Education

2020 – 2025
B.Sc. in Computer Science & Engineering
Sylhet Engineering College (Affiliated with Shahjalal University of Science and Technology)
CGPA: 3.58/4.0
2019
Higher School Certificate
Syed Hatem Ali College, Barisal
GPA: 5.0/5.0
2017
Secondary School Certificate
Barisal Zilla School, Barisal
GPA: 5.0/5.0

Research Interests

• Generative AI
• Deep Learning
• Computer Vision
• Image Processing
• Explainable & Interpretable AI
• Real-time AI Systems
• Medical Imaging & AI for Health Care
• Multimodal Learning
• Trustworthy AI

Conference Papers

• [IEEE SPICSCON 2026]   RefineDiff: Latent Diffusion-Based Feature Refinement for Medical Image Segmentation.
  Sabib Ahmed, et al.
  IEEE SPICSCON: IEEE International Conference on Signal Processing, Information, Communication and Systems 2026
  Submitted
  📄 Details

  DiffRefine is a lightweight, plug-and-play latent diffusion module that enhances medical image segmentation by performing iterative denoising directly in the bottleneck feature space. Unlike conventional diffusion approaches that require architectural redesign or operate in pixel space, it integrates seamlessly into existing encoder–decoder models. This latent-space refinement improves structural consistency and boundary delineation, leading to more complete segmentation. Extensive experiments show consistent performance gains across multiple architectures, highlighting its generalizability and practical efficiency.
• [IEEE ICEFronT 2026]   Parameter-Efficient Fine-Tuning with Adapters for Sorghum Disease Classification.
  Sabib Ahmed, et al.
  ICEFronT: IEEE International Conference on Engineering and Frontier Technologies 2026 (ICEFronT)
  Under Review
  📄 Details

  This paper introduces PEFT-B3, one of the first systematic applications of adapter-based parameter-efficient fine-tuning (PEFT) for sorghum disease classification, showing that large pretrained vision models can be adapted to agricultural disease detection without full fine-tuning. Its novelty lies in demonstrating that training only lightweight adapters (15.58% of parameters) can match full fine-tuning performance (99.58% accuracy) while dramatically reducing computational cost, making deployment feasible in resource-constrained agricultural settings. The strength of the work is not just high accuracy, but the broader idea that parameter-efficient adaptation can replace conventional full-model retraining in agricultural AI, introducing a scalable and deployable paradigm largely unexplored in this domain.
• [IEEE ICEFronT 2026]   Ensemble-ASNet: An Optimized Weighted CNN Ensemble for Multi-Class Kidney Lesion Classification Using CT Images
  Sabib Ahmed, et al.
  ICEFronT: IEEE International Conference on Engineering and Frontier Technologies 2026 (ICEFronT)
  Under Review
  📄 Details

  This work proposes Ensemble-ASNet, a novel optimized weighted ensemble framework that improves multi-class kidney lesion classification by learning optimal model contributions through constrained F1-score–driven soft voting, rather than relying on equal or heuristic fusion. By selectively combining complementary representations from InceptionV3, VGG16, and MobileNetV2, the model reduces overfitting, class imbalance, and single-model bias while achieving highly robust performance. Unlike prior approaches, it tightly integrates performance-aware ensembling with Grad-CAM++–based interpretability, ensuring that predictions are both highly accurate and clinically meaningful, making it a reliable and practically deployable diagnostic system.
• [IEEE QPAIN 2026]   Evaluating Stable Diffusion with LoRA for Class-Imbalanced Medical Image Augmentation.
  Sabib Ahmed, et al.
  2026 IEEE 2nd International Conference on Quantum Photonics, Artificial Intelligence & Networking (QPAIN)
  Accepted
  📄 Abstract

