Leukocytes Image Classification Using Optimized Convolutional Neural Networks

• Developed a fast and accurate hybrid approach for leukocyte classification. • Achieved fast convergence of hyperparameters with a random search approach. • Demonstrated superior performance of an optimized convolutional neural network. Hematologic diseases and blood disorders can be studied through the microscopic or chemical examination of blood smear images. Many researchers work on identifying, counting, and classifying different types of blood cells as a theoretical and practical problem that is crucial for disease diagnosis and treatment planning. There are various approaches to classify blood cells such as thresholding, morphological operators, segmentation, edge-based techniques, region-based techniques, and hybrid approaches. Each of these techniques has several limitations in effectively classifying different types of cells; however, methods based on deep learning (DL) have remarkably contributed to the progress of blood cell classification by combining feature extraction, feature selection, and classification into one interconnected step. This study develops a hybrid approach of DL and optimization for accurate and efficient classification of four types of leukocytes: neutrophils, eosinophils, lymphocytes, and monocytes. Model hyperparameters are optimized using grid search (GS) and random search (RS), in which a convolutional neural network (CNN) is used to classify leukocytes. CNNs work through pattern recognition to detect significant features that help distinguish different classes. The blood cell count and detection (BCCD) dataset provides basic information, but the data is insufficient and highly unbalanced for CNNs to accurately classify the images, so the data is augmented to improve model performance. This segmentation-free optimized CNN achieved a classification accuracy of 97% for the validation set, which includes 2,487 cell images, and 99% for the training set, which includes 9,966 cell images. The model reached a sensitivity and specificity of 94% and 98%, respectively. RS accelerates the process of hyperparameter optimization while achieving the same accuracy as GS. The results are compared with the results accomplished by recent CNN models on the BCCD database using seven performance measures and demonstrate the superior performance and competence of the proposed method. This research study develops a fast and accurate approach for leukocyte classification and can be beneficial for other image classification tasks and help clinicians in diagnosing blood diseases.