Comparative analysis of diverse architectures for accurate blood cancer cell classification

Abstract

Blood cancer cell diagnosis is crucial in medical diagnostics. It demands accurate classification of blood cell images. Proper classification of blood cancer cells is fundamental for accurately diagnosing the specific type and subtype of blood cancer as well as efficient treatment planning. It provides specific information, helping healthcare professionals predict the prognosis, survival rates, and the potential for disease recurrence. On this basis, deep learning models have demonstrated remarkable performance. In this paper, we have introduced a comprehensive research into blood cancer cell classification, employing a diverse dataset encompassing various blood cell types. We explore the effectiveness of VGG19 with batch normalization, Vision Transformer (ViT), Ensemble Adversarial Inception-ResNetV2, DenseNet201, and ResNeXt50 architectures in this challenging task. In order to enhance the model performance, we integrate essential data preprocessing techniques, including resizing, cropping, and normalization. Additionally, novel data augmentation strategies, such as random cropping, and flipping are introduced to augment the training dataset and improve model generalization. Remarkably, VGG19 with batch normalization has shown tremendous success by achieving 99.86% accuracy. Moreover, the DenseNet201 model has performed brilliantly by achieving an accuracy of 98.55%. The ResNeXt50 model shows excellent performance with an accuracy of 98.31%. The Vision Transformer (ViT) achieves a solid accuracy of 98.91%. Lastly, Ensemble Adversarial Inception-ResNet V2 also performed no-tably achieving 96.38%. In this context, VGG19 with batch normalization is able to show more excellent performance than other models by achieving the highest accuracy.