Certain investigations on supervised and unsupervised machine learning algorithms for predicting the nature of tumors in multimodal data

Abstract

The latest statistics from World Health Organization (WHO) for the year 2018 shows that cancer is the second leading cause of deaths worldwide It accounts for 96 million deaths that occurred worldwide It has also been estimated that worldwide there would be 236 million new cases of cancer each year by 2030 Early detection of cancer is very important It helps the physicians/radiologists in treatment The cancer will be completely curable if it is identified in the initial stages Cancer identification essentially comprises tests including Magnetic Resonance Imaging (MRI) Nerve test Biopsy etc newlineIn this thesis three novel techniques are presented for tumor classification in MRI These techniques aim for improved accuracy achievement of dimensionality reduction and decrease in computational cost factors in tumor classification Based on these factors three techniques are proposed: 1 Multilayer Extreme Learning Machine ML ELM based on Possibilistic Fuzzy C Means Clustering 2 Stacked Extreme Learning Machine based on subtractive clustering method and 3 Higher Order Neural Network based on optimization algorithms The proposed methods can segment and classify multimodal tumor data from MRI as cancerous or normal The techniques can act as a valuable decision support system for tumor diagnosis and can aid in further treatment Such techniques are motivated by the need for automation in diagnostic measures and providing adequate assistance in tumor classification newline newline