Optimised feature selection for early cancer detection

Abstract

Global Cancer Incidence, Mortality and Prevalence (GLOBOCAN) status report for the year of 2020, suggests the occurrence of 10.0 million cancer deaths and 19.3 million new cancer cases. Clearly, cancer incidence and mortality are rapidly growing worldwide. Also, the leading causes of cancer deaths are found to be lung cancer and breast cancer. Cancer cells are having the probability of spreading to other parts of the body too. Most chronic cancers are not curable, but some can be controlled for a few months or years. Also, there is a possibility of high rate of relapse of the disease. These remissions can be partial or complete. But, if detected early, certain cancers can be treated by surgery, chemotherapy, and radiation therapy. This research work focuses on detecting cancer in its early stage so that right measures can be taken to combat the disease. In this attempt to create a beneficial working model, the combination of Artificial Neural Network (ANN), Convolution Neural Network, Graph based Neural Network with Genetic Algorithm (GA) have proven to be successful. As a proof of concept, we present a combination of feature selection techniques that can effectively reduce the feature set and optimize the classification techniques. The proposed method, when applied on a benchmark dataset, gave a higher accuracy by selecting most relevant 7 features out of 10 with an accuracy of 95.7%. Using Convolution Neural Network, the accuracy improved to 98.3% with optimal hyperparameter tuning.