Research Projects
(Project 01) iMRI: Integrated Magnetic Resonance Imaging
Magnetic resonance (MR) imaging is an essential tool for diagnostics and health management. Wait times are long and increasing for MR examinations in Canada, averaging approximately 9.3 weeks in 2019. Long wait times adversely impact personalized healthcare by delaying subsequent health services, leading to late diagnosis, poorer patient outcomes, and increased cost to both individuals and the health care system.
According to the 2019 “The Value of Radiology report,” the direct annual costs to the Canadian economy related to MR imaging diagnostics is $700M. The indirect annual costs incurred due to excessive MR wait times are estimated to be an additional $700M.
The time required to complete an MR exam often exceeds 45 minutes, but deep-learning-based image reconstruction methods have shown favourable results to reduce MR imaging examination times by reconstructing images from under-sampled acquisitions. These sophisticated reconstruction methods can increase patient throughput and reduce wait times. In practice, existing deep-learning models for MR reconstruction do not consider existing redundancies in the data, such as multi-sequence and multi-visit data (i.e., personalized data). These models often do not generalize well between different scanners. Automated deep learning (AutoDL) provides a framework to develop algorithms capable of fine-tuning image reconstruction models to specific users and use-cases without the intervention of a data scientist.
This project will develop AutoDL reconstruction methods that incorporate redundancies in the data, such as past subject-specific information, to make MR diagnostics more efficient. We propose developing a software application called MRIntelligence, which combines these innovations to reduce MR examination times by a factor of ten and to expedite scans' analysis, thus significantly improving personalized healthcare delivery and reducing MR-related expenses.
(Figure Caption) Pilot results of a brain MR imaging exam accelerated by a factor of 20. On the top left, a conventional, deep-learning-based MR image reconstruction approach and its magnified residual map. In the bottom left, the iMRI reconstruction On the right, the fully sampled reconstruction reference. Higher SSIM and pSNR, and lower NRMSE correspond to higher-fidelity reconstructions.
(Project 02) Detection of Regional Biomarkers of Brain Ageing using Magnetic Resonance Imaging
Global brain age prediction from MR images and its comparison with chronological age has proven to be a reliable biomarker for brain disorders, like Parkinson’s and Alzheimer’s disease, but also other conditions like Down’s Syndrome and HIV have been linked to the exacerbation of the brain ageing process.
The brain ageing models first proposed used handcrafted features, such as volumes of cortical structures and image texture, to develop a regression model to estimate brain age. With the success and rapid growth of deep learning, brain age prediction models shifted towards using convolutional neural networks (CNNs) for the brain age prediction task. The advantages of CNNs are that they can learn the features directly from the data (i.e., no need to handcraft features), and these deep learning models often produce more accurate predictions than traditional methods. Several deep learning models report an average brain age prediction error < 2 years.
Age-related brain changes are characterized by region-specific and nonlinear patterns of processes, such as cell growth and synaptic pruning, and widespread brain atrophy that happens during brain ageing. Although accurate deep-learning-based global brain age prediction models can indicate signs of accelerated brain ageing and brain disorders, they lack the spatial specificity to highlight which regions of the brain are the most affected.
The concept of regional brain age prediction based on MR image features is new, and it overcomes the limitations of having a single global index. The regional brain age prediction methodology proposed and investigated in this study is novel and will create an advanced paradigm for spatially resolving brain ageing mechanisms based on imaging features. This new paradigm will expand the field to new exciting directions that will allow us to understand better the normal brain ageing mechanisms and how different disorders affect the human brain.
I anticipate that regional brain age predictions will be a superior biomarker compared to global brain age prediction across many disorders. The ability to detect regions of the brain that show signs of accelerated ageing will allow us to better understand the mechanism of disease, and to develop and verify the efficacy of new therapies and interventions that can counter the effects of accelerated ageing.
(Project 03) Machine Learning FOR HETEROGENEOUS BRAIN MRI: BRIDGING THE GAP TO GENERALIZAble MODELS
Magnetic resonance (MR) imaging has become a key technology for brain imaging, resulting in massive databases, rapidly increasing the need for big data analytics, robust pooling, and harmonization, especially for data acquired across diverse cohorts. A barrier to the success of these techniques is the inherent variation between image acquisition protocols and different equipment, resulting in a lack of reproducible results. It has been shown that even when care is taken to standardize acquisitions, changes in hardware, software, or protocol design can lead to differences in quantitative results and loss of consistency. As a result, the quantitative utility of MR in multisite or long-term studies is dramatically impacted.
Machine learning (ML) has been extensively investigated for MR imaging analysis with multiple goals, such as quantitative analysis of structures or abnormalities and progress evaluation over time. Yet only a limited number of applications are now in use outside the research environment. A key reason for that is the poor generalizability of the models to data from different sources or acquisition domains. Developing new methods to handle this diverse MR imaging data is crucial for achieving accurate models and broadening their usage. Our long-term research program goal is to tackle the current limitations of the broader use of ML for medical imaging, focusing on the challenges of conducting large and multisite studies, using data harmonization and domain adaptation strategies of ML models that allow generalization from one dataset to another, avoiding domain-specific decision-making.
I anticipate a significant improvement in the generalization capability of ML tools developed for brain MR imaging applications. My findings will significantly impact the research area by allowing the usage of such models in larger, heterogeneous datasets and best practices when translating learning from one application to another using the proposed data harmonization and domain adaptation strategies. While my short-term goal is to work with MR imaging, the proposed strategies would be substantial for translation to other applications for other medical images and other computer vision applications.
(Project 04) What goes where? A Garbage Classification system based on Images and Natural Language
Every year tons of garbage are inappropriately disposed of in the wrong trash bin causing a huge environmental footprint. Sorting the garbage properly across recyclable, compost and landfill trash bins is of uttermost importance to reduce this footprint. This project focuses on using image recognition and natural language processing for real-time classification of different types of garbage, so they can be correctly sorted for recycling, composting or landfill. The goal is to build a system in which the user can show an object to a camera or take a picture and optionally input some keywords that describe the object into a terminal or through voice recognition. Then, the system will tell the user in which bin the garbage should be disposed of, or if the user should take the garbage to a waste drop-off location. The system can also tell the user that the garbage can be taken to a recycling center for a refund. This is very common in the city of Calgary with the so-called “Bottle Depots”, where the user can recycle empty plastic/glass bottles, beer cans and drink packings, among other similar items. This system will help increase the percentage of garbage that’s recycled since throwing away recyclable garbage into a bin designed for non-recyclable items will cause it to end up in a landfill or a compost centre.
(Project 05) fairness in machine learning
One of the most crucial aspects to consider while working on Machine Learning models is how bias and fairness in various stages can affect outcomes for different user groups. This especially becomes more important in the context of AI in healthcare, where patients' privacy should be treated with utmost care. In this project, we intend to develop interpretable machine learning models while identifying and reducing sources of bias in the data and using features through methods such as aggregation and proper evaluation among different demographic groups.
We prioritize equity, diversity, and inclusion to develop trustworthy ML algorithms for medical imaging applications. We expect to present reliable and non-biased results for equity deserving groups, male and female participants of any race, all age ranges in the adult lifespan, and people with varying educational backgrounds. Our work can be translated to datasets other than the one we used during development, such as multi-site heterogeneous data, simulating larger clinical trials. We intend to make code available, making our findings reproducible and providing straightforward benchmarking.
(Project 06) Machine Learning for pediatric brain Mr data
Details coming soon...