DEEP LEARNING FOR BLOOD CELL CLASSIFICATION: AN EXPLAINABLE ARTIFICIAL INTELLIGENCE APPROACH FOR DIAGNOSTICS IN HEMATOLOGY

ALISON HENRIQUE MARCELINO; Claudia Stoeglehner Sahd; Heron dos Santos Lima; Elisangela Ap. da Silva Lizzi; Cristhiane Goncalves; MARCELLA SCOCZYNSKI RIBEIRO MARTINS; Heron dos Santos Lima

Título do Trabalho

Autores

ALISON HENRIQUE MARCELINO
Claudia Stoeglehner Sahd
Heron dos Santos Lima
Elisangela Ap. da Silva Lizzi
Cristhiane Goncalves
MARCELLA SCOCZYNSKI RIBEIRO MARTINS
Heron dos Santos Lima

Modalidade

Poster

Área temática

Systems Biology and Modeling

Data de Publicação

02/10/2025

País da Publicação

Brazil | Brasil

Idioma da Publicação

Inglês

Página do Trabalho

https://www.even3.com.br/anais/xmeeting-2025/1112366-deep-learning-for-blood-cell-classification--an-explainable-artificial-intelligence-approach-for-diagnostics-in-

ISBN

978-65-272-1722-0

Palavras-Chave

Blood cell classification, Deep learning, SHapley Additive exPlanations, Explainability,Neural network

Resumo

The microscopic analysis of peripheral blood smears is considered the gold standard for detecting various hematological disorders. However, it is a time-consuming, repetitive process, prone to subjective interpretation, as different operators may reach different conclusions from the same sample. This study proposes a deep learning-based approach to enhance the accuracy of blood cell classification and segmentation, thereby reducing the workload of pathologists and improving diagnostic efficiency. We employed the deep neural network model ResNet50 to classify images into six cell types—basophils, eosinophils, erythroblasts, lymphocytes, monocytes, and platelets—complemented by SHAP (SHapley Additive exPlanations) to provide explainability to the model's decisions. The "Blood Cells Image Dataset," comprising 17,092 expert-annotated images from the Hospital of Barcelona using the CellaVision platform, was used in this study. The model achieved a test accuracy of 97.33% (98.93% in training and 96.13% in validation), with minimal overfitting, as evidenced by the small gap in loss values (0.55 in training vs. 0.60 in testing). Class-wise performance was notable, with eosinophils reaching an F1-score of 0.98 and platelets 0.99. The model also showed strong discriminative capability for challenging morphologies such as lymphocytes and monocytes, with F1-scores of 0.95 and 0.96, respectively. SHAP analysis revealed clinically meaningful patterns: lymphocytes were characterized by high nuclear density and an elevated nucleus-to-cytoplasm ratio, while platelet recognition required fewer features due to their distinctive morphology. Furthermore, the model exhibited high predictive confidence, assigning 94.52% probability to lymphocyte classifications. These findings support the model's potential as a reliable tool for augmenting diagnostic accuracy in hematological assessments.

Título do Evento: 21º Congresso Brasileiro de Bioinformática: X Meeting 2025
Cidade do Evento: João Pessoa
Título dos Anais do Evento: Proceedings of the 21st Brazilian Congress on Bioinformatics (X-Meeting)
Nome da Editora: Even3
Meio de Divulgação: Meio Digital

Como citar

MARCELINO, ALISON HENRIQUE et al.. DEEP LEARNING FOR BLOOD CELL CLASSIFICATION: AN EXPLAINABLE ARTIFICIAL INTELLIGENCE APPROACH FOR DIAGNOSTICS IN HEMATOLOGY.. In: Proceedings of the 21st Brazilian Congress on Bioinformatics (X-Meeting). Anais...João Pessoa(PB) Centro Universitário - UNIESP, 2025. Disponível em: https//www.even3.com.br/anais/xmeeting-2025/1112366-DEEP-LEARNING-FOR-BLOOD-CELL-CLASSIFICATION--AN-EXPLAINABLE-ARTIFICIAL-INTELLIGENCE-APPROACH-FOR-DIAGNOSTICS-IN-. Acesso em: 02/01/2026