Unsupervised Clustering Algorithm as Region of Interest Proposals for Cancer Detection Using CNN

Deep learning methods are now getting lot of attention due to their success in many fields. Computer-aided bio-medical image analysis systems act as a tool to assist medical practitioners in correct decision making. Use of Deep Learning algorithms to predict Cancer at early stage was highly promoted by Kaggle Data Science Bowl 2017 competition and Cancer Moonshot Initiative. In this article, we have proposed a novel combination of unsupervised machine learning tool which is modularity optimization based graph clustering method and Convolutional Neural Networks (CNN) based architectures for lung cancer detection. The unsupervised clustering method helps in reducing the complexity of CNN by providing Region of Interest (ROI) proposals. Our CNN model has been trained and tested on LUNA 2016 dataset which contains Computed Tomography (CT) scans of lung region. This method provides an approximate pixel-wise segmentation mask along with the class label of the ROI proposals.