  Deep learning models for medical image analysis are frequently constrained by severe class imbalance and data scarcity, especially in clinical settings where positive cases are rare. Although GAN-based synthesis and classical augmentation have been investigated to address these problems, they often have poor scalability under real-world restrictions, low diversity, and training instability. In this study, we thoroughly assess text-to-image Stable Diffusion as a useful data augmentation technique for class-imbalanced medical imaging. Using renal stone CT imaging as a case study, we optimize Stable Diffusion with Low-Rank Adaptation (LoRA) to produce high-quality synthetic samples exclusively for the minority class. Instead of suggesting a new classifier, we evaluate the created data's usefulness using many standardized downstream evaluation backbones, such as CNN and transformer-based models, to guarantee robustness among representation learning paradigms. Comprehensive tests and ablation studies on synthetic-to-real data ratios show that while excessive dependence on synthetic data worsens performance due to distributional bias, moderate diffusion-based augmentation (25–50%) consistently increases recall and F1-score. Qualitative analysis and Fr'echet Inception Distance enhance the realism of the generated visuals. Overall, this study demonstrates that, under realistic data and computational restrictions, Stable Diffusion, when modified via LoRA, can function as an efficient but limited data augmentation technique for class-imbalanced medical imaging.
• [IEEE ICECTE 2026]   Dual-Stage Vision Transformer (DS-ViT) with Attention-Guided Patch Aggregation for Alzheimer's Disease Classification
  Sabib Ahmed
  5th International Conference On Electrical, Computer, & Telecommunication Engineering (ICECTE)
  Accepted
  📄 Abstract

  Alzheimer's disease (AD) is a progressive neurological illness that needs to be diagnosed early and accurately in order to be effectively treated. This study proposes a novel Dual-Stage Vision Transformer (DS-ViT) architecture that combines CNN-based local patch embedding with Transformer-based global attention to enable the model to capture both fine-grained structural information and inter-regional interdependence within brain MRI scans. An Alzheimer's MRI dataset with four stages of dementia was utilized to train and assess the proposed DS-ViT model. An early stopping strategy was used to reduce overfitting; as validation performance plateaued, training was stopped at the 83rd epoch. According to experimental data, DS-ViT outperforms the standard ViT-Base/16 and traditional CNNs (VGG16, MobileNetV2, InceptionV3), achieving an accuracy of 95.40%. Additionally, the model regularly demonstrates strong and dependable classification performance across all classes with excellent precision, recall, and F1-scores. An important development in AI-driven diagnostic systems, the DS-ViT framework provides a flexible and modular design that may be applied to further medical imaging and neurodegenerative disease detection tasks.
  [Paper]
• [ICERIE 2025]   An Innovative Approach to Smoker Status Prediction with ML Algorithms in Bangladesh
  Sabib Ahmed, et al.
  8th International Conference on Engineering Research, Innovation and Education (ICERIE 2025)
  Accepted
  📄 Abstract

  In Bangladesh, smoking continues to be a significant issue for public health. As for the divisions in age, currently, about 18% of adults smoke, together with older and younger men, including 36.2% of men and 0.8%. Moreover, smoking, including both the use of non-combustible tobacco and different shades of age, particularly smokeless tobacco products, including betel quid and zarda, are significantly common among middle-aged to elderly men and women. Such an extensive consumption pattern leads to significant health complications, contributing to more than 100,000 deaths in the country. About 7% of teenagers use smoking and tobacco products, hinting at negligence of youth tobacco use. In order to approximate and forecast smoking patterns, this study approaches the possibility and effectiveness of some machine learning algorithms for predicting the smoking status of patients based on their age, height, weight, waist circumference, visual acuity in both eyes, hearing capability in both ears, systolic and relaxation. We use seven machine learning algorithms: Naïve Bayes, logistic regression classifier, random forest classifier, XGBoost classifier, Decision Tree , KNeighbors and a new approach called HEStacked to classify the smoking status of patients. Each algorithm is tested with different efficiency measures, such as accuracy, and confusion matrices, after being trained on a balanced dataset. Result illustrates the effectiveness of the HEStacked Classifier, which showed an impressive precision of around 88.03% in correctly classifying smoker status. This model highlights the important role of training and targeted intervention measures, which will be a great asset for public health activities aimed at the prevention and problem of diseases associated with smoking in more general. These findings can be used as a promising tool for the prediction of smoking status based on health risk perceptions of Bangladesh.
  [Paper]

