Hybrid deep learning model for acute lymphoblastic leukemia (ALL) detection using pre-trained ResNet-50 and Vision Transformer architecture
Jason Hendrawan, Timoteus Fabian Halim, Jackson Bhakti Gusman, Widodo Budiharto · Procedia Computer Science · 2025
Early and accurate detection of cancer such as Acute Lymphoblastic Leukemia (ALL) is important for effective treatments and improved survival rates. Traditional diagnostic method such as manual analysis of blood smear images are time-consuming and vulnerable to human error. To overcome those weaknesses, this study proposes a hybrid deep learning framework which combines a Convolutional Neural Network (CNN), specifically a pretrained ResNet-50, with a Vision Transformer (ViT) to address both local feature extraction and global context analysis in microscopic blood images. The architecture mimics the approach of expert hematopathologists by extracting spatial features with ResNet-50 and capturing long-range dependencies using ViT’s self-attention mechanisms. Our model’s diagnostic precision is enhanced by considering both local cellular characteristics and their broader spatial relationships. As a result, this model achieves an exceptional classification accuracy of 98% in distinguishing leukemic from healthy cells on the pediatric-focused C-NMC 2019 dataset containing over thousands of microscopic blood smear images, while maintaining computational efficiency with 37-second training epochs which is a critical advantage for clinical deployment in resource-constrained settings. This research shows the effectiveness of integrating CNNs and Transformers in medical imaging, particularly for pediatric ALL diagnosis and offers a robust and scalable solution for pediatric ALL diagnosis.