Tauhidur Rahman Sakib Google Scholar

~/research/publications

Publications

15 papers: 11 published across IEEE, Elsevier, IET and SSRN, plus 4 accepted and coming online soon. First author on 3, supervisor and corresponding author on 10, and top contributor on 4. Tap or swipe any card to read its abstract, then open the paper or download the PDF.

IEEEDec 2025
First authorSupervisorCorresponding author
IEEEPublished
Operationally Robust Machine Learning for Pre-Launch Missile Failure Prediction

Abstract—Pre-launch failures are rare, costly and hard to predict from imbalanced sensor logs. This work builds an operationally robust pipeline using SMOTE-based imbalanced learning and ensemble models, evaluated with ROC analysis and strict validation so predictions stay reliable under real operating conditions.

Index Terms—pre-launch prediction, SMOTE, imbalanced learning, ROC analysis.

I. Introduction

Missile reliability depends on catching faults before launch. Failure events are rare, which makes standard classifiers overconfident on the majority class.

Fig. 1.
II. Related Work
III. Method
IV. Results
2025 IEEE International Conference on Agentic AI (ICA)

Operationally Robust Machine Learning for Pre-Launch Missile Failure Prediction

2025 IEEE International Conference on Agentic AI (ICA)

Tap or swipe to read the abstract

IEEEFeb 2026
First authorSupervisorCorresponding author
IEEEPublished
CDFNet Based Fusion Network for Interpretable Alzheimer’s Disease Prediction Using Hybrid Imaging and Clinical Meta-Ensemble Learning

Abstract—Early Alzheimer’s prediction needs evidence from both brain imaging and clinical data. This study presents CDFNet, a multimodal fusion network that combines deep radiomic features from neuroimaging with clinical biomarkers through a meta-ensemble, with interpretability built in for clinical trust.

Index Terms—CDFNet, multimodal fusion, neuroimaging, meta-ensemble, deep radiomics.

I. Introduction

Alzheimer’s disease progresses silently for years. Combining imaging with clinical markers promises earlier and more trustworthy prediction.

Fig. 1.
II. Related Work
III. Method
IV. Results
2026 IEEE 18th International Conference on Advanced Trends in Radioelectronics, Telecommunications and Computer Engineering (TCSET)

CDFNet Based Fusion Network for Interpretable Alzheimer’s Disease Prediction Using Hybrid Imaging and Clinical Meta-Ensemble Learning

2026 IEEE 18th International Conference on Advanced Trends in Radioelectronics, Telecommunications and Computer Engineering (TCSET)

Tap or swipe to read the abstract

IEEEFeb 2026
First authorSupervisorCorresponding author
IEEEPublished
Predicting Super-Class Drug Mechanisms from Large-Scale IC50 Cell-Line Sensitivity Profiles

Abstract—Knowing how a drug acts speeds up discovery. This work predicts super-class mechanisms of action from large-scale IC50 cell-line sensitivity profiles, using autoencoder embeddings, resampling and ensemble learning to reach 91.82% accuracy on phenotypic data.

Index Terms—mechanism of action, IC50, autoencoder, embeddings, SMOTE, ensemble.

I. Introduction

Mechanism-of-action labelling is slow and expensive in the lab. Phenotypic response data offers a scalable computational alternative.

Fig. 1.
II. Related Work
III. Method
IV. Results
2026 IEEE 18th International Conference on Advanced Trends in Radioelectronics, Telecommunications and Computer Engineering (TCSET)

Predicting Super-Class Drug Mechanisms from Large-Scale IC50 Cell-Line Sensitivity Profiles

2026 IEEE 18th International Conference on Advanced Trends in Radioelectronics, Telecommunications and Computer Engineering (TCSET)

Tap or swipe to read the abstract

IEEEJun 2026
SupervisorCorresponding author
IEEEPublished
A Hybrid Dual-Backbone Deep Learning Framework for Lumbar Spinal Stenosis Classification from MRI with External Validation

Abstract—The automated diagnosis of lumbar spinal stenosis from MRI remains challenging due to anatomical complexity. SpineNet-Fusion combines EfficientNet-B3 and DenseNet-121 to capture global and fine-grained features, reaching 97.4% cross-validation accuracy and 92.9% on external clinical data, with Grad-CAM highlighting relevant regions.

