3D Reconstruction of Lung Tumour Using Deep Auto-encoder Network and a Novel Learning- Based Approach

Abstract

Lung cancer is a common dangerous cancer among men and women worldwide. Using the information about the 3D shape of the lung tumours is useful for determining the cancer type and drug delivery problems. This chapter aims to propose a novel approach for 3D tumour reconstruction from a sequence of 2D parallel CT images. To achieve this goal, we first preprocessed CT images before implementing DBSCAN clustering for lung segmentation. We defined efficient features that made the results more accurate and improved the speed of the DBSCAN algorithm. Next, we designed a deep autoencoder network to extract useful features from each cluster. Then classifications methods are applied to classify tumours among the other clusters. By extracting the tumour area from 2D images, we can construct the 3D shape of tumours using the Marching Cubes algorithm. A novel stochastic approach is proposed to interpolate some intermediate slices between available slices to improve the accuracy of the ultimate 3D shape. Complexity and errors are reduced in the presented approach compared to the previous methods. Finally, results indicate that our approach is more automatic and accurate than the other 3D lung tumour modelling approaches.