Md Shihab Reza - Research

Attention and Residual Mechanism-based CNN Architecture (ARC-Net) with Enhanced Fairness Generalization for Deepfake Facial Image Detection

This research aims to propose a novel Attention and Residual-enhanced Convolutional Neural Network, named ARC-Net, for efficient DeepFake image detection. ARC-Net is designed not only to achieve high performance but also to provide meaningful attention to crucial facial regions often manipulated in DeepFakes. To address the geographical bias commonly found in existing DeepFake datasets, this study has collected 500 facial images from Bangladeshi individuals, promoting dataset diversity and inclusivity.
Authors: Md Shihab Reza, Farhana Elias, Monirul Islam Mahmud, Nova Ahmed

Publication Status

Under Review (PLOS One):
[1] Md Shihab Reza, Farhana Elias, Monirul Islam Mahmud, and Nova Ahmed. 2025. Attention and Residual Mechanism–based CNN Architecture (ARC-Net) with Enhanced Fairness Generalization for Deepfake Facial Image Detection.

Code Availability: Github

DFU_DIALNet: Towards Reliable and Trustworthy Diabetic Foot Ulcer Detection with Synergistic Confluence of Grad-CAM and LIME

This study introduces DFU-DIALNet, a novel framework for Diabetic Foot Ulcer detection that achieves 99.33 % accuracy on a merged dataset (500 collected images + DFUC2021). Tested on DFUC2020 and Kaggle DFU, it generalizes to unseen ulcerated and healthy feet. By integrating Grad-CAM and LIME, we enhance interpretability, and a real-time web app demonstrates its reliability and practical deployability.
Authors: Monirul Islam Mahmud, Md Shihab Reza, Mohammad Olid Ali Akash, Farhana Elias, Nova Ahmed.

Publication Status

[1] Mahmud MI, Reza MS, Akash MOA, Elias F, Ahmed N (2025) DFU_DIALNet: Towards reliable and trustworthy diabetic foot ulcer detection with synergistic confluence of Grad-CAM and LIME. PLoS One 20(9): e0330669. https://doi.org/10.1371/journal.pone.0330669

Code Availability: Github

Efficient Cloud Segmentation and Multiregional Inclusive Smartphone Dataset

This research, conducted in collaboration with UCD and Dr. Soumyabrata Dev, develops a deployable, efficient U-Net–based computer vision model for cloud segmentation. For the first time in remote sensing history, we are creating a multi-regional smartphone-captured dataset, enabling broad geographic coverage and inclusivity.

Authors: Ifran Rahman Nijhum, Md Shihab Reza and Soumyabrata Dev.

Publication Status

Accepted:
[1] Ifran Rahman Nijhum, Md. Shihab Reza, and Soumyabrata Dev. (2025). TINY-UCLOUDNET: Memory-Efficient Ground-Based Cloud Segmentation Network. In International Geoscience and Remote Sensing Symposium (IGARSS),
Under Review (IEEE Geoscience and Remote Sensing Letters):
[1] Ifran Rahman Nijhum, Md Shihab Reza, and Soumyabrata Dev. 2025. Squeeze-Expand-Excite: Lightweight U-Net for Cloud Segmentation with a Multi-Regional Nighttime Smartphone Dataset.

Code Availability: Github

A Novel Clinical Dataset with XAI-Integrated Ensemble and Large Language Models for Interpretable Dengue Detection: A Comparative Analysis of Reduced-Feature AI Classifiers

This research aims to construct a novel clinical dataset with a minimized feature set for dengue detection and to evaluate both traditional machine learning algorithms and large language models (LLMs) on this data. In dengue-endemic regions where patient records are scarce, these specialized ML and LLM approaches are critical for robust performance on feature-reduced datasets. We further apply global XAI techniques, Permutation Feature Importance and Morris Sensitivity Analysis to elucidate each feature’s contribution across the entire cohort, and pioneer the use of LLMs in dengue research for the first time.
Authors: Muhammad Zawad Mahmud, Md. Shihab Reza, Shahran Rahman Alve, Samiha Islam, and Riasat Khan.

Publication Status

Under Review (Array):
[1] Muhammad Zawad Mahmud, Md Shihab Reza, Shahran Rahman Alve and Samiha Islam 2024. A Novel Clinical Dataset with XAI-Integrated Ensemble and Large Language Models for Interpretable Dengue Detection: A Comparative Analysis of Reduced-Feature AI Classifiers

Explainable Machine Learning for Fair and Transparent Trustworthy Credit Scoring

This research develops a transparent hybrid credit-scoring model by integrating XGBoost and deep neural networks with Linear Discriminant Analysis for feature reduction and benchmarking it against six classical algorithms. It pioneers a dual-layer interpretability framework, using LIME for local explanations and Morris Sensitivity Analysis for global feature insights, to ensure fairness, transparency, and trust in high-stakes lending. We are currently conducting multi-domain validation of Bayesian-optimized stacking ensembles to deliver a next-generation credit-risk model with granular explainability and robust statistical inference across diverse borrower populations.
Authors: Md Shihab Reza, Monirul Islam Mahmud, Farhana Elias , Ifti Azad Abeer, and Nova Ahmed.

