Lightweight EfficientNetB3 Model Based on Depthwise Separable Convolutions for Enhancing Classification of Leukemia White Blood Cell Images

Acute lymphoblastic leukemia (ALL) is a type of leukemia cancer that arises due to the excessive growth of immature white blood cells (WBCs) in the bone marrow. The ALL rate for children and adults is nearly 80% and 40%, respectively. It affects the production of immature cells, leading to an abnormality of neurological cells and potential fatality. Therefore, a timely and accurate cancer diagnosis is important for effective treatment to improve survival rates. Since the image of acute lymphoblastic leukemia cells (cancer cells) under the microscope is complicated to recognize the difference between ALL cancer cells and normal cells. In order to reduce the severity of this disease, it is necessary to classify immature cells at an early stage. In recent years, different classification models have been introduced based on machine learning (ML) and deep learning (DL) algorithms, but they need to be improved to avoid issues related to poor generalization and slow convergence. This work enhances the diagnosis of ALL with a computer-aided system that yields accurate results by using DL techniques. This research study proposes a lightweight DL-assisted robust model based on EfficientNet-B3 using depthwise separable convolutions for classifying acute lymphoblastic leukemia and normal cells in the white blood cell images dataset. The proposed lightweight EfficientNet-B3 uses less trainable parameters to enhance the performance and efficiency of the leukemia classification. Furthermore, two publicly available datasets are considered to evaluate the effectiveness and generalization of the proposed lightweight EfficientNet-B3. In addition, different measures are employed, such as accuracy, precision, recall, and f1-score, to evaluate the effectiveness of the proposed and baseline classifiers. In addition, a detailed analysis is given to evaluate and compare the performance and efficiency of the proposed with existing pre-trained and ensemble DL classifiers. Experimental results show that the proposed model for image classification achieves better performance and outperforms the existing benchmark DL and other ensemble classifiers. Moreover, our finding suggests that the proposed lightweight EfficientNet-B3 model is reliable and generalized to facilitate clinical research and practitioners for leukemia detection.