Comparative Evaluation of YOLOv5, YOLOv8, and YOLOv11 for Real-Time Leukemia Cell Detection in Blood Smear Images

Acute Lymphoblastic Leukemia (ALL) is a rapidly progressing hematological malignancy that predominantly affects the paediatric population and demands early, accurate diagnosis for favourable clinical outcomes. Conventional diagnostic approaches rely on manual microscopic examination of Peripheral Blood Smear (PBS) images, a process that is labour-intensive, highly subjective, and prone to human error, particularly in resource-limited healthcare environments. This paper presents a novel hybrid deep learning architecture CNN GRU with Attention designed for automated binary classification of PBS images into Healthy and Leukemia categories. The proposed system integrates a pre-trained ResNet-50 Convolutional Neural Network (CNN) for spatial feature extraction from blood cell morphology, a Gated Recurrent Unit (GRU) network for modelling sequential dependencies across spatial regions of the extracted feature maps, and an Attention Mechanism that focuses the model on diagnostically critical regions such as abnormal blast cells. The model is trained on the publicly available Kaggle ALL Blood Cell Cancer dataset comprising over 4,000 PBS images across four classes (Benign, Early Pre-B, Pre-B, Pro-B), converted to binary labels. The system achieves a training accuracy of approximately 88% and a validation accuracy of 88–90%, demonstrating strong generalisation. A full-stack web application Leukemia AI is implemented using FastAPI (backend) and Next.js with Tailwind CSS (frontend), enabling real-time image upload, inference, and confidence-score reporting. The proposed hybrid system offers a clinically relevant, accessible, and computationally efficient solution for AI-assisted leukemia screening.