Undergraduate Thesis & First Paper

• This was my Undergraduate Thesis which was supervised by Dr. Mir Md. Jahangir Kabir Sir. 

• Abstract: Alzheimer's disease (AD) is a widespread neurological condition affecting millions globally. It gradually advances, leading to memory loss, cognitive deterioration, and a substantial decline in overall quality of life for those affected. AD patients experience memory decline, eroding cherished memories and straining relationships, while daily tasks become challenging. Numerous investigations have been conducted in this field, as the timely identification of Alzheimer's disease at its initial stage is of the utmost importance. A major limitation in this field is the predominant emphasis on using single fine-tuned CNN architecture or comparing pre-trained and custom CNN models for Alzheimer's detection, often on small datasets, which neglects a more comprehensive approach. Using smaller datasets can negatively impact deep learning modeling accuracy due to overfitting, limited representation, and poor generalization. This study addresses the current research problems and proposes an ensemble approach that combines predictions from various pre-trained models, including DenseNet-121, EfficientNet B7, ResNet-50, VGG-19, and a Custom CNN. The model averaging ensemble method was applied, a subset of the Stacking Ensemble, to two ADNI datasets, with Dataset-I being the larger. The goal was to assess the efficacy of this ensemble approach for accurate multiclass classification on ADNI datasets, where it successfully identified all classes despite differing sample volumes. A vast experiment was conducted on two distinct and widely recognized real-world datasets, resulting in accuracies of 99.96% and 98.90% respectively. Finally, the outcome of the research compared with recent research findings demonstrates the potential of our approach in advancing Alzheimer's disease detection by outperforming other benchmark approaches by a significant margin. 

Published Papers

2. Exploring Deep Convolutional Neural Networks: A Grad-CAM Enhanced Comparative Study for Automated COVID-19 Diagnosis from Chest X-ray Images

Abstract: The emergence of COVID-19 in late 2019 sparked a global health crisis. Identifying infections quickly is key to providing proper care and limiting transmission. The Reverse Transcription-Polymerase Chain Reaction (RT-PCR) test, a widely recognized primary diagnostic method for detecting COVID-19, faces challenges associated with cost and compatibility in resource-constrained environments. This issue is resolved by chest X-ray imaging since it is non-invasive and inexpensive. In this domain, research often utilizes pre-trained models on relatively small datasets of chest X-ray images, limiting the potential for optimal generalization and accuracy achievable with larger training sets. Also, these research works limit the exploration of crucial information on salient regions within CXR images, constraining the depth of analysis and interpretation. This study addresses current research challenges through a detailed comparative analysis of convolutional neural network (CNN) architectures, including pre-trained models like Inception V3, Xception, DenseNet-121, VGG-19, and a Custom CNN developed from scratch. The evaluation focuses on their effectiveness in detecting COVID-19 using a relatively vast and comprehensive dataset. The evaluation focuses on assessing the efficacy of these models in detecting COVID-19 across binary-class and multi-class classifications within a large dataset, offering valuable insights into their overall effectiveness. Grad-CAM is employed in this study, which visualizes picture salient areas affecting deep neural network predictions and improves model interpretability. The heatmaps reveal the model’s decision-making processes, making their behavior more understandable to medical practitioners. With resulting accuracies of 99.13% and 97.66% in binary class and multi-class classification, Inception V3 demonstrated the most optimal performance. This proposed approach surpasses existing benchmark methods, underscoring its superiority.

3. Detection of Fake News with RoBERTa Based Embedding and Modified Deep Neural Network Architecture

Abstract: The spread of fake news has emerged as a critical challenge in the era of information and digital connectivity. The consequences of misinformation can be profound, affecting public opinion, policy decisions, and even public health. Therefore, detecting and reducing the spread of fake news is an essential issue that requires reliable and precise solutions. Historically, the field of fake news detection has witnessed notable advancements, but several shortcomings have persisted. Many earlier approaches struggled to attain the requisite levels of accuracy. Another formidable obstacle that these earlier approaches encountered was the dynamic and ever-evolving nature of the tactics employed by those who spread fake news. They employ sensational language, which uses phrases with high emotional content to attract readers and elicit a strong emotional response. Recognizing these deficiencies, we present an innovative solution that leverages the state-of-the-art RoBERTa model and a meticulously modified deep neural network architecture. Our approach stands out by not only recognizing the urgency of the fake news detection problem but also by proposing architectural enhancements. It specifically targets the inadequacies of prior methods. We introduce attention mechanisms designed to identify subtle cues indicative of misinformation. The feature extraction techniques capture the nuanced patterns that fake news articles often follow. These architectural refinements make our model extremely effective and achieve an accuracy of 99.76%. The comprehensive evaluations demonstrate that our RoBERTa-based model consistently outperforms previous state-of-the-art fake news detection models, emphasizing the crucial role of advanced language models in combating misinformation.

