A Comparative Evaluation of Transfer Learning Techniques for White Blood Cell Detection

Standard deep-learning (DL) and machine learning (ML) approaches are valuable frameworks in computer vision that improve the accuracy of medical image diagnosis and classification, including the identification of microscopic blood cells. This study examines the detection and categorization of acute leukemia in detail. Improving patient prognosis and treatment options requires early diagnosis, categorization, and accurate detection of white blood cells. Developing a precise and effective model for identifying and classifying malignancies in white blood cell (WBC) images remains challenging despite the widespread use of microscopes for blood cell examination and advancements in AI-based detection techniques. To address these challenges, the authors utilized a total of 48,000 images, comprising both public and private sources, after augmenting three classes of white blood cell (WBC) cells. This study aims to evaluate whether Vision Transformers can match or surpass the performance of convolutional neural networks (CNNs) for WBC classification using large-scale datasets, and to determine whether spatial inductive biases inherent in CNNs offer a measurable advantage, and achieved accuracies of 95.07%, 95.27%, 82.66%, 92.00%, and 83.59%., also creating an Ensemble model that combine three models(VGG19, Xception, and U-Net) running on gpu.v2-4090x4 server with RAM 384 GB because of volume of data which achieving accuracy of 92%, Also study proposed an architecture for deep learning that automatically recognizes and classifies WBC images, categorizing them into five types. The models for detection and classification techniques were also assessed using accuracy, F1 score, recall, and precision for each class of WBC.