Hybrid Tumor Segmentation Framework Integrating Generative Adversarial Networks with Topological Active Nets

Earlier detection of Lung cancer involves segmenting tumor regions in whole body Positron Emission Tomography (PET)/ computed tomography (CT) images. Compared to segmenting tumor on isolated regions, segmenting tumor on whole body region is complex due to its higher metabolic similarity with surrounding tissues. Though many methods have been proposed to segment lung tumors, their tumor detection performance outside thorax region is poor. Thus, they cannot be used successfully for segmenting tumor in whole body regions. This work proposes a deep learning integrated with topological active net (TAN) for accurate segmentation of tumor from whole body scans. TAN is applied at first level to get approximate boundaries for legions which is then processed by deep learning Generative Adversarial Network (GAN) to get the accurate location of legion by filtering the noises.