Publication Status

Published:
[1] Md Shihab Reza, Monirul Islam Mahmud, Ifti Azad Abeer, and Nova Ahmed. 2024. Linear Discriminant Analysis in Credit Scoring: A Transparent Hybrid Model Approach. International Conference on Computer and Information Technology (ICCIT). https://doi.org/10.1109/ICCIT64611.2024.11022149
Under Review (SN Computer Science):
[1] Monirul Islam Mahmud, Md Shihab Reza, Farhana Elias, Ifti Azad Abeer, and Nova Ahmed. 2025. Multi-Domain Validation of Bayesian Optimized Stacking Ensembles for Next-Generation Credit Risk Modeling with Granular Explainability and Robust Statistical Inference. Preprints.

Available online as a preprint at Preprints. https://www.preprints.org/manuscript/202510.0484/v1

Optimizing Stroke Detection: An Analysis of Different Feature Selection Approaches

This research aims to optimize automated stroke detection by systematically comparing filter, wrapper, and embedded feature-selection approaches and integrating the most effective subset into classification models. By benchmarking six machine-learning algorithms on both full and reduced feature sets, we identify a lean feature panel that achieves high diagnostic accuracy with lower computational overhead. We further employ global explainability techniques, including Permutation Feature Importance and SHAP analysis, to clarify each feature’s impact across patient cohorts and ensure an interpretable, trustworthy detection pipeline.
Authors: Monirul Islam Mahmud, Md Shihab Reza, Syeda Shabnam Khan.

Publication Status

Published:
[1] Monirul Islam Mahmud, Md Shihab Reza, and Syeda Shabnam Khan. 2024. Optimizing Stroke Detection: An Analysis of Different Feature Selection Approaches. In Companion of the 2024 ACM International Joint Conference on Pervasive and Ubiquitous Computing (UbiComp Companion ’24). https://dl.acm.org/doi/10.1145/3675094.3677602

Machine Learning Meets Transparency in Osteoporosis Risk Assessment: A Comparative Study of ML and Explainability Analysis

This research aims to evaluate six machine learning classifiers, including Random Forest, Logistic Regression, XGBoost, AdaBoost, LightGBM and Gradient Boosting, on clinical, demographic and lifestyle data for early osteoporosis risk assessment with hyperparameters optimized via GridSearchCV. It applies SHAP, LIME and Permutation Feature Importance to elucidate model decisions, identifying age, hormonal changes and family history as primary risk predictors XGBoost achieved the highest performance with 91% accuracy, 0.92 precision, 0.91 recall and a 0.90 F1 score, demonstrating a transparent and trustworthy pipeline suitable for clinical deployment
Authors: Farhana Elias , Md Shihab Reza , Muhammad Zawad Mahmud , Samiha Islam , Shahran Rahman Alve.

Publication Status

Published:
[1] Farhana Elias , Md Shihab Reza , Muhammad Zawad Mahmud , Samiha Islam , Shahran Rahman Alve. 2025. Machine Learning Meets Transparency in Osteoporosis Risk Assessment: A Comparative Study of ML and Explainability Analysis. 2025 International Conference on Quantum Photonics, Artificial Intelligence, and Networking (QPAIN). https://doi.org/10.48550/arXiv.2505.00410 & https://doi.org/10.1109/QPAIN66474.2025.11172213

Advance Transfer Learning Approach for Identification of Multiclass Skin Disease with LIME Explainable AI Technique.

This research aims to develop transfer-learning models for multiclass skin-disease detection on the HAM10000 dataset by evaluating five pretrained architectures: ResNet50, InceptionV3, VGG16, VGG19 and MobileNetV2. It benchmarks both accuracy and inference efficiency, identifying ResNet50 as the most accurate with 99% accuracy and MobileNetV2 as a lightweight alternative with 97.5% accuracy. The study provides guidance on selecting clinically deployable models that balance performance and computational constraints.
Authors: Muhammad Zawad Mahmud, Md Shihab Reza, Samiha Islam, and Shahran Rahman Alve.

Publication Status

Accepted:
[1] Muhammad Zawad Mahmud, Md Shihab Reza, Samiha Islam, and Shahran Rahman Alve. 2024. Advance Transfer Learning Approach for Identification of Multiclass Skin Disease with LIME Explainable AI Technique. International Conference on Computer and Information Technology (ICCIT). https://doi.org/10.1109/ICCIT64611.2024.11021908

Unsupervised Domain-Adaptive 3D Segmentation of Alveolar Bone Defects in CBCT Scans

This research aims to investigate unsupervised segmentation of alveolar bone defects from CBCT scans and explore domain adaptation strategies using synthetic data. We will assemble diverse real-patient CBCT datasets alongside simplified synthetic variants to train 3D deep learning models and employ Gradient Reversal Layers, Siamese-network architectures, and backbones such as PointNet, PointNet++ and DGCNN to minimize the domain gap. Finally, we will validate on four unseen natural CBCT volumes with and without domain adaptation to assess defect segmentation performance in the absence of human-annotated 3D ground truth.
Authors: Md Shihab Reza*, Monirul Islam Mahmud*, Muhammad Zawad Mahmud*, and Saif Ahmed.(*These authors contributing equally)

Publication Status

Work on Progress:

[1] Md Shihab Reza, Monirul Islam Mahmud, Muhammad Zawad Mahmud, Saif Ahmed. 2025. Unsupervised Domain-Adaptive 3D Segmentation of Alveolar Bone Defects in CBCT Scans.

Research Collaboration

Research Grants & Funding

Page updated

Google Sites

Report abuse