Embedded Files
Presentation and Projects
Poster Replica
Idea Presentation:
A Research on COVID-19 for determining the potential risk factors of being infected with COVID-19 in students of the Department of Statistics at the University of Rajshahi. (Academic Project)
Location: Faculty of Science; University of Rajshahi.
Poster Replica
Poster Presentation:
Comparative study of different machine learning algorithms in big data analytics towards the Fourth Industrial Revolution.
Location: Diamond Jubilee Celebration and Third Reunion Poster Exhibition (2023); Department of Statistics, University of Rajshahi.
(Get the first prize for the excellent Poster)
Publications
(Presented) Gourab Sarker, Md. Mesbahul Alam. (2023); Comparative Study of Convolutional Neural Network Architectures on COVID-19 Classification;
International Conference on The Role of Science and Technology towards 4IR.
Abstract:
In the years 2020 - 2021, coronavirus disease (COVID-19) was a name of fear and we still do not have a clear picture about it. As it is unknown and deadly for humans, it needs to get diagnosed hurriedly. Recently, deep learning (DL), and precisely convolutional neural networks (CNNs), have achieved good results for the classification of COVID-19 and non-COVID-19. This research is a comparison study for the six well-known CNN architectures: VGG16, ResNet50, InceptionV3, InceptionResNetV2, DenseNet201, and Xception. For this purpose, the models were evaluated through a multiclass statistical analysis based on accuracy and AUC. The dataset used for the experiments contained two different classes of COVID-19 and non-COVID-19 from Kaggle. The results reveal that the VGG16 architecture is found to be the best having test accuracy of 83.89%. Therefore, it is reasonable to conclude that the VGG16 model is a beneficial tool for doctors and medical professionals in assisting the identification and prevention of the diseases stated.
Keywords: cnns, covid-19, deep learning, image_net, accuracy.
(Presented) Gourab Sarker, Samiul Islam, Md. Mesbahul Alam. (2024); Improving Skin Lesion Classification with Deep Learning Techniques;
International Conference on The Role of Statistics and Data Science in 4IR.
Abstract:
This study explores the application of pre-trained deep-learning models for classifying skin lesions into four distinct categories: Chickenpox, Measles, Monkeypox, and Normal skin conditions. Leveraging the InceptionV3 architecture, we achieved an impressive accuracy of 91.67%, surpassing the performance of other architectures, including DenseNet201, Xception, and InceptionResNetV2. Our methodology emphasizes the importance of utilizing a diverse dataset combined with innovative data augmentation techniques to address generalization concerns that often hinder model performance in real-world clinical scenarios. The enhanced robustness of our models not only improves their reliability but also their applicability in dermatological diagnostics. Furthermore, we employed layer-wise accuracy heatmaps to provide valuable insights into InceptionV3's decision-making processes, illustrating its ability to discern intricate features within skin lesion images. This dual approach significantly contributes to the existing literature on skin lesion classification and paves the way for future research aimed at enhancing the diagnostic accuracy and interpretability of deep learning models in healthcare settings. As artificial intelligence continues to transform medical diagnostics, our findings underscore the potential for effective, efficient, and accessible skin lesion diagnosis using advanced machine learning techniques.
Keywords: skin lesion classification, deep learning, cnn, data augmentation, medical imaging.
M.Sc. Thesis
Breast Cancer Histopathological Image Classification with Convolutional Neural Networks
Abstract:
Breast cancer is a significant global health issue that requires accurate and efficient diagnostic tools. This research paper focuses on the crucial requirement for robust histopathological image classification models and explores the capabilities of convolutional neural networks (CNNs) in addressing this need. The study conducted experimental analyses on the BreakHis dataset, examining magnifications of 40x, 100x, 200x, and 400x. Six predefined CNN models (VGG16, Xception, ResNet50, EfficientNetB7, MobileNet, and DenseNet201) were compared with our proposed model. The results of the study reveal interesting patterns, particularly the consistent values of accuracy and precision across epochs, accompanied by nearly perfect recall. The performance of the models is presented in detail, demonstrating accuracy ranging from 67.70% to 96.88%, precision between 0.677 and 0.987, and mean absolute error (MAE) fluctuating from 0.193 to 0.322 across different magnifications. VGG16, Xception, ResNet50, EfficientNetB7, and DenseNet201 exhibit stability in terms of accuracy and precision. MobileNet, on the other hand, shows increased accuracy at 400x magnification and consistently performs well across all magnifications among the predefined architectures. Comparative analyses highlight the superior performance of the proposed model, achieving an accuracy of 96.88% at 400x magnification. In-depth discussions delve into the peculiar stability observed in the metrics, emphasizing the need for further investigation into the learning dynamics and potential implications for model generalization. The strengths, limitations, and clinical implications of the study are thoroughly examined, revealing the potential of the proposed model to achieve exceptional diagnostic accuracy. This thesis contributes to the ongoing discourse on utilizing advanced deep-learning techniques for breast cancer histopathological image classification. The presented findings, methodologies, and insights lay a solid foundation for future research endeavors aimed at enhancing diagnostic exactness in breast cancer pathology.
Keywords: cnns, breast_cancer, deep_learning, histopathology, imagenet.
Page updated
Google Sites
Report abuse