New Developments in Computational Methods for CNN-Based Models-Based Blood Cell Classification

Abstract

Recent years have seen a surge of interest in using advanced computational analysis for applications like medical imaging as the dataset sizes required also increased. This paper proposes a new method to detect and classify different type of blood cell based on Convolutional Neural Network (CNN) model. We show in a case study, that our approach does outperform traditional methods on blood cell images with approximately 1000 new and unseen cells. Blood and its constituents act as important markers for various physiological conditions that are vital in humans. Pathologists have traditionally used optical microscopy of blood images to solve problems that are due to diseases, such as leukemia and aid. Recently, data-driven approaches have been abandoned by some and replaced with computer aided diagnostic systems designed for blood disorders from microscopic images. The use of CNNs as feature extractors means that we no longer need to manually handcraft features, allowing the model to automatically learn and determine relevant image attributes all by itself. This also provides learning extracted features and makes it easy for integrating with traditional machine learning algorithms (SVM, KNN, Random Forest) to combine the merits of deep learning and classical methods. In addition, having a measure of how well classifiers are doing in five raw image features is very useful to have an idea about the relevance/faults in each feature which prevent properly categorizing tasks. Our study provides insights to better understand the informativeness of image-based features and recommendations on their application specifically in classification tasks.

Note:
Funding Information: There is no funding information to disclose.

Declaration of Interests: There are no competing interests to declare.

Keywords: Blood Cell Classification, Convolutional Neural Networks, Machine Learning, Medical Imaging, Diagnostic Systems