OPTIMIZED DEEP LEARNING PIPELINE FOR CLASSIFYING LEUKEMIA AND HEALTHY CELLS WITH HIGH RELIABILITY

Abstract

 Leukemia remains a critical health concern requiring early and accurate detection to improve patient outcomes. Traditional diagnostic methods often suffer from limitations in precision and scalability. To address these challenges, this study proposes an optimized deep learning pipeline designed for reliable classification of leukemia (ALL) and healthy (Hem) cells using microscopic blood smear images. The approach uses a conditional GAN (cGAN) for synthetic data generation, enriching the dataset with diverse, realistic images. A hybrid model combining Swin Transformer and EfficientNetV2 is employed for feature extraction, followed by Bayesian hyperparameter optimization to enhance model performance. Classification is performed using an Attention-based Denoising Autoencoder (DAE) with fully connected layers. The methodology is validated on two datasets comprising 12,529 real images and an additional 8000 cGAN-generated images. The proposed pipeline achieved 99.85% accuracy on dataset 1 and 99.24% accuracy on dataset 2, demonstrating a 1.95% improvement over existing methods.