CANCER DETECTION FOR LOW GRADE SQUAMOUS ENTRAEPITHELIAL LESION

The National Cancer Institute estimates in 2012, about 577,190 Americans are expected to die of cancer, more than 1,500 people a day. Cancer is the second most common cause of death in the US, accounting for nearly 1 of every 4 deaths. Cancer diagnosis has a very important role in the early detection and treatment of cancer. Automating the cancer diagnosis process can play a very significant role in reducing the number of falsely identified or unidentified cases. The aim of this thesis is to demonstrate different machine learning approaches for cancer detection. Dr. Tawfik, pathologist from University of Kansas medical Center (KUMC) is an inventor of a novel pathology tissue slicer. The data used in this study comes from this slicer, which successfully allows semi-automated cancer diagnosis and it has the potential to improve patient care. In this study the slides are processed and visual features are computed and the dataset is made from scratch. After features extraction, different machine learning approaches are applied on the dataset which has shown its capability of extracting high-level representations from high-dimensional data. Support Vector Machine and Deep Belief Networks (DBN) are the concentration in this study. In the first section, Support vector machine is applied on the dataset. Next Deep Belief Network which is capable of extracting features in an unsupervised manner is implemented and with back-propagation the network is fine tuned. The results show that DBN can be effective when applied to cytological cancer diagnosis by increasing the accuracy in cancer detection. In the last section a subset of DBN features are selected and then appended with raw features and Support Vector Machine is trained and tested with that. It shows improvement over the first section results. In the end the study infers that Deep Belief Network can be successfully used over other leading classification methods for cancer detection