Performance analysis on hybrid deep learning model for classification of microscopic blood smear images

Abstract

Haematological diagnosis using manual analysis of peripheral blood smears using microscopes is time-consuming, subjective, and heavily relies on the skills of the expert; this study explores whether hybrid convolutional-transformer can better enhance automated microscopic analysis. accuracy and ability of multi-classification WBC. Comparative analysis is performed with the help of a large publicly available dataset of microscopic images of WBCs where four models are being trained and tested based on a standardized protocol: a ResNet-50-based convolutional neural network, a Vision Transformer (ViT-Base/16), a feature-level model that fuses CNN with ViT embeddings, and a proposed hybrid-CNN-ViT architecture that adds a transformer module to a CNN backbone. With the experimental results, both CNN and ViT baselines perform well, but the hybrid CNN-ViT model is superior and offers the best performance, i.e. macro F1-score of 0.980 and macro-AUC of 0.998 with statistically significant performances over the competing architectures. Interpretability analysis also shows that the hybrid model is efficient in integrating finer-level local morphological information, e.g. nuclear texture, cytoplasmic granularity, with global contextual based information, e.g. nucleus-to-cytoplasm percentage and cell general morphology. These results imply that were utilized hybrid CNNtransformer systems can help to increase the accuracy of automated WBC screening and facilitate laboratory processes in clinical haematology. The main contribution of this study is a comprehensive and statistically validated comparison of CNN, transformer, fusion, and hybrid designs, together with a morphology-aware hybrid architecture that establishes a new performance benchmark for automated WBC classification.