Computing and Information Systems - Theses
Permanent URI for this collection
Search Results
1 - 2 of 2
-
ItemTowards Robust Medical Machine Learning( 2022)Machine learning systems have been developed to address a number of problems in numerous domains, among which medical solutions have been facilitated by machine learning approaches for decades. These approaches play important roles in automated disease diagnosis, medical image processing, and auxiliary surgical operation, etc. Despite the highly efficient diagnosis benefited from machine learning approaches, these methods may not be robust to common challenges in practical scenarios, such as special while crucial characteristics of medical data, annotation variations from multiple experts, noisy annotations, and multi-source datasets. Such problems impede machine learning methods from being applied accurately and safely to medical tasks. In the thesis, we introduce special while important medical problems that were not brought into the spotlight before. We then provide corresponding robust machine learning solutions for each problem when existing machine learning methods degrade significantly in these tasks. Specifically, the first problem is the similarity analysis for time series with large discontinuities, which is common in surgical time series. We thus propose a robust distance measurement for time series with large discontinuities when they disable the accurate measurement of local characteristics using existing algorithms. Second, surgical policies provided by different surgeons for the same patient/surgery may not be exactly the same. We then propose the reward-penalty Dice loss (RPDL) to learn non-unique surgical segmentation regions for deep vision networks. RPDL is robust to varying annotations for the same input, which enables the comprehensive learning of models from multiple experts. Third, medical datasets might be composed of limited examples and noisy annotations, making it challenging to train deep learning models. To address this challenge, we propose alpha-IoU, a family of power Intersection over Union (IoU) losses for bounding box (bbox) regression. We show that alpha-IoU losses are more robust to small datasets and noisy bboxes in lesion detection. Fourth, large-scale medical datasets are often collected from different institutions cooperated by a number of experts. In this case, we build a one-stage framework SpineOne for detecting degenerative discs and vertebrae from spinal MRIs, which implements both the keypoint localization and classification tasks simultaneously. SpineOne is a robust detector to multi-source MRI slices with various scales, numbers and quality. All four proposed machine learning approaches outperform existing baselines by a noticeable margin in specific medical tasks. In summary, four medical issues are thoroughly investigated in the thesis, i.e., the distance measurement for surgical time series with large discontinuities, the surgical region segmentation with a variety of clinician annotations, the lesion detection with limited examples and noisy bboxes, and the anatomical keypoint detection with multi-source medical data. Towards more robust medical machine learning, we then propose one robust machine learning approach for each corresponding problem.
-
ItemLearning to generalise through features( 2020)A Markov decision process (MDP) cannot be used for learning end-to-end control policies in Reinforcement Learning when the dimension of the feature vectors changes from one trial to the next. For example, this difference is present in an environment where the number of blocks to manipulate can vary. Because we cannot learn a different policy for each number of blocks, we suggest framing the problem as a POMDP instead of the MDP. It allows us to construct a constant observation space for a dynamic state space. There are two ways we can achieve such construction. First, we can design a hand-crafted set of observations for a particular problem. However, that set cannot be readily transferred to another problem, and it often requires domain-dependent knowledge. On the other hand, a set of observations can be deduced from visual observations. This approach is universal, and it allows us to easily incorporate the geometry of the problem into the observations, which can be challenging to hard-code in the former method. In this Thesis, we examine both of these methods. Our goal is to learn policies that can be generalised to new tasks. First, we show that a more general observation space can improve the performance of policies tested in untrained tasks. Second, we show that meaningful feature vectors can be obtained from visual observations. If properly regularised, these vectors can reflect the spacial structure of the state space and used for planning. Using these vectors, we construct an auto-generated reward function, able to learn working policies.