Machine learning for medical applications with limited data: Incorporating domain expertise and addressing domain-shift

• Author / Creator

  Vega Romero, Roberto Ivan

• Machine learning has the potential to help medical experts to deliver better healthcare. There are, however, important technical challenges that need to be solved before we can develop reliable models for clinical practice, including: (1) Limited number of labeled instances, (2) Uncertainty of the labels used during training, and (3) Differences between the distributions that generated the training and test data. This dissertation focuses on strategies for effectively applying machine learning under these circumstances.For learning models from a limited number of labeled instances, we propose incorporating domain expert knowledge during the training process. This domain expertise can be encoded in the form of probabilistic labels, which provide more information per instance than the commonly used categorical labels, or by using machine learning to extend the medical models currently used by human experts. We demonstrate the effectiveness of the probabilistic labels in three medical image classification tasks: for diagnosing hip dysplasia, fatty liver, and glaucoma. We observed gains up to 22% in terms of classification accuracy when compared with the use of categorical labels. We also show how to use machine learning to extend an SIR epidemiological model for predicting the evolution in the number of people infected with COVID-19, achieving state-of-the-art results in terms of mean absolute percentage error (MAPE) in data from the United States and Canada.For addressing the uncertainty around the labels, we use probabilistic graphical models. Instead of providing a point-estimate, probabilistic models predict an entire probability distribution, which accounts for the uncertainty in the data. Probabilistic models are a key component of the probabilistic labels mentioned above, and they also allow the incorporation of human decision making for tracking the number of new infections when using machine learning with the SIR model.Finally, a consequence of training machine learning models with a limited number of labeled instances is that the training set might not be an accurate reflection of the data used during inference -- in particular, the test set might not follow the same probability distribution that generated the training data. This means that a predictor learned from one dataset might do poorly when applied to a second dataset. This problem is known as batch effects or dataset shift, while approaches to correct for the discrepancies in these probability distributions fall under the umbrella term domain adaptation. Depending on the assumptions on what causes the discrepancy, these problems might be studied under specific names, such as covariate-shift, class-imbalance, etc.Here, we first propose an algorithm for domain-shift adaptation when the discrepancy between distributions is caused by linear transformations, and then empirically show that of style transfer technique can alleviate domain-shift caused by changes in texture. We provide empirical results for the task of segmentation of the hip in ultrasound images, with gains of up to 20% in terms of Dice score when applying style-transfer for unsupervised domain adaptation.Although all the applications in this dissertation are related to the medical domain, we expect that the techniques shown here are applicable when: (1) expert knowledge can be encoded as probabilities, (2) there exist a parametric model currently used by domain experts for analyzing a phenomenon, and/or (3) the discrepancy between the source and target domains is caused by affine transformations or differences in texture.

• Subjects / Keywords

  • Machine Learning
  • Medical applications
  • Domain adaptation

• Graduation date

  Fall 2022

• Type of Item

  Thesis

• Degree

  Doctor of Philosophy

• DOI

  https://doi.org/10.7939/r3-nq89-3145

• License

  This thesis is made available by the University of Alberta Library with permission of the copyright owner solely for non-commercial purposes. This thesis, or any portion thereof, may not otherwise be copied or reproduced without the written consent of the copyright owner, except to the extent permitted by Canadian copyright law.