Index Terms—lumbar spinal stenosis, MRI, EfficientNet, DenseNet, Grad-CAM.

I. Introduction

Lumbar spinal stenosis is a common cause of back pain and disability. Reading MRI scans varies between clinicians, which motivates robust automated classification.

Fig. 1.
II. Related Work
III. Method
IV. Results
2026 International Conference on Cybersecurity, Digital Forensics, and AI Applications (ICCSDFAI)

A Hybrid Dual-Backbone Deep Learning Framework for Lumbar Spinal Stenosis Classification from MRI with External Validation

2026 International Conference on Cybersecurity, Digital Forensics, and AI Applications (ICCSDFAI)

Tap or swipe to read the abstract

IEEEJun 2026
SupervisorCorresponding author
IEEEPublished
A Hybrid DenseNet–Vision Transformer Framework for Diabetic Retinopathy Classification

Abstract—Diabetic retinopathy is a major cause of vision loss and needs accurate, timely diagnosis. DenseViT-DR combines DenseNet121 and a Vision Transformer for multi-class severity grading from fundus images, reaching 96.33% accuracy on APTOS 2019 and 90.40% on external clinical data.

Index Terms—diabetic retinopathy, DenseNet, vision transformer, fundus images.

I. Introduction

Diabetic retinopathy is among the leading causes of preventable blindness. Manual grading is slow, so automated classification can support earlier diagnosis.

Fig. 1.
II. Related Work
III. Method
IV. Results
2026 International Conference on Cybersecurity, Digital Forensics, and AI Applications (ICCSDFAI)

A Hybrid DenseNet–Vision Transformer Framework for Diabetic Retinopathy Classification

2026 International Conference on Cybersecurity, Digital Forensics, and AI Applications (ICCSDFAI)

Tap or swipe to read the abstract

ElsevierSep 2025
Coding & analysis
ElsevierPublished
Multi-modal predictive modeling of schizophrenia severity: Leveraging liver function indicators and cognitive scores with random forest and SVM

Abstract—Emerging evidence links schizophrenia to hepatic dysfunction via the gut-liver-brain axis. This study proposes a multimodal framework combining liver biomarkers (ALT, AST, bilirubin, albumin, INR) with cognitive scores, modelling severity with Random Forest regression and SVM classification.

Index Terms—schizophrenia severity, liver biomarkers, cognitive assessment, random forest, SVM.

I. Introduction

Schizophrenia involves both cognitive deficits and systemic physiological changes, yet liver markers are rarely used in severity prediction.

Fig. 1.
II. Related Work
III. Method
IV. Results
Psychiatry Research: Neuroimaging

Multi-modal predictive modeling of schizophrenia severity: Leveraging liver function indicators and cognitive scores with random forest and SVM

Psychiatry Research: Neuroimaging

Tap or swipe to read the abstract

IEEEJun 2025
Top contributor
IEEEPublished
Hybrid Email Filtering Using TF-IDF and BERT Enhanced Spam Detector

Abstract—Spam keeps evolving faster than rule-based filters. This work introduces TBESD, a hybrid detector that pairs fast TF-IDF features with contextual BERT representations, achieving close to 99% accuracy with a high AUC on benchmark email data.

Index Terms—spam detection, TF-IDF, BERT, hybrid model, AUC.

I. Introduction

Email remains a primary attack vector. Combining lexical and contextual signals catches spam that either method misses alone.

