Grant: Master's Award: Canada Graduate Scholarships
University of Alberta
$17,500.00
Data-Efficient Medical Image Classification using Learned Uncertainty
EF 2022-23
Résumé
In recent years, medicine has witnessed the promising potential of machine learning (ML). ML, particularly deep ML, is well-suited to medical diagnostics given the high dimensionality and growing availability of patient data. Despite increased data curation for such tasks, producing informative labels for medical data remains expensive and resource-intensive. We seek to improve diagnostic classification performance when given only binary labels, specifically investigating if learning uncertainty can improve model generalization and sample efficiency. Previous research has shown that using medical expert-curated features to generate probabilistic labels can improve model accuracy trained on relatively small datasets. However, this method relies on domain experts and does not address the challenges of obtaining labels for medical data. Our work instead explores the philosophy of separating knowledge acquisition (learning) from utilization (predicting), such that learning informative features and labels is autonomous and relies only on the given input data and its labels. We perform an empirical evaluation on the accuracy improvement and data-efficiency of different representations of input and label uncertainty, as well as establish a theoretical foundation for encoding this uncertainty within ML architectures and training regimes. We also suggest alternative implementations, such as transfer learning, that can achieve the same goals under a different formulation. Our principal motivation is to improve diagnostic classifiers for the clinical setting, focussing on reliability and scalability to support widespread adoption in healthcare. Our implementation reduces the minimum required data to train strong classifiers while reducing the labeling burden on medical experts.
Programme
Master's Award: Canada Graduate Scholarships
Bénéficiaire
Emplacement
Alberta, Canada
Type de bénéficiaire
Academia
Type
Grant
Source
CIHR Funding Decisions Database
Tel que publié
Id
486617
Name
Du Tian Xiang
Region
Alberta
Country
Canada
Orgtype
Institution Paid; Research Institution
Prcname2
Special Cases - Awards Programs
Projecttitle
Data-Efficient Medical Image Classification using Learned Uncertainty
Programtype2
Studentship Programs
Orgnameinp2
University of Alberta
Cihramount2
$17,500
Cihrequipment2
$0
Programname2
Master's Award: Canada Graduate Scholarships
Keyworddelim
Convolutional Neural Networks; Deep Learning; Diagnostic Classification; Image Classification; Irritable Bowel Disease; Machine Learning; Model Distillation; Probabilistic Labels; Sample Efficient; Transfer Learning
Abstract
In recent years, medicine has witnessed the promising potential of machine learning (ML). ML, particularly deep ML, is well-suited to medical diagnostics given the high dimensionality and growing availability of patient data. Despite increased data curation for such tasks, producing informative labels for medical data remains expensive and resource-intensive. We seek to improve diagnostic classification performance when given only binary labels, specifically investigating if learning uncertainty can improve model generalization and sample efficiency. Previous research has shown that using medical expert-curated features to generate probabilistic labels can improve model accuracy trained on relatively small datasets. However, this method relies on domain experts and does not address the challenges of obtaining labels for medical data. Our work instead explores the philosophy of separating knowledge acquisition (learning) from utilization (predicting), such that learning informative features and labels is autonomous and relies only on the given input data and its labels. We perform an empirical evaluation on the accuracy improvement and data-efficiency of different representations of input and label uncertainty, as well as establish a theoretical foundation for encoding this uncertainty within ML architectures and training regimes. We also suggest alternative implementations, such as transfer learning, that can achieve the same goals under a different formulation. Our principal motivation is to improve diagnostic classifiers for the clinical setting, focussing on reliability and scalability to support widespread adoption in healthcare. Our implementation reduces the minimum required data to train strong classifiers while reducing the labeling burden on medical experts.
Competitiondate
202212
Approvedterm2
1 yr 0 mth
Pinamesdelim
Du, Tian Xiang
Approvedterm
1 an 0 mois
Cihrcontribution2
~ $17,500~ $0~