BSc in CSE
BRAC University
Computer Science Engineer | Data Science & AI Enthusiast
Dhaka, Bangladesh
Hi, I'm Zulkar Nain.
Computer Science Engineer exploring Data Science, AI, Computer Vision, and Healthcare Analytics.
I am interested in applying computational methods and machine learning to real-world problems, especially in healthcare and data-driven systems.
About Me
Building thoughtful systems at the intersection of computing and healthcare.
I am a Computer Science & Engineering graduate from BRAC University with interests in Data Science, Machine Learning, Computer Vision and Healthcare Analytics.
My academic and practical experience has involved building software systems, experimenting with machine learning models, working with computer vision techniques, and exploring how data-driven approaches can address real-world problems.
I am particularly interested in healthcare applications of AI and data science, with a specific focus on responsible and meaningful use of computational methods.
My professional background includes work as a Production Approval Analyst / Order Management professional, along with freelance experience on Fiverr and Upwork where I collaborated with international clients and delivered project-based work independently.
VC's List
Recognized during undergraduate studies
4+ Years
Freelancing experience
Experience
Professional experience shaped by analysis, coordination, and independent delivery.
IDL Bangladesh
My current work focuses on order management, production approval, operational coordination, and maintaining accuracy in business workflows. I value structured communication and reliable execution in collaborative environments.
CVIS Research Lab, BRAC University
Conducted deep analysis of structured medical image datasets from WHO and achieved 98% AUC in early cervical cancer detection using deep learning and machine learning approaches.
Fiverr • Upwork
I have worked with international clients on freelance projects, built software and web solutions, communicated directly with clients, and delivered work independently over several years. My Fiverr work has exceeded $10,000 over approximately four years.
Research & Academic Work
Selected research contributions in computer vision, healthcare AI, and applied machine learning.
Enhancing VIA Screening for Cervical Cancer: A Comprehensive System Integrating Image Processing, Risk Factors, and Follow-up Facilitation
This research explored the application of computer vision and machine learning to Visual Inspection with Acetic Acid (VIA) screening, with emphasis on medical image analysis, risk-factor modeling, and decision-support workflows.
The study examined methods including VGG16, ResNet50, YOLO, and Random Forest to support healthcare analytics and improve the reliability of screening-oriented AI systems in clinical research contexts.
View ThesisMedical Image Analysis
Computer Vision
Machine Learning
Healthcare AI
Clinical Decision Support
Conference presentations and manuscripts under publication
-
ViTamin-XAI: A Hybrid Vision-Transformer Network with Local Interpretable Model-Agnostic Explanations for Robust and Interpretable Diabetic Retinopathy Screening
Vision Transformer • Explainable AI • Diabetic Retinopathy
Presented at ICECI 2026, IEEE; online copy pending. -
Teaching Machines to Feel Faster: DistilBERT for Eleven-Way Emotion Classification
NLP • DistilBERT • Emotion Classification
Presented at ICBDAIA 2026; online copy pending.
Peer-reviewed publications and accepted revisions
-
UniFusion: Unified Dense-Attention Framework for Efficient Multimodal Medical Image Fusion in Brain Tumor Classification
Multimodal Medical Imaging • Attention Mechanisms • Brain Tumor Analysis
Accepted with minor revision — PLOS One -
Attention-Gated DenseNet with Class-Weighted Focal Loss for Multi-Class Cervical Cancer Cell Classification on Pap-Smear Images
Deep Learning • Medical Imaging • Cervical Cancer Screening
Accepted with minor revision — Scientific Reports (Nature)
Healthcare Analytics
Using data science to understand healthcare-related problems and support thoughtful analysis.
Computer Vision
Image understanding, object detection, and visual analysis for applied research problems.
Machine Learning
Predictive modeling, classification, and feature analysis for structured and image-based data.
Health Data Science
Applying statistical and computational methods to healthcare data with care and rigor.
AI for Healthcare
Exploring responsible applications of AI in medical and healthcare contexts.
