http://dspace.daffodilvarsity.edu.bd:8080/handle/123456789/16 2026-04-26T02:04:42Z http://dspace.daffodilvarsity.edu.bd:8080/handle/123456789/17043 Deep Transfer Learning Framework for Alzheimer’s Disease Classification with Explainable AI Insights Prema, Sadia Tasnim Alzheimer’s disease (AD) is a progressive neurodegenerative disorder that requires early and reliable diagnosis, especially in low-resource settings. This study proposes an explainable deep learning framework for multi-class classification of AD stages from MRI brain images. A publicly available four-class Mendeley MRI dataset (Non- Demented, Very Mild, Mild, and Moderate-Demented) was preprocessed through resizing, normalization, augmentation, and an 80/10/10 train–validation–test split. Four transfer learning models GoogLeNet, DenseNet121, ResNet101, and VGG16 were fine- tuned using the Adam optimizer and evaluated with accuracy, precision, recall, F1-score, AUC, and mean categorical hinge loss. Among all architectures, GoogLeNet achieved the best performance with 0.98 accuracy, macro-F1 of 0.98, AUC of 1.00, and the lowest hinge loss (0.048), clearly outperforming the other models. Explainable AI techniques, Grad-CAM and LIME, were applied to highlight discriminative brain regions, consistently focusing on clinically relevant structures such as the hippocampus, temporal lobe, and ventricles. These pictorial elucidations uphold the clinical suitability of the model’s decisions. All in all, the suggested GoogLeNet based architecture is a mixture of high strong interpretable diagnostic accuracy, proving that it can be a assistance screening device in MRI-based AD detection under data constrained clinical. such environments as Bangladesh. Thesis Report 2025-12-27T00:00:00Z http://dspace.daffodilvarsity.edu.bd:8080/handle/123456789/17042 Adaptive Multi-Objective Waste Intelligence System: Hierarchical Transfer Learning for Real-time Recyclable Material Recognition with Edge Computing Deployment Hossain, MD Sifat The current waste crisis in the world, which generates over 2.12 billion tons annually, requires highly advanced technological interventions. This thesis introduces AMWIS, a novel hierarchical transfer learning framework for real-time recyclable material recognition that is capable of deployment with edge computing.Problematic: Traditional manual waste sorting is inefficient, error-prone, and very labor-consuming. Binaryclassification and limited multi-class classification models (usually 4-6 categories) cannot reflect thediversity of real-world wastes.Proposed Solution: AMWIS bridges this gap with the deployment of complete 9-class waste classification,namely Cardboard, Food Organics, Glass, Metal, Miscellaneous Trash, Paper, Plastic, Textile Trash, Vegetation, using transfer learning on the Kaggle Waste Classification Dataset composed of 4,000 images. Key Contributions:1. Methodological Innovation: Hierarchical transfer learning combining EfficientNet-B3, MobileNetV3, and Vision Transformer architectures with adaptive ensemble fusion 2. Detailed Classification: 9-class taxonomy reflecting real composition of waste streams 3. Practical Deployment: Edge computing architecture enables real-time inference for resource-constrained devices. 4. Environmental Impact: Quantified resource recovery and circular economy benefits Results: AMWIS achieved 94.7% accuracy on the validation set with an inference time of 156ms per image, hence is ready for production at waste management facilities. Significance: This research connects existing works such as Aral et al. (2022) and Bircanoglu et al. (2019) to the needs of practical deployment. Unique contributions are made on hierarchical learning, ensemble optimization, and edge deployment architecture. Thesis Report 2025-12-27T00:00:00Z http://dspace.daffodilvarsity.edu.bd:8080/handle/123456789/17041 Sequence-Based Prediction of Amyloid Proteins Using a Hybrid CNN- GRU Deep Learning Architecture Anu, Muhsana Saima Amyloid fibrils formed by misfolded proteins are central to the pathology of several neurodegenerative disorders, including Alzheimer’s and Parkinson’s disease. Reliable in silico prediction of amyloidogenic proteins and peptides can greatly reduce experimental burden and guide mechanistic studies. Existing computational tools are dominated by hand-crafted sequence descriptors coupled with shallow machine-learning classifiers or ensemble models. While these approaches have achieved high accuracy, they often struggle to capture long-range residue dependencies and contextual patterns that underlie aggregation propensity.This study proposes iAmyloid_PepCG, a sequence-based predictor that integrates multiple engineered features with a hybrid Convolutional Neural Network–Gated Recurrent Unit (CNN–GRU) architecture. Protein/peptide sequences were collected from publicly available benchmark datasets and encoded into a diverse feature space including amino-acid composition, composition–transition– distribution (CTD/CTDC/CTDD), dipeptide composition, pseudo amino-acid composition, physicochemical property (PCP) vectors, and contextual embeddings from transformer models (ESM, ProtBERT, ProtALBERT). A two-stage evaluation was performed: (i) 10-fold cross-validation on the training set and (ii) assessment on an independent hold-out test set.The proposed hybrid CNN–GRU model (iAmyloid_PepCG) achieved an independent-test accuracy of 95.45%, sensitivity of 100%, F1- score of 0.9333, Matthews correlation coefficient (MCC) of 0.9037, Cohen’s kappa of 0.8991, and area under the ROC curve (AUC) of 0.9714, outperforming classical ML baselines and several state-of-the- art amyloid predictors on the same benchmarks.Cross-validation accuracy reached 78.18% with an AUC of 0.8861, indicating stable generalisation.These findings demonstrate that combining local pattern extraction by CNN with long-range dependency modelling by GRU, applied to a rich multi- view feature representation, yields a powerful framework for amyloid protein prediction. Thesis Report 2025-12-27T00:00:00Z http://dspace.daffodilvarsity.edu.bd:8080/handle/123456789/17040 ParKNGo : A Parking Management System Ridoy, Md. Alamin The ParKNGo Parking Management System is a full-stack web-based application developed to address common urban parking problems by providing a smart and digital solution. The system connects customers who are searching for parking spaces with hosts who own available parking spots, making the entire parking process simple, fast, and reliable. Users can register and log in as Customers, Hosts, or Admins, and each role has clearly defined functionalities. Customers can search for nearby parking spots, apply filters such as location, price, and availability, and complete bookings through an easy-to-use interface. Hosts can add new parking spots, update availability, activate or deactivate listings, view customer bookings, and track their total earnings through a dashboard. Admins are responsible for managing users, approving parking listings, monitoring bookings, and ensuring smooth system operation. The application is designed to be secure, scalable, and responsive, with features such as real-time availability updates, efficient booking management, and a clean, user-friendly interface. Overall, the ParKNGo platform streamlines urban parking by improving convenience forcustomers, creating earning opportunities for hosts, and providing effective administrative control over the system. Project Report 2025-12-27T00:00:00Z