Research ArticleUnder Review

Rule-based BloodCell Interpretable and Actionable Hematological Classification using ConvNeXt-Tiny

A ConvNeXt-Tiny model classifies eight peripheral blood cell types at 0.98 accuracy, paired with LIME visual explanations and a rule-based module that turns each prediction into non-diagnostic clinical guidance.

Muhammad Adnan Tariq, Sheikh Muhammad Saqib, Tariq Shahzad, Muhammad Usman Tariq, Tehseen Mazhar, Habib Hamam

Corresponding author: Tehseen Mazhar

Expert Systems With ApplicationsElsevier2026
Accuracy
0.98
Held-out test set; 0.974 ± 0.003 under 5-fold CV
Weighted F1
0.98
Held-out test set (3,422 images)
MCC
0.974
Cohen's Kappa 0.974
Images
17,107
8 classes, 13,685 train / 3,422 test

Abstract

The automated identification and classification of peripheral blood cells is critical for diagnosing a range of hematological disorders, from systemic infections to leukemia. Such classification supports diagnosis by providing differential cell counts and by detecting abnormal cell morphology, rather than diagnosing a disease directly. Traditional manual microscopy is labor-intensive and prone to subjective error, while existing deep learning "black-box" models often lack the transparency required for clinical adoption. This study proposes an interpretable, research-oriented decision-support framework leveraging the ConvNeXt-Tiny architecture, optimized with tuned hyperparameters and an input resolution of 64×64 pixels to enable rapid, lightweight training. The model was trained on a dataset of 17,107 images, using 13,685 for training and 3,422 for testing across eight classes: Basophils, Eosinophils, Erythroblasts, Immature Granulocytes (IG), Lymphocytes, Monocytes, Neutrophils, and Platelets. On the held-out test set, the model achieved an overall accuracy of 0.98 and a weighted F1-score of 0.98. Reliability was assessed with a Matthews Correlation Coefficient (MCC) of 0.9737 and a Cohen's Kappa of 0.9737, indicating strong agreement with the expert-labeled ground truth. A key contribution is the integration of Explainable AI (XAI) through LIME (Local Interpretable Model-Agnostic Explanations), which highlights influential superpixels to spatially validate cellular morphology. These visual explanations are complemented by a rule-based clinical decision-support (expert-system) module that maps the predicted cell class to predefined, non-diagnostic clinical guidance. By combining deep-learning feature extraction with rule-based recommendations drawn from standard hematology references, the framework provides a transparent and lightweight research prototype intended to assist, rather than replace, expert hematological assessment.

Overview

A lightweight ConvNeXt-Tiny framework for interpretable peripheral blood cell classification, combining LIME-based explainability with a rule-based clinical decision-support module that maps predictions to non-diagnostic clinical guidance.

Key Contributions

  • 1High-capacity blood-cell classification framework based on the ConvNeXt-Tiny architecture, achieving 0.98 accuracy across eight cell classes.
  • 2Unified single-stage pipeline that eliminates dependency on segmentation-first preprocessing.
  • 3Integration of LIME-based XAI for region-level visual explanations of cellular morphology.
  • 4Rule-based clinical decision-support module that maps predicted cell class to predefined, non-diagnostic clinical guidance.
  • 5Rigorous large-scale evaluation on 17,107 peripheral blood cell images with reliability metrics (Cohen's Kappa, MCC, Hamming Loss).

Figures

Pipeline diagram: the blood cell image dataset is normalised and resized to 64 by 64, fed through a ConvNeXt-Tiny network, and the trained model output branches into per-class confidence scores, LIME explanation analysis, and a rule module producing suggestions.
Figure 1. Proposed methodology: images are normalised and resized to 64×64, classified by a ConvNeXt-Tiny backbone, then passed through LIME explanation and a rule-based module that turns the predicted class into non-diagnostic guidance.
A blood smear image passes through the trained ConvNeXt-Tiny model to a LIME view outlining the cell region in yellow, yielding a basophil prediction at 0.87 confidence and a panel of recommended guidance points.
Figure 17. LIME analysis on a basophil image. Highlighted superpixels show the morphology driving a 0.87-confidence prediction, which the rule-based module converts into recommended clinical guidance.

Methodology

ConvNeXt-Tiny backbone trained on 64×64 peripheral blood smear images (13,685 train / 3,422 test) with tuned hyperparameters, followed by LIME-based XAI overlays and a rule-based decision-support module that maps predictions to predefined, non-diagnostic clinical guidance.

Datasets

  • Blood Cells Image Dataset

    17,107 peripheral blood smear images across eight classes: basophils, eosinophils, erythroblasts, immature granulocytes, lymphocytes, monocytes, neutrophils and platelets.

Results & Findings

Overall accuracy of 0.98, weighted F1-score of 0.98, Cohen's Kappa of 0.9737, MCC of 0.9737, and Hamming Loss of 0.0365 across eight blood cell types, with 1.00 confidence scores observed in classes such as Eosinophils and Erythroblasts.

Conclusion

Integrating ConvNeXt-Tiny with LIME-based Explainable AI and a rule-based decision-support module produces a transparent, clinically actionable framework for automated blood cell classification, moving beyond black-box models to support (not replace) expert hematological assessment.

Author Affiliations

  1. 1COMSATS University Islamabad, Sahiwal Campus, Pakistan
  2. 2Gomal University, DI Khan, Pakistan
  3. 3NCBA&E, Lahore, Pakistan
  4. 4Abu Dhabi University, UAE
  5. 5Université de Moncton, Canada
  6. 6University of Johannesburg, South Africa
  7. 7IITG, Gabon
  8. 8Bridges for Academic Excellence - Spectrum, Tunisia

Explore More Publications

Check out my other research publications in AI, machine learning, and emerging technologies.

View All Publications