Fig. 1.
II. Related Work
III. Method
IV. Results
2025 2nd International Conference on Next-Generation Computing, IoT and Machine Learning (NCIM)

Hybrid Email Filtering Using TF-IDF and BERT Enhanced Spam Detector

2025 2nd International Conference on Next-Generation Computing, IoT and Machine Learning (NCIM)

Tap or swipe to read the abstract

SSRN2026
SupervisorCorresponding author
SSRNPublished
A Lightweight Network Matches Larger Models in Cross-Dataset Polyp Segmentation under a Controlled Multi-Seed Protocol

Abstract—Cross-dataset generalization and deployment cost are two largely independent obstacles to the clinical translation of polyp segmentation. We present WaveLite-DG, a compact encoder–decoder addressing both. It combines a MobileNetV3-Large encoder, Wavelet Frequency Modules that perturb approximation-band feature statistics during training while enhancing detail-band structure, a gated context scan at the bottleneck, and a boundary-supervised decoder. We derive in closed form the energy this perturbation injects. Trained on Kvasir-SEG and CVC-ClinicDB, with CVC-ColonDB, ETIS-LaribPolypDB and CVC-300 withheld from every fitting and selection decision, WaveLite-DG reaches a three-seed mean macro unseen Dice of 0.8300 ± 0.0004, the highest of six architectures retrained under one protocol without photometric augmentation and statistically equivalent to Polyp-PVT and VM-UNet within a pre-set margin of 0.02. With colour jitter, or with the perturbation added to those two baselines, they lead by under 0.008. In every regime tested WaveLite-DG stays within 0.008 of the best model with one third or less of the parameters (7.68 M) and one fifth or less of the computation (1.93 giga multiply–accumulate operations), and it is 0.0427 more accurate than the best of four lightweight networks. Approximation-band perturbation adds +0.0651 macro Dice without photometric augmentation but only +0.0185 with colour jitter, which recovers most of the gain, and perturbing the undecomposed feature at matched energy shows no detectable difference. The contribution therefore lies in efficiency at parity rather than in a new mechanism.

Index Terms—polyp segmentation, domain generalization, wavelet transform, lightweight networks.

I. Introduction

Colonoscopy models often fail on hospitals they were not trained on, and large models are costly to deploy.

Fig. 1.
II. Related Work
III. Method
IV. Results
SSRN preprint

A Lightweight Network Matches Larger Models in Cross-Dataset Polyp Segmentation under a Controlled Multi-Seed Protocol

SSRN preprint

Tap or swipe to read the abstract

IEEENov 2025
Top contributor
IEEEPublished
Multi-Class Retinal Disease Detection Using VGG-19 with K-Fold Cross-Validation

Abstract—Retinal diseases need early screening at scale. This study fine-tunes VGG-19 with transfer learning for multi-class retinal disease detection from fundus images, validated with K-fold cross-validation for reliable performance estimates.

Index Terms—retinal disease, VGG-19, transfer learning, fundus images, K-fold.

I. Introduction

Fundus imaging is cheap and widely available, making it ideal for automated screening of common retinal diseases.

Fig. 1.
II. Related Work
III. Method
IV. Results
2025 IEEE International Conference on Biomedical Engineering, Computer and Information Technology for Health (BECITHCON)

Multi-Class Retinal Disease Detection Using VGG-19 with K-Fold Cross-Validation

2025 IEEE International Conference on Biomedical Engineering, Computer and Information Technology for Health (BECITHCON)

Tap or swipe to read the abstract

IETMar 2025
Top contributor
IETPublished
Mental Health Evaluation During Internet Blackouts: A Machine Learning Approach

Abstract—Internet blackouts during the 2024 quota movement in Bangladesh cut people off overnight. This study surveys emotional, behavioural and environmental effects and applies machine learning to evaluate the mental-health impact of sudden disconnection.

Index Terms—internet blackout, mental health, machine learning, survey.

I. Introduction

The internet is now essential to daily life. Its sudden loss during civil unrest creates measurable psychological stress.

Fig. 1.
II. Related Work
III. Method
IV. Results
International Conference on Green Energy, Computing and Intelligent Technology (IET Conference Proceedings)

Mental Health Evaluation During Internet Blackouts: A Machine Learning Approach

International Conference on Green Energy, Computing and Intelligent Technology (IET Conference Proceedings)

Tap or swipe to read the abstract

BRAC2024
Top contributor
BRACPublished
Analyzing Financial Stress and Its Impact on Mental Health Among Bangladeshi Undergraduate Students Using SPSS and SmartPLS

Abstract—Undergraduates face rising tuition, unstable finances and job insecurity. Using 448 survey responses, this study models how financial resources and perceived instability affect stress, well-being and academic performance with SPSS and SmartPLS.

