A hybrid explainable deep learning framework for blood cancer classification using CNN-based feature embeddings and random forest decision models
5 affiliations
- 1AIML Department, Symbiosis Institute of Technology, Symbiosis International (Deemed University)
- 2Computer Science and Engineering, Pranveer Singh Institute of Technology
- 3Department of Artificial Intelligence and Machine Learning, Pranveer Singh Institute of Technology
- 4Department of Computer Science and Engineering, Koneru Lakshmaiah Education Foundation
- 5IT Department, UIET CSJM University
Abstract
The precision and early detection of subtypes of acute lymphoblastic leukaemia (ALL) in peripheral blood smear images are crucial for efficient clinical practice. Traditional deep learning methods tend to be challenging in terms of model interpretation and are often reliant on large-scale annotated datasets, which compromises their practicality in diagnostic applications. To address the gap in diagnostic practicability, a novel hybrid and explainable deep learning architecture is presented, combining a convolutional neural network (CNN) model with a Random Forest (RF) classifier within a latent space framework. With a custom-trained CNN model having three convolution and pooling layers in a four-class ALL dataset consisting of 3242 images, a validation accuracy of 76.35% is obtained. The final 256-dimensional latent space representation in the CNN model is used as a representation space for an RF classifier to boost overall accuracy to 83% and significantly improve the discriminant capability of malignant subtypes. Moreover, to analyse model diagnosis in a transparent and interpretable way, a novel multi-explainability framework is incorporated, combining the advantages of Grad-CAM, explaining model behaviour in spatial images, with other methods, including LIME and SHAP, to uncover model explanations in the deep latent space. By the use of Grad-CAM, highlighting morphologically critical areas in images like the nucleus and cytoplasm boundary in blood smear images is achieved. In addition, using LIME and SHAP, attributing model diagnostic capability to latent variables in deep learning is shown to harness sparse latent space variables in model diagnosis to significantly improve model practicability in clinical applications. The proposed approach is treatable efficiently, has better model practicability in generalised applications compared to standalone deep learning models, and is entirely transparent in diagnostic applications.
Related
- An Explainable Disease-Aware Intelligent Histopathological Representation Learning Framework for Generalizable Breast Cancer Classification
- Hybrid deep learning–driven explainable AI framework for fault detection and classification in smart power grids
- Explainable computational framework for subtype classification and biomarker identification in head and neck squamous cell carcinoma using machine learning
- Explainable machine learning for breast cancer prediction in resource-constrained settings: A multi-algorithmic framework integrating shap-based transparency with clinical decision support
- CERVEX: A Foundation-Model Framework With Built-In Explainable AI for Automated Cervical Cytology Classification From Pap Smear Images
BibTeX
@article{ansari2026hybrid,
title = {{A hybrid explainable deep learning framework for blood cancer classification using CNN-based feature embeddings and random forest decision models}},
author = {Zulfikar Ali Ansari and Hemlata Pant and Nayancy and M. N. V. Kiranbabu and Sanjeet Kumar},
journal = {Scientific Reports},
year = {2026},
month = jul,
doi = {10.1038/s41598-026-63617-6},
url = {https://doi.org/10.1038/s41598-026-63617-6},
note = {CC BY},
}
RIS
TY - JOUR
TI - A hybrid explainable deep learning framework for blood cancer classification using CNN-based feature embeddings and random forest decision models
AU - Zulfikar Ali Ansari
AU - Hemlata Pant
AU - Nayancy
AU - M. N. V. Kiranbabu
AU - Sanjeet Kumar
PY - 2026
DA - 2026/07/01
JO - Scientific Reports
DO - 10.1038/s41598-026-63617-6
UR - https://doi.org/10.1038/s41598-026-63617-6
AB - The precision and early detection of subtypes of acute lymphoblastic leukaemia (ALL) in peripheral blood smear images are crucial for efficient clinical practice. Traditional deep learning methods tend to be challenging in terms of model interpretation and are often reliant on large-scale annotated datasets, which compromises their practicality in diagnostic applications. To address the gap in diagnostic practicability, a novel hybrid and explainable deep learning architecture is presented, combining a convolutional neural network (CNN) model with a Random Forest (RF) classifier within a latent space framework. With a custom-trained CNN model having three convolution and pooling layers in a four-class ALL dataset consisting of 3242 images, a validation accuracy of 76.35% is obtained. The final 256-dimensional latent space representation in the CNN model is used as a representation space for an RF classifier to boost overall accuracy to 83% and significantly improve the discriminant capability of malignant subtypes. Moreover, to analyse model diagnosis in a transparent and interpretable way, a novel multi-explainability framework is incorporated, combining the advantages of Grad-CAM, explaining model behaviour in spatial images, with other methods, including LIME and SHAP, to uncover model explanations in the deep latent space. By the use of Grad-CAM, highlighting morphologically critical areas in images like the nucleus and cytoplasm boundary in blood smear images is achieved. In addition, using LIME and SHAP, attributing model diagnostic capability to latent variables in deep learning is shown to harness sparse latent space variables in model diagnosis to significantly improve model practicability in clinical applications. The proposed approach is treatable efficiently, has better model practicability in generalised applications compared to standalone deep learning models, and is entirely transparent in diagnostic applications.
SN - 2045-2322
ER -