Hyperparameter Selection in Deep Learning Model for Classification of Philadelphia-Chromosome Negative Myeloproliferative Neoplasm

Abstract

Myeloproliferative neoplasm (MPN), is a rare type of cancer compare to tumorous cancer. Thus, delay diagnosis is common, resulting in high potentially preventable complications. This is due to the dependency on physician clinical experience and cognitive level, which may cause serious mistakes such as mistreatment or misdiagnosis. In recent years, deep learning has reached new peak in cancer classification with better expediency and accuracy than physician diagnosis. In this study, three types of MPNs namely polycythemia vera (PV), essential thrombocythemia (ET) and primary myelofibrosis (PMF) were classified using convolution neural network. In order to find the best classification output, hyperparameter tuning was applied by tweaking optimizer, input size and number of epochs. Model that shows good performance was RMSprop optimizer with input size 32 × 32 and 120 number of epochs which produced testing accuracy of 91.64%. Hence, this study proved deep learning algorithm can be applied to classify three types of MPNs with appropriate hyperparameter tuning.