3D Lung Tumor Segmentation System Using Adaptive Structural Deep Belief Network

Deep Learning has a hierarchical network architecture to represent the complicated feature of input patterns. The adaptive structural learning method of Deep Belief Network (Adaptive DBN) has been developed and achieved the refined classification probabilities on some benchmarks, which can discover an appropriate number of hidden neurons for given input data in a Restricted Boltzmann Machine (RBM) by the neuron generation-annihilation algorithm, and can obtain appropriate number of hidden layers in DBN by the layer generation algorithm. In this chapter, a new segmentation method using Adaptive DBN is proposed to extract lung tumor from 3D CT images. The developed model was evaluated for the 156 cases collected by an open data in NSCLC Radiogenomics. As a result, our model showed 0.884 Dice coefficient for the test data, which was higher value than the 3D U-Net provided by MONAI. In addition, a framework which realizes a series of medical imaging analysis for medical professionals in clinics is proposed as the utilization of our Adaptive DBN model, from collecting data to model training, inference, and re-training for new data.