A Convolutional Neural Network Framework for Automated Leukemia Classification from Blood Smear Images

Abstract The increasing demand for reliable and efficient diagnostic support in hematology has motivated the application of deep learning techniques to medical image analysis. Among blood disorders, leukemia remains a critical condition that requires timely and accurate classification of subtypes for effective treatment planning. In this study, we propose the design of a convolutional neural network architecture, referred to as ModifiedNet, aimed at classifying four leukemia subtypes plus healthy cells from peripheral blood smear images. The network is developed based on architectural principles drawn from AlexNet, GoogLeNet, and ResNet, incorporating depthwise separable convolutions, dilated convolutions, and residual connections to enhance feature extraction efficiency while maintaining considerable network depth with reduced parameter count. Training was performed on an augmented dataset using standardized hyperparameters to ensure fair comparison with baseline architectures. ModifiedNet achieved superior performance with 97.24% test accuracy, outperforming established models including AlexNet (94.22%), GoogLeNet (95.1%), and ResNet-18 (93.94%). The model demonstrated strong discriminative ability across all leukemia subtypes, with particularly robust performance in distinguishing healthy cells from malignant variants. These results validate the effectiveness of the proposed architectural modifications for automated leukemia classification tasks.