Canada Spends Logo

Explore our Budget 2026 analysis

Grant: Master's Award: Canada Graduate Scholarships

University of Toronto

$17,500.00

General Pre-Trained Model for Magnetic Resonance Image Processing

FY 2021-22

Summary
Machine learning holds a great promise in biomedical research and clinical diagnosis. However, the state-of-the-art machine learning models depend on a large quantity of high-quality data, multiple orders of magnitude more than that typically available in a biomedical setting. To overcome this limitation, I propose to use a pretrainable machine learning model that can leverage the vast amounts of publicly available, unlabeled, raw data in addition to the limited amount of available labeled data in the medical setting.MRI scans are widely used in clinical and research settings. Accurate and timely diagnosis and assessment of neurological disorders using MRI scans require reliable biomarkers such as brain morphometrics. Many of such biomarkers must be computed from segmentations of the MRI scan. Currently, the gold standard for MRI segmentation is manual segmentation which is very time consuming and labor intensive. This cost of data and limited quantity significantly limits the accuracy and scale of machine learning models.In the machine learning field of Natural Language Processing, researchers were able to utilize vast amounts of raw, unlabeled data through a transfer learning process known as pre-training. Most notably is the Generative Pre-trained Transformer 3, which achieved unrivaled performance in this field. Utilizing this pretraining process, I will develop machine learning models that can be pre-trained on the large, unlabeled MRI datasets. After pretraining on unlabeled data, the model can be modified to be fine tuned on expert labelled datasets. This type of pre-trained model can then be used on any task involving MRI scans. Such a model can allow the clinical benefits of machine learning to be applied to fields with relatively insufficient training data.
Program
Master's Award: Canada Graduate Scholarships
Location
Ontario, Canada
Recipient Type
Academia
Type
Grant
Source
CIHR Funding Decisions Database

As published

Id
466989
Name
Zhu Yi Cheng
Region
Ontario
Country
Canada
Orgtype
Institution Paid; Research Institution
Prcname2
Special Cases - Awards Programs
Projecttitle
General Pre-Trained Model for Magnetic Resonance Image Processing
Programtype2
Studentship Programs
Orgnameinp2
University of Toronto
Cihramount2
$17,500
Cihrequipment2
$0
Programname2
Master's Award: Canada Graduate Scholarships
Keyworddelim
Deep Learning; Image Segmentation; Machine Learning; Medical Image Processing; Pretrainable Model; Transfer Learning; Unsupervised Training
Abstract
Machine learning holds a great promise in biomedical research and clinical diagnosis. However, the state-of-the-art machine learning models depend on a large quantity of high-quality data, multiple orders of magnitude more than that typically available in a biomedical setting. To overcome this limitation, I propose to use a pretrainable machine learning model that can leverage the vast amounts of publicly available, unlabeled, raw data in addition to the limited amount of available labeled data in the medical setting.MRI scans are widely used in clinical and research settings. Accurate and timely diagnosis and assessment of neurological disorders using MRI scans require reliable biomarkers such as brain morphometrics. Many of such biomarkers must be computed from segmentations of the MRI scan. Currently, the gold standard for MRI segmentation is manual segmentation which is very time consuming and labor intensive. This cost of data and limited quantity significantly limits the accuracy and scale of machine learning models.In the machine learning field of Natural Language Processing, researchers were able to utilize vast amounts of raw, unlabeled data through a transfer learning process known as pre-training. Most notably is the Generative Pre-trained Transformer 3, which achieved unrivaled performance in this field. Utilizing this pretraining process, I will develop machine learning models that can be pre-trained on the large, unlabeled MRI datasets. After pretraining on unlabeled data, the model can be modified to be fine tuned on expert labelled datasets. This type of pre-trained model can then be used on any task involving MRI scans. Such a model can allow the clinical benefits of machine learning to be applied to fields with relatively insufficient training data.
Competitiondate
202112
Approvedterm2
1 yr 0 mth
Pinamesdelim
Zhu, Yi Cheng
Approvedterm
1 an 0 mois
Cihrcontribution2
~ $17,500~ $0~