Abstract:

Abstract White Blood Cells build the base of the human immunity and hence hold a critical place in haematological disease diagnosis. Since there exist 12 distinct types in white blood cells which vary by only a small margin, propose a hybrid classification model which combines deep learning techniques along with optimization algorithms for achieving higher performance. The proposed model utilizes high-performance, state-of-the-art technologies. A total of 1460 images are obtained from the standard Kaggle database. The input images are preprocessed using bilateral filter and contrast limited adaptive histogram equalization algorithm is applied for contrast enhancement. The preprocessed imaged are then segmented using UNet architecture of convolutional neural networks. Features are then extracted using Capsule Net machine learning approach. Finally, WBC images are classified into five types namely Eosinophils, Neutrophils, Basophils, Lymphocytes and Monocytes using stacked sparse auto encoder and optimized with Mayfly optimization algorithm. The proposed model is compared with existing algorithms like Support Vector Machine, DenseNet, Inceptionv3, ResNet, Convolutional Neural Network and is found to have superior performance. It achieves an accuracy of 97.79%, precision score of 97.40%, Recall of 97.40%, specificity of 97.17%, F1-Score of 97.4% and ROC value of 0.998.

Keywords: