Publications

Performance Analysis of Machine Learning Algorithms in Chronic Kidney Disease Prediction

The paper suggests a computer-aided design to predict chronic kidney disease (CKD) using eight machine-learning algorithms. It evaluates their performance on a UCI machine learning repository dataset and employs random imputation and mean/mode sampling to handle missing data. Feature encoding is used to convert categorical data into numerical values. Results show Random Forest and Logistic Regression achieve the highest accuracy at 99%, followed by AdaBoost, XGBoost, Naive Bayes, Decision Tree, and SVM. KNN performs the least well with 73% accuracy.

Ongoing Research

Segmentation-Enhanced Deep Learning Utilizing SAM for Freshwater Fish Species Classification in Resource-Constrained Regions

Developing an innovative approach for freshwater fish species classification, particularly in regions like Bangladesh, we're integrating image segmentation techniques into deep learning models to enhance accuracy and resilience. Utilizing the "segment-anything-model," we're generating precise mask images for training a U-Net model specifically tailored for fish segmentation. This strategy focuses on localized features to improve classification accuracy and mitigate deceptive practices. With our dataset of 2300 images, we're conducting an analysis to evaluate the effectiveness of this segmentation-enhanced approach, with promising preliminary results indicating potential accuracy improvements.

Deep Learning-Based Mobile Application for Skin Cancer Detection: Enhancing Accuracy and Efficiency Using Diverse Datasets

Developing a mobile application capable of accurately identifying skin abnormalities can aid in timely medical intervention. Leveraging deep learning techniques, such as classification algorithms, can enable precise skin cancer diagnosis. In this study, we evaluate skin lesion classification using the HAM10000 dataset, comprising nearly 10,000 dermatoscopic images. We assess three neural networks EfficientNetV2, MobileNetV2, and ResNetV2, and explore segmentation using the Unet architecture to enhance model precision. Preliminary results indicate promising outcomes for skin cancer classification. By selecting the top-performing model, MobileNet, we plan to develop a mobile application for diagnosis, incorporating diverse datasets to improve accuracy and usability.