Computer-Aided Diagnosis System for Blood Diseases Using EfficientNet-B3 Based on a Dynamic Learning Algorithm

• Published in: Diagnostics (Basel) Vol. 13; no. 3; p. 404
• Main Authors: Abd El-Ghany, Sameh, Elmogy, Mohammed, El-Aziz, A. A. Abd
• Format: Journal Article
• Language: English
• Published: Switzerland MDPI AG 22.01.2023; MDPI
• Subjects: Accuracy; acute lymphoblastic leukemia (ALL); Algorithms; Annealing; Blood; Blood diseases; Blood platelets; Bone marrow; Cancer; computer-aided diagnostic (CAD); Computer-aided medical diagnosis; Datasets; deep learning; Diagnosis; Disease; EfficientNet-B3; Immune system; learning rate (LR); Leukemia; Machine learning; malaria parasite; Methods; Morphology; Neural networks; Saudi Arabia; learning rate (LR); deep learning; malaria parasite; EfficientNet-B3; computer-aided diagnostic (CAD); acute lymphoblastic leukemia (ALL)
• ISSN: 2075-4418; 2075-4418
• DOI: 10.3390/diagnostics13030404

The immune system’s overproduction of white blood cells (WBCs) results in the most common blood cancer, leukemia. It accounts for about 25% of childhood cancers and is one of the primary causes of death worldwide. The most well-known type of leukemia found in the human bone marrow is acute lymphoblastic leukemia (ALL). It is a disease that affects the bone marrow and kills white blood cells. Better treatment and a higher likelihood of survival can be helped by early and precise cancer detection. As a result, doctors can use computer-aided diagnostic (CAD) models to detect early leukemia effectively. In this research, we proposed a classification model based on the EfficientNet-B3 convolutional neural network (CNN) model to distinguish ALL as an automated model that automatically changes the learning rate (LR). We set up a custom LR that compared the loss value and training accuracy at the beginning of each epoch. We evaluated the proposed model on the C-NMC_Leukemia dataset. The dataset was pre-processed with normalization and balancing. The proposed model was evaluated and compared with recent classifiers. The proposed model’s average precision, recall, specificity, accuracy, and Disc similarity coefficient (DSC) were 98.29%, 97.83%, 97.82%, 98.31%, and 98.05%, respectively. Moreover, the proposed model was used to examine microscopic images of the blood to identify the malaria parasite. Our proposed model’s average precision, recall, specificity, accuracy, and DSC were 97.69%, 97.68%, 97.67%, 97.68%, and 97.68%, respectively. Therefore, the evaluation of the proposed model showed that it is an unrivaled perceptive outcome with tuning as opposed to other ongoing existing models.