Hyperparameter Tuning in Deep Learning Approach for Classification of Classical Myeloproliferative Neoplasm

Abstract

Histopathology images are an essential resource for defining biological compositions or examining the composition of cells and tissues. The analysis of histopathology images is also crucial in supporting different class of disease including for rare disease like Myeloproliferative Neoplasms (MPN). Despite technological advancement in diagnostic tools to boost procedure in classification of MPN, morphological assessment from histopathology images acquired by bone marrow trephine (BMT) is remained critical to confirm MPN subtypes. However, the outcome of assessment at a present is profoundly challenging due to subjective, poorly reproducible criteria and highly dependent on pathologist where it caused interobserver variability in the interpretation. To address, this study developed a classification of classical MPN namely polycythemia vera (PV), essential thrombocythemia (ET) and primary myelofibrosis (MF) using deep learning approach. Data collection was undergoing several image augmentations processes to increase features variability and expand the dataset. The augmented images were then fed into CNN classifier followed by implementation of cross validation method. Finally, the best classification model was performed 95.3% of accuracy by using Adamax optimizer. High accuracy and best output given by proposed model shows significant potential in the deployment of the classification of MPN and hence facilitates the interpretation and monitoring of samples beyond conventional approaches.