There are many active research projects accessing and applying shared ADNI data. Use the search above to find specific research focuses on the active ADNI investigations. This information is requested annually as a requirement for data access.
| Principal Investigator | |
| Principal Investigator's Name: | mohammad nili |
| Institution: | Tehran |
| Department: | AI |
| Country: | |
| Proposed Analysis: | Our central membership includes 4 participants (3 AI students and 1 Cognitive Science student) under supervision of a neuroscientist professor. Considering the following two papers which were solid in the results (Jodko-Władzińska et al., 2020; Llinás et al., 2020) that used magnetography methodology, we were inspired to used deep learning and magnetography to detect anomalies in standard MRI datasets. There are 3 major themes of brain disorders in our work (Parkinson, Alzheimer, and Schizophrenia) for the topographic organization of the brain using deep-learning, neural networks. By attending to the appendix of Paul Lauterbur and the Invention of MRI book (1971) we know “the basic idea of magnetography is imposing a sequence of magnetic field gradients on an object to convert the frequency of magnetic resonance signals (MRI) to measure of its location within the object.” There is also Relaxation magnetography which defined as “sharp distinction may be made between regions of different relaxation time by an appropriate variation on any of the magnetographic experiment.” Such imaging can be obtained without additional methodological steps such as the use of contrast media. Intuitively, magnetography is brain activity based on MEG recordings addressed as functional tomography also called “functional tomogram”. the flow of transmembrane current that is the basis for the functional tomograms, represented magnetic fields In both nerve and muscle tissue. Learning normal anatomy of a healthy brain (learning to compress and recover) is the main idea of Unsupervised Anomaly Detection (UAD) based on Deep-learning in MRI data. Early works such as Auto- encoders and Latent Variable models are classic. Generative Adversarial Network, Reconstruction- based, Monte Carlo, Gradient-based, Restoration-based methods have better results (Baur, Denner, et al., 2021). we can also develop classic architectures for deep convolutional neural networks (DCNN) like VGG16 (Matsuo et al., 2020). In contrast, assuming Gaussian distribution for segmentation fMRI signals which are not necessarily suitable. Deep learning frameworks such as RNN and LSTM and GRU in a data-driven way to detect temporally dynamic functional states are suitable NNs for sequence modeling (Li & Fan, 2018). Also, we could use the Bayesian approaches and generate ensemble architectures, Deep Siamese networks for representation learning (Alaverdyan et al., 2020). Both unsupervised and supervised U-Net automatically detect deviations indicative of lesions or edema on abnormal MR images (Baur, Wiestler, et al., 2021). By attending to (Muñoz-Ramírez et al., 2022) we could apply recent deep learning techniques to detect anomalies in brain diffusion tensor imaging (DTI) in neurodegenerative diseases such as Parkinson's disease (PD). by using the Gaussian mixture model (GMM) we could estimate the density with the distributions of the extracted features (Kim et al., 2021). to accelerate the speed of MRI scans some real-time approaches based on reconstructive techniques were introduced. These methods reduce the number of slices and required repetition/echo times per slices such as few- shot learning, transfer learning, on-device learning introduced for motion correction, anomaly detection, and anatomical segmentation (Getty et al., 2021). We can conclude from previous works that new architectures and techniques based on Game Theory such as GAN networks and energy base loss function and also using meta-learners can be helpful in future works on MRI images for segmentation and anomaly detection. |
| Additional Investigators |