Medical Imaging
Developing interpretable and clinically relevant imaging systems for diagnostic support and healthcare AI applications.
Diabetic Retinopathy (DR) remains a leading cause of preventable blindness globally. While deep learning has achieved expert-level performance in diagnosis, the “black-box” nature of these models hinders clinical adoption. Furthermore, pure Vision Transformers (ViTs) often require massive datasets to learn local features efficiently. This paper introduces ViTaminXAI, a novel framework leveraging the ViTamin architecture — a hybrid model combining Mobile Convolution Blocks (MBConv) and Transformer blocks — to achieve state-of-the-art performance in binary DR classification. We integrate Local Interpretable Model-agnostic Explanations (LIME) to provide pixel-level interpretability through perturbation-based analysis. Experimental results on an augmented retinal dataset demonstrate a validation accuracy of 99.44%, with a test accuracy of 99.44%, significantly outperforming traditional CNN baselines. The LIME visualizations validate that the model correctly identifies pathological features such as microaneurysms and hemorrhages, ensuring transparency and trustworthiness for medical deployment. The model contains 86.98 million trainable parameters and converges in approximately 7.14 hours on dual-GPU hardware.
Index Terms: Diabetic Retinopathy, Vision Transformers, Hybrid Architecture, Explainable AI, LIME, Deep Learning, Medical Image Analysis.
In this work, focused on multi-class emotion recognition on the Text Emotion dataset of ~106,000 English sentences annotated with 11 emotions. A number of pretrained transformer encoders — in particular DistilBERT-base, BERT-base and RoBERTa-base — are fine-tuned and compared with classic as well as ensemble baselines such as k-NN, logistic regression, naive Bayes classifiers, decision trees, AdaBoost and gradient boosting. All models are trained by subword tokenization and supervised cross-entropy, evaluated with accuracy, weighted precision, recall, F1-score and confusion matrices. DistilBERT-base attains the highest performance overall — 0.89 accuracy, 0.89 weighted F1-score on the test set — and yields a lower computational budget compared to the larger counterparts. Per-class results demonstrate high robustness on frequent emotions (neutral, enthusiasm, hate and relief) and solid effort for challenging classes, making lightweight transformers feasible for real-world applications.
Keywords: emotion recognition, text emotion classification, transformer models, DistilBERT, efficient NLP, social media analytics, sentiment analysis.
Deep learning-based brain tumor classification benefits substantially from fusing complementary diagnostic information from CT and MRI scans, but current state-of-the-art methods are limited by large parameter counts (>8M), slow inference (>70ms per image), and training instability from adversarial optimization. We introduce UniFusion, a parameter-efficient convolutional neural network that leverages dense connectivity for improved gradient flow, incorporates a dual-channel Convolutional Block Attention Module (CBAM) for spatial-channel feature refinement, and utilizes multi-strategy fusion combining maximum, additive, and learned adaptive weighting operations. UniFusion was rigorously evaluated on 4,318 paired CT-MRI brain images (stratified by patient: 70% training, 15% validation, 15% test) aggregated from the Cancer Imaging Archive, BraTS2021, and institutional cohorts (IRB Protocol #2024-AI-378). In systematic comparisons against four representative models (UNetFusion, Temporal Decomposition Network, ResNetFusion, DeepFuse) under uniform training protocols, UniFusion achieved 95.67% test accuracy (95% CI: 93.84–97.50%), 96.15% precision, 95.24% recall, and 95.69% F1-score — statistically equivalent to the most recent state-of-the-art (McNemar's test, p = 0.78) — while reducing parameters by 50.4% (4.24M vs. 8.55M), accelerating inference by 38.7% (46ms vs. 75ms on NVIDIA Tesla P100), and ensuring stable training without adversarial complexity. Ablation studies demonstrate synergistic architectural contributions: dense connectivity (91.82% baseline accuracy), CBAM attention (+3.41%), and multi-strategy fusion (+0.44%). These results establish UniFusion as a clinically deployable solution for robust, real-time multimodal brain tumor classification.