Index Terms—financial stress, mental health, SmartPLS, structural equation modelling.

I. Introduction

Financial pressure is a growing concern for students in Bangladesh, with effects that reach both well-being and grades.

Fig. 1.
II. Related Work
III. Method
IV. Results
BRAC University · Undergraduate thesis

Analyzing Financial Stress and Its Impact on Mental Health Among Bangladeshi Undergraduate Students Using SPSS and SmartPLS

BRAC University · Undergraduate thesis

Tap or swipe to read the abstract

Accepted2026
SupervisorCorresponding author
ManuscriptAccepted
TriFusion-ResViTNet: A Hybrid Feature-Level Fusion Framework for MRI-Based Rare Neurological Disease Classification

Abstract—Rare neurological disorders are hard to classify from MRI because patterns overlap and data is limited. TriFusion-ResViTNet fuses ResNet50 spatial features, Vision Transformer context and radiomics at the feature level for robust classification.

Index Terms—MRI classification, neurological disease, vision transformer, ResNet50, radiomics.

I. Introduction

Rare diseases suffer from small datasets, which makes single-backbone models unstable.

Fig. 1.
II. Related Work
III. Method
IV. Results
Accepted · online soon

TriFusion-ResViTNet: A Hybrid Feature-Level Fusion Framework for MRI-Based Rare Neurological Disease Classification

Accepted · online soon

Tap or swipe to read the abstract

Accepted2026
SupervisorCorresponding author
ManuscriptAccepted
A Gated Three-Stream Fusion Framework for Automated Brain-Stroke Classification from MRI

Abstract—Separating haemorrhagic and ischemic stroke quickly is critical for emergency care. This gated three-stream framework learns ResNet, Vision Transformer and radiomic features and fuses them with gated attention for explainable stroke classification.

Index Terms—stroke classification, MRI, vision transformer, ResNet-50, gated-attention fusion.

I. Introduction

Stroke treatment depends on subtype, and lesion appearance varies widely between patients.

Fig. 1.
II. Related Work
III. Method
IV. Results
Accepted · online soon

A Gated Three-Stream Fusion Framework for Automated Brain-Stroke Classification from MRI

Accepted · online soon

Tap or swipe to read the abstract

Accepted2026
SupervisorCorresponding author
ManuscriptAccepted
Attention-Enhanced Deep Residual Learning for Multi-Stage Liver Fibrosis Classification with External Validation

Abstract—Staging liver fibrosis from ultrasound suffers from operator dependence and device variability. ResFib-SE combines a ResNet50 backbone with squeeze-and-excitation attention for METAVIR staging, validated on external data.

Index Terms—liver fibrosis, ultrasound, ResNet50, squeeze-and-excitation, METAVIR.

I. Introduction

Adjacent fibrosis stages look very similar on ultrasound, which makes manual staging inconsistent.

Fig. 1.
II. Related Work
III. Method
IV. Results
Accepted · online soon

Attention-Enhanced Deep Residual Learning for Multi-Stage Liver Fibrosis Classification with External Validation

Accepted · online soon

Tap or swipe to read the abstract

Accepted2026
SupervisorCorresponding author
ManuscriptAccepted
CBAM-Net: An Attention-Enhanced Lightweight Convolutional Network for Steel Surface Defect Classification with Ablation and Robustness Analysis

Abstract—Automated steel inspection is challenged by irregular defects, complex textures and lighting changes. CBAM-Net, a lightweight CNN with CBAM attention after every block, is studied through ablation and robustness analysis for industrial deployment.

Index Terms—steel surface defects, CBAM, attention, lightweight CNN, Grad-CAM.

I. Introduction

Factory inspection needs models that are both accurate and cheap enough to run on the line.

Fig. 1.
II. Related Work
III. Method
IV. Results
Accepted · online soon

CBAM-Net: An Attention-Enhanced Lightweight Convolutional Network for Steel Surface Defect Classification with Ablation and Robustness Analysis

Accepted · online soon

Tap or swipe to read the abstract