<![CDATA[Automated blood cell classification supports hematological diagnosis by providing objective and efficient analysis, but end-to-end deep learning models often require substantial computational resources that limit deployment on low-resource clinical devices. This study evaluates whether frozen deep features extracted from EfficientNetV2B0 or ResNet50 provide better separability for the eight BloodMNIST classes, and examines which classical classifier offers the most practical balance of accuracy, model size, and training time. The BloodMNIST dataset, consisting of 11,959 training images, 1,712 validation images, and 3,421 test images, is processed using data augmentation and Otsu-based unsupervised segmentation before the resulting masks are replicated into three channels and passed into pretrained ImageNet CNNs used strictly as frozen feature extractors. The extracted features are classified using Support Vector Machine with grid search, K-Nearest Neighbor, Artificial Neural Network, and Random Forest, with performance assessed through accuracy, precision, recall, and F1-score. EfficientNetV2 with Support Vector Machine achieves the highest performance, reaching 76.8% test accuracy, 75.3% precision, 72.6% recall, and a 73.6% F1-score, while EfficientNetV2 with Artificial Neural Network provides a comparable 76.2% accuracy and a 73.0% F1-score with a compact 2 MB model size. These findings highlight a clear trade-off between accuracy, model size, and computational cost, demonstrating that hybrid deep-feature pipelines offer lightweight and effective solutions for blood cell classification in resource-constrained clinical settings]]>
Copyrights © 2025
