Projects
Medical Image Segmentation Transformer with Convolutional Attention Mixing (CAM) Decoder (Ongoing)
Summary: One of the familiar and promising deep learning approaches used for medical image segmentation is transformers, as they can capture long-range dependencies among the pixels by utilizing self-attention. Despite being successful in medical image segmentation, transformers need help in capturing local contexts of pixels in multimodal dimensions. Hence, we are developing lightweight transformers with memory-enabled decoders and attention-mixing decoders for precise medical image segmentation.
Attentional Homogeneous-Padded Composite Model for Respiratory Anomalies Prediction
Summary: Early diagnosis, treatment, and regular monitoring can limit respiratory disease spread and adverse effects. However, more trained physicians are needed to ensure early diagnosis and treatment. This can be overcome by automating auscultation to identify anomalies like crackles, wheezes, and both. In this work, we propose a hybrid deep learning model combining ResNet34, a CNN model, a feature extractor, and Long Sort-term Memory (LSTM) as a predictor. We have also added an attention layer in the feature extractor to make the model learn the critical regions of the feature vector. Our model outperforms the recent state-of-the-art models and improves the performance of 2-class- and 4-class anomaly predictions. GitHub
A Machine Learning Approach for Real-Time Detection of Blood Glucose Concentration using IR
Summary: The study aims to develop an automated system using Conv-LSTM to identify diabetes from clinical and physical data. The system extracts significant features from the Pima Indian Diabetes Database using the Boruta algorithm and performs hyperparameter optimization using the Grid Search algorithm. The Conv-LSTM model achieved the highest accuracy of 97.26% using cross-validation technique, outperforming other models. The system can aid in early detection and diagnosis of diabetes.
An Automated Approach for Dairy Delivery Date Prediction using Artificial Intelligence
Summary: The Forest Deep Neural Network (FDNN) model integrates neural network architecture with a supervised random forest feature detector to predict cattle's daily and hourly calving time. The model is developed using physical activity data collected from 45 Holstein-Friesian cows 15 days before their calving time. Statistical analysis was carried out over the data and five classifiers, including FDNN, DNN, RF, DT, and SVM, were applied to the datasets. Hyperparameter optimization was conducted using Grid Search approach to filter out the optimal parameter configurations. The FDNN model outperformed the other four classifiers in terms of accuracy, sensitivity, specificity, and ROC score for predicting cattle's daily and hourly calving time. The study concludes that this model could help improve animal welfare and increase efficiency in the agricultural industry.