Alison Marcelino

Brazil

Morphological Classification of Blood Cells with ResNet50: An Explainable Approach Integrating SHAP and Grad-CAM

Claudia Stoeglehner Sahd, Claudia Stoeglehner Sahd, Heron Oliveira dos Santos Lima, Cristhiane Gonçalves, Marcella Scoczynski, Federal University of Technology- Paraná- UTFPR

Abstract

Background

Microscopic examination of peripheral blood smears remains the gold-standard method for identifying a variety of hematologic disorders. However, it is a labor-intensive procedure, subject to interobserver variability and dependent on substantial expertise. In this scenario, a deep-learning–based approach is proposed to enhance the accuracy and standardization of blood-cell morphological classification, incorporating interpretability mechanisms.

Methods

The ResNet50 architecture, optimized through transfer learning, was employed to classify images into six cellular categories: basophils, eosinophils, erythroblasts, lymphocytes, monocytes, and platelets. Training and validation were performed using the Blood Cells Image Dataset, which contains 17,092 images annotated by hematologists from the Hospital of Barcelona and acquired through the CellaVision platform. To ensure transparency in the model’s decisions, two interpretability techniques were applied: SHAP, for identifying decisive morphological attributes, and Grad-CAM, for generating discriminative activation maps.

Results

The network achieved an overall accuracy of 97.33% on the test set, with consistent performance across training and validation, indicating a low tendency toward overfitting. Class-wise F1-scores were high, with eosinophils (0.98) and platelets (0.99) standing out. Morphologically more challenging classes, such as lymphocytes and monocytes, also demonstrated robust performance, with F1-scores between 0.95 and 0.96. The interpretability analyses were consistent with classical hematologic criteria, such as nuclear density in lymphocytes and characteristic granulation patterns in eosinophils and platelets.

Conclusions

The findings demonstrate the potential of integrating deep learning with interpretability methods as a reliable tool to support hematologic diagnosis. The proposed approach contributes to greater precision, transparency, and efficiency in routine laboratory workflows, reinforcing its applicability in clinical settings that require standardized morphological evaluation.