PERFORMANCE EVALUATION OF CNN MODELS USING TRANSFER LEARNING AND ENSEMBLE APPROACHES FOR AUTOMATED LEUKEMIA DIAGNOSIS

Abstract

A significant portion of deaths related to cancer globally are caused by Leukemia, one of the most serious hematological malignancies. It is distinguished by the rapid growth of premature lymphocytes, which disrupts the regular operation of the bone marrow and blood. To increase survival rates and ensure prompt treatment, early and accurate diagnosis is essential. The majority of modern diagnostic techniques, however, rely on the laborious, ineffective, and error-prone manual interpretation of peripheral blood smear (PBS) images. Convolutional neural networks (CNNs), in particular, provide a dependable and automated substitute in deep learning. In this work, we utilize a carefully selected PBS dataset representing four main leukemia types collected in Bangladesh to evaluate six CNN architectures: VGG19, InceptionV3, MobileNetV2, Xception, DenseNet-201, and SecrensNet152. InceptionV3, MobileNetV2, DenseNet-201, VGG19, and SecrensNet152 were all subjected to transfer learning to enhance model generalization. To improve performance, we also developed an ensemble model called DEX, which combines DenseNet-121, EfficientNet-B7, and Xception. With an astounding accuracy of 99%, the trial findings show that DEX outperformed any of the separate CNN models. Accuracy improvements of up to 16% were observed with transfer learning compared to baseline models. These results open the door for the incorporation of CNN ensembles into real-time clinical decision support systems and demonstrate their potential for providing extremely accurate leukemia diagnoses.