recurrent 3D-DenseUNet, a novel fusion of Convolutional and Recurrent neural network. Our approach is to train this network using image-volumes with tumor only slices of size (256 X 256 X 8). A data-driven adaptive weighting method is also used in our approach to differentiate between tumorous and non-tumorous image-slices, which shows more promise than crude intensity thresholding of 0.70, that we have used in this work for competition purpose. Our model has been trained and tested on the NSCLC-Radiomics dataset of 260 patients, provided by The Cancer Imaging Archive (TCIA) for 2018 IEEE VIP Cup. In this dataset, our proposed approach achieves an average dice score of 0.74, mean surface distance of 1.719 and 95% Hausdorff distance of 7.249.