Projects
Projects
Embedded Files
Research projects
Research projects
I am currently working on this project where my teammates and I leverage the Diffusion models to get the brain tumor segmentation from MRI images. We introduced a new method that eliminates the need for hyper-parameter tuning of the Diffusion models. Our model can automatically identify the best noise level and timestamps for image inference for the diffusion model. The pre-print is available on the archive titled AnoFPDM: Anomaly Segmentation with Forward Process of Diffusion Models for Brain MRI.
Using Large-scale Contrastive Language-Image Pre-training to Maximize Brain MRI-Based Headache Classification
In this project, I am using the BioMedCLIP model which is pre-trained on the PMC-15M dataset for classifying different headaches from MRI images. This task is challenging for the clinicians as there is very little difference between the healthy and headache patients' MRI scans. Hence I am using the Deep learning methods to find the latent space differences using MRI images and text data. We fine-tuned the BioMedCLIP model on our private dataset from Mayo Clinic and achieved a remarkable performance increase in terms of accuracy, sensitivity, and AUC scores. The results from the model show that using a multimodal model helps to get a better representation for classification.
Early human illness diagnosis has been demonstrated to be more accurate when deep learning methods are used. In the case of diagnosing a brain tumor, when even a little misdiagnosis might have serious consequences, accuracy is especially important. Disclosure of brain tumors in medical images is still a difficult task. Brain MRIs are notoriously imprecise in revealing the presence or absence of tumors. Using MRI scans of the brain, a Convolutional Neural Network (CNN) was trained to identify the presence of a tumor in this research. Results from the CNN model showed an accuracy of 99.17%. The CNN model's characteristics were also retrieved.
Worked on this project related to Virtual Network Functions (VNF). Our goal was to increase the energy efficiency of Virtual Network Functions and network operators to turn off the virtual instances of the network services according to the service demand. When the demand is low, these can be turned off to save significant energy. We proposed deep learning-based VNF fore-casting methods to forecast vCPU requirements for different service functions depending on the traffic. Our results show promising accuracy and fewer errors. This paper is published in the Journal of Networks and Systems Management.
Comparison of Shilling Attacks on Recommender Systems
I along with my undergraduate thesis partner worked on this research work that included a comparative study of shilling attacks on the recommender system. It was an extended work of Dr. Felice Antonio Merra, who studied the effects of different shilling attacks such as the Random attack and Bandwagon attack. We compared his results with a new form of attack which is called love hate attack. For measuring the performance of the attack, we chose three recommendation models: ItemKnn, UserKnn, and Neural collaborative filtering. In this project, we used the publicly available dataset from Library Thing to build fake shilling profiles and injected those profiles into the recommender system. Finally, we measured the hit rate of the initially targeted items and compared the results with those from the paper of Dr. Felice Antonio Merra. We found that the hit rate of the LoveHate attack is comparatively higher than the other forms of attack.
Page updated
Google Sites
Report abuse