Automated detection of leukemia from white blood cell (WBC) images holds immense potential for aiding medical professionals in timely diagnosis and treatment. However, existing deep learning models often suffer from computational inefficiency, hindering their practical deployment in resource-constrained environments. In this study, we present a tailored solution by proposing a lightweight variant of the EfficientNetB3 architecture, optimized for leukemia WBC image classification.

Our approach capitalizes on the inherent efficiency of depthwise separable convolutions, strategically integrated into the backbone of the network. By employing this technique, we aim to streamline model parameters and computations while preserving discriminative features critical for accurate classification. We meticulously curate a dataset comprising diverse WBC images, meticulously annotated with various leukemia subtypes and disease stages, ensuring comprehensive model training and evaluation.

Through extensive experimentation, we demonstrate the efficacy of our lightweight model in achieving competitive performance metrics, including accuracy, precision, and recall, when compared to conventional architectures. Moreover, we conduct thorough ablation studies to dissect the impact of depthwise separable convolutions on model efficiency and classification efficacy.

Our findings underscore the significant advancements enabled by our proposed approach, offering a compelling solution for efficient and accurate leukemia detection. This work contributes to the ongoing efforts in leveraging deep learning for enhancing medical diagnostics, paving the way for scalable and accessible solutions in clinical practice.

KEYWORDS : Leukemia, white blood cell images, deep learning, EfficientNetB3, lightweight architecture, computational efficiency, depthwise separable convolutions, classification, dataset curation, performance metrics, ablation studies, medical diagnostics, scalable solutions