Peer-Reviewed Journal Article

• [Springer SIVP (Q2)]   Clinically Explainable Cross-Attention CNN–ViT Framework for Automated Multi-Class Retinal Disease Screening Using OCT
  Sabib Ahmed, et al.
  Signal, Image and Video Processing, Springer
  Schimago Ranking: Q2
  Under Review
  📄 Details

  This paper proposes a cross-attention–based CNN–ViT hybrid framework that is novel in how it actively fuses local lesion features and global contextual information, rather than simply combining them through standard concatenation. By allowing CNN features (fine-grained retinal details) to dynamically attend to transformer representations (global structure), the model captures subtle inter-class differences that conventional architectures miss. Its second key innovation is a unified explainability mechanism (Grad-Attention Map) that merges Grad-CAM and transformer attention into a single, clinically interpretable visualization, providing both precise lesion localization and global reasoning in one view. Validated on a relatively complex 8-class OCT classification task, the framework stands out for jointly advancing feature fusion, interpretability, and clinical reliability, making it a strong and practically meaningful contribution beyond incremental accuracy gains.
• UDF-Net: Uncertainty–Diversity Guided Multi-Level Feature Fusion for Skin Cancer Classification
  Sabib Ahmed
  Manuscript in Preparation, 2026
  Under Review
  📄 Details

  Developed UDF-Net, a dual-backbone framework for skin cancer classification that uses multi-level feature fusion with confidence- and diversity-guided weighting to improve performance and reliability, along with uncertainty estimation for better model calibration and interpretability.

Undergraduate Thesis

• A Study on Real-Time Pavement Crack Detection Using YOLOv9: Challenges and Opportunities
  Role: Lead Researcher
  📄 Abstract

  For efficient road maintenance and infrastructure safety, pavement crack detection is essential. Nevertheless, manual inspection techniques are time-consuming, prone to mistakes, and unsuitable for extensive monitoring. An automated crack detection system utilizing the cutting-edge deep learning model YOLOv9 is presented in this work. Four different types of cracks were identified in the 1,200 annotated pavement images that made up the custom dataset: pothole, longitudinal, transverse, and alligator. Real-world variations in road conditions, surface textures, and lighting environments are captured in the dataset. This dataset was used to train and assess YOLOv9, which produced an F1-score of 73.2% and a mean Average Precision (mAP) of 70.4%. These outcomes reveal how well YOLOv9 performs in providing precise, scalable, and instantaneous crack detection. By facilitating prompt and well-informed decision-making by road authorities and civil engineers, the suggested system has the potential to greatly improve road safety and maintenance workflows.

Applied ML Projects

• Face Mask Detection
  Developed a YOLOv8-based object detection model capable of identifying three categories — mask, no mask, and bad mask — from both images and videos. The system can be used for real-time monitoring in public spaces.
  Tools: Python, PyTorch, OpenCV, Google Colab
• Brain Tumor Classification
  Designed a CNN model to classify brain tumors from CT and MRI images. Deployed as a Streamlit web app for real-time prediction and visualization.
  Tools: Python, TensorFlow, NumPy, Pandas, OpenCV, Streamlit
• Vehicle Detection
  Implemented a YOLOv12 model to detect 17 distinct vehicle types in images and videos. Suitable for traffic analysis, surveillance, and transportation monitoring.
  Tools: Python, PyTorch, OpenCV, Google Colab

Technical Skills

Languages
Python, C++, C, Java, JavaScript, HTML, CSS
ML & Deep Learning
PyTorch, TensorFlow, Keras, Scikit-learn, CUDA, XGBoost
Computer Vision
OpenCV, YOLO, CNN, Transformers, Roboflow
Data Science
Pandas, NumPy, Matplotlib, Seaborn, MySQL
Tools
VS Code, Jupyter Notebook, Anaconda, Git, GitHub, Linux (Ubuntu), LaTeX

Contact