4. Adapting Contextual Embedding to Identify Sentiment of E-commerce Consumer Reviews with Addressing Class Imbalance Issues

Abstract: Understanding consumer attitudes toward specific products is crucial for boosting sales in the e-commerce industry. To effectively target customers with popular products based on reviews, the classification of consumer feedback becomes imperative. However, classifying product reviews can be challenging, particularly when dealing with imbalanced data labels, which often result in suboptimal classification performance. This study builds upon previous efforts that utilized the Amazon Fine Food Reviews dataset for classification tasks. While these prior attempts showed promise, they were hindered by either poor embeddings or the prevalent class imbalance issue. In response, this research tries to solve these problems by using word embeddings with RoBERTa, a pre-trained transformer-based language model, to classify reviews. Additionally, the XGBoost classifier was implemented, along with embeddings from the language model. Losses were first calculated with equal weights for all class labels, and a re-weighted loss was subsequently adopted to balance the impact of each class on the loss function during training. The incorporation of RoBERTa and XGBoost, along with the class label re-weighting, contributed to improved capturing of intricate word relationships within reviews. As a result, this approach achieves significantly improved accuracy in both binary and multiclass classifications compared to earlier endeavors. Notably, it attained an impressive accuracy of 83.84% in multiclass classification and 93.29% in binary classification tasks, marking a substantial advancement in the field of consumer review analysis.

• Paper Under Review at F1000 Research (Q1)

• Contribution: Co-author

• Preprint:  Link 

6. Machine Learning-Based Prediction of Type II Diabetes: Leveraging High-Quality Datasets and Key Variables for Early Detection and Clinical Management. 

Abstract: Diabetes, a chronic and widespread disease marked by elevated blood sugar levels, poses severe health risks such as kidney failure, heart attacks, blindness, and limb amputation. Its increasing prevalence globally underscores the urgent need for early detection and intervention. This study investigates the performance of various machine learning algorithms in predicting diabetes, emphasizing the significance of timely prognostication to inform clinical decisions and aid prevention efforts. With the escalating impact of unhealthy lifestyles and dietary choices, diabetes has become a leading cause of mortality. Leveraging high-quality datasets like Vanderbilt and PIMA India, recent research identifies key variables such as glucose, pregnancy, BMI, diabetic pedigree function, and age for accurate Type II diabetes prediction. The study employs Logistic Regression, demonstrating the potential of machine learning models in forecasting the likelihood of diabetes onset, thereby contributing to proactive healthcare management and prevention strategies. 

• Contribution : Co-author & Corresponding author

• Paper under revision at Healthcare Technology Letters (Q3) 

7. Inductive and Transfer Learning based hybrid model Techniques for Accurate and Automated   Diagnosis of Neurological Diseases

• Paper under revision at Brain & Behaviour (Q2).

• Contribution: Co-author & Corresponding author.

8.Human Stress Detection Technologies: An In-depth Comparison of Machine Learning Algorithms and Applications

Abstract: Workload pressure, examination stress, family responsibilities, and a variety of other factors all contribute to an increase in stress in the body. Stress weakens the human mind and body by accelerating several health disorders. Therefore, the number of approaches for early projection of stress plays a vital role in the healthcare sector. Stress prediction techniques can be broadly categorized as questionnaire-based, where a psychiatrist provides a feedback form to the user to identify the status of stress. In sensor-based stress measurement methods, stress will be measured by some symbolic constraints like heart rate, skin conductance, pupil diameter, and a list of questions. For collecting the crucial data for stress detection, Skin Temperature (ST), Electrocardiogram (ECG), and Electrodermal Activity (EDA) are keenly monitored during the experiment. In this manuscript, a comparative study was performed using the results of various techniques to predict human stress. This article represents various techniques for stress prediction and various opportunities to improve their performance 

• Paper under minor revision at The Journal of Engineering(Q2).

• Contribution: Co-author.

9. An Insight into Content-Based Image Retrieval Techniques, Datasets, and Evaluation Metrics

Abstract: The goal of a Content-Based Image Retrieval (CBIR) framework is to enable users to efficiently retrieve images from a large database based on the visual content of the images, rather than relying on metadata or annotations. Content-based image retrieval (CBIR) systems are becoming more and more popular finding their applications in a wide variety of fields like heath care, e-commerce, law enforcement, and searching digital libraries. Computing machines with CUDA architecture have powered deep learning-based techniques for efficient CBIR and, as such, CBIR systems have become fast with more accurate query results. This work is intended to provide an introduction to CBIR systems and different feature and learning-based techniques to perform CBIR. An overview of different datasets, evaluation metrics, and pros and cons of different CBIR techniques is presented. The paper concludes by discussing current research challenges and future opportunities to improve and apply CBIR to various fields.

• Paper under review at Journal of Optics.

• Contribution: Co-author.