LeukemiaCellNet: a hybrid CNN-transformer architecture for accurate classification of leukemia blood cells

Abstract

Timely and precise leukemia diagnosis remains a challenge in healthcare. This study improves accuracy utilizing powerful AI technologies, with a focus on deep learning models. This study examines CNN architectures, including VGG., Xception, Inception, ResNet, and DenseNet, together with transformer-based models like ViT, DaViT and MaxViT. A hybrid model merging ViT and ResNet50 was proposed, utilizing CNNs for image feature extraction and transformer networks for feature representation. Analysis using a medical-image dataset assessed the F1 score, recall, and precision of the model. Meanwhile, the model ResNet acquired an accuracy of 54.61% but the transformer models, MaxVit and Davit achieved 53%. However, ResNet50- ViT(Hybrid model) excelled all models, by attaining 92.87% accuracy with wellbalanced recall and precision. The findings illustrate the promise of fully automatic leukemia diagnosis as potentially disruptive, with the possibility of using hybrid AI models to address healthcare diagnostic problems in an accurate and timely manner. The combination of state-ofthe- art convolutional and attention mechanisms provides adequate solutions to the complexities associated with leukemia, setting a new standard for AI in hematology and opening new opportunities for the development of new automated diagnostic systems.