Cervical cancer remains a leading cause of cancer-related mortality among women worldwide, despite being largely preventable through early cytological screening. The Papanicolaou (Pap) smear test is the cornerstone of cervical-cancer screening programs but is constrained by labor-intensive manual interpretation, inter-observer variability, and a non-trivial false-negative rate. Computer-aided diagnosis (CAD) systems built upon deep convolutional neural networks (CNNs) offer a promising solution, but existing approaches typically rely on either heavy parameter-rich backbones unsuited for deployment in resource-constrained settings, or use loss functions that treat all training samples equivalently and therefore fail to address inherent difficulty variation across cervical-cell morphologies. In this work, we propose a lightweight Attention-Gated DenseNet121 (AG-DenseNet) trained with a Class-Weighted Focal Loss (CW-FL) for fine-grained classification of cervical cells into five clinically meaningful categories. We integrate Convolutional Block Attention Module (CBAM)-style channel- and spatial-attention gates after each dense block of DenseNet121, introducing fewer than 0.15 million additional parameters while explicitly suppressing background clutter and amplifying nucleus–cytoplasm regions critical for cytological grading. We adopt a class-weighted focal loss formulation that combines inverse-frequency alpha-balancing with a focusing parameter, jointly down-weighting easy examples and steering optimization toward morphologically ambiguous classes. Experiments on the SIPaKMeD dataset comprising 4,049 single-cell images demonstrate that the proposed method attains a test accuracy of 97.20%, a macro-F1 of 97.22%, Cohen's kappa of 0.9650, Matthews correlation coefficient of 0.9651, and a macro one-versus-rest area-under-ROC of 0.9988, while requiring only 7.26 million parameters. The model achieves perfect classification of Parabasal cells and exceeds 98% AUC on every class. Calibration analysis yields an expected calibration error of 0.0417, indicating well-calibrated confidence estimates suitable for clinical triage. Qualitative attention-map evidence confirms that the proposed gating mechanism concentrates on diagnostically relevant cellular regions. The framework offers a favorable accuracy-efficiency trade-off and constitutes a practical building block for automated cervical-cancer screening pipelines.
Keywords: cervical cancer, Pap smear, SIPaKMeD, deep learning, DenseNet, attention mechanism, focal loss, medical image classification.
Projects
Selected public repositories and research-focused projects from GitHub.
Enhancing VIA Screening for Cervical Cancer
ResearchUndergraduate thesis project using computer vision models and regression analysis to improve screening workflows in healthcare.
MediAid
Healthcare SoftwareA healthcare-focused software project exploring technology-driven approaches to improving access to and management of healthcare-related information.
PCOS Risk Predictor
Healthcare AnalyticsThis project aims to detect at-risk PCOS patients to support better healthcare service and prioritization for assessment.
ML Deployment with FastAPI
DeploymentA project that serves an ONNX-optimized image classification model as a web service with FastAPI, Docker, and Kubernetes.
FileDao-P2P
Web AppA peer-to-peer file sharing application built with Python to enable direct file transfer across devices on the same network.
Sobuy
RepositoryA complete e-commerce platform from scratch using Python Flask, establishing a robust two-tier architecture for Admin and Customer roles.
Skills
Technologies I work with across research, development, and applied data work.
Programming & Web
- Python
- HTML
- CSS
- JavaScript
- Django
- Flask
- Bootstrap
Data Science & Machine Learning
- Machine Learning
- Computer Vision
- VGG16
- ResNet50
- YOLO
- Random Forest
- SVM
- Logistic Regression
Data & Development Tools
- MySQL
- Git
- GitHub
- Docker
- Power BI
- Cursor
- Railway
- DigitalOcean
- Kaggle
- Vercel
Education
Academic foundation in computing with a strong interest in applied research.
BRAC University
Bachelor of Science in Computer Science & Engineering
My education strengthened my interest in computing, applied analytics, machine learning, and data-driven problem solving. I also received recognition on the VC's List during undergraduate study.