According to the latest CNS Tumor guidelines 2021 of World Health Organization (WHO), brain tumors need to be classified based on its genetic mutation information along with radiological and histopathological information. But genetic testing is costly and not accessible to many parts of the world. Hence we are trying to mitigate the problem by predicting the genetic mutation status from the histopathological images of the tumor. This will help reduce the treatment cost. From a technical point of view, this is challenging as we are trying to predict multiple genetic mutations from a single model, i.e., this is a multi label classification problem which is more challenging than mult class classification problem.

Parkinson’s disease (PD) is a neurodegenerative condition that affects millions of individuals worldwide. One of its main symptoms is motor impairment, which includes tremors, rigidity, and bradykinesia. Accurate and early estimation of the PD stage is crucial for effective diagnosis and disease management. In this work, we propose a hybrid ML-DL approach to estimate different (as many as four) stages of PD. We utilize multi-modal data and apply late fusion. Our strategy takes advantage of the complementary information offered by multiple modalities, including neuroimaging data (specifically T1-weighted MRI) and clinical assessment data. The clinical assessment is based on MDS-UPDRS features and the H&Y scale. We apply various ML models on clinical data and select the best-performing classifier. We investigate the impact of different loss functions on the performance of our DL model on the MRI data. Our findings demonstrate that the focal loss function outperforms other loss functions in terms of accuracy. Furthermore, we demonstrate that there are significant differences in MRI scans across various stages of PD using statistical analysis. We have evaluated our proposed method on the publicly available PPMI dataset. The experimental results show that our approach surpasses several state-of-the-art methods for PD stage estimation, achieving an impressive accuracy of 98.08%.

Parkinson’s disease is the second most common degenerative disease caused by the loss of dopamine-producing neurons. The substantia nigra region is deprived of its neuronal functions, causing striatal dopamine deficiency. Clinical diagnosis reveals a range of motor to non-motor symptoms in these patients. In subjects with Parkinson's disease, magnetic resonance imaging (MR) imaging is able to capture the structural changes in the brain due to dopamine deficiency. In this work, the CNN model is used to try to analyze MR brain images to tell the difference between healthy control subjects and PD subjects. 

Recently, deep learning (DL)-based image compression (CS) algorithms have been developed that provide high reconstruction quality with minimal computing cost. However, present DL-based image CS approaches ignore interframe signals, resulting in inefficiency when applied to video CS. A convolutional neural network (CNN) is used to analyze intraframe and interframe correlations in this work. First breaks the video sequence into numerous groups of pictures (GOPs), with the first frame being a keyframe sampled more than the others. In a GOP, a convolution layer proposes block-based framewise sampling, which optimizes the sampling matrix. The framewise initial reconstruction utilizing a linear CNN, which efficiently uses intraframe correlation, is first considered. Then comes deep reconstruction with multilayer feature compensation, which compensates non-keyframes with keyframes. The current non-keyframe can be recovered using both keyframe and previously recovered non-keyframes. By this approach, the recovered non-keyframes may have higher PSNR than when the non-keyframes recovered from only keyframes.  Also, we will consider SSIM as a loss function. We will add an encryption model after compression.

The success of human action recognition in recent years has been mostly driven by a large amount of accessible labelled data. Supervised-based action recognition approaches assume that training data (a source domain) and testing data (a target domain) are sampled independently and identically distributed from the same distribution space. In practice, the training and testing data are usually collected from the related domain but under different distributions, which is called a "domain shift." The performance of models degenerates significantly when applied to a new test domain with different distributions due to domain shift. This greatly limits the application of current action recognition models.