University of Twente Student Theses


Machine learning in nuclear medicine: detection of Parkinson's disease, myocardial ischemia and major adverse cardiac events using support vector machines

Dotinga, Maaike (2019) Machine learning in nuclear medicine: detection of Parkinson's disease, myocardial ischemia and major adverse cardiac events using support vector machines.

[img] PDF
Abstract:The use of machine learning (ML) in medical imaging is rapidly increasing. Its ability of mapping the relationship between a defined input and output through learning by examples makes ML useful for a wide range of applications in radiology and nuclear medicine. As multiple parameters can be integrated into a single outcome, it particularly has great potential to enhance diagnosis and prognosis to facilitate patient management. In this thesis, we focus on the use of support vector machines (SVMs) to detect Parkinson’s disease (PD), myocardial ischemia and major adverse cardiac events (MACE). In the first study, we aimed to develop and validate a linear SVM classifier to discriminate PD from non-PD based on I-123 FP-CIT SPECT striatal uptake ratios, age and gender. Its generalizability was assessed using previously unseen datasets from two centers using comparable acquisition and image processing methods, thereby comparing prediction performances of the derived model between sites. Expert nuclear medicine physicians scored I-123 FP-CIT SPECT scans of both datasets as either PD or non-PD after which their prediction performance was compared to that of the SVM model to assess its clinical value. We found that comparable prediction performances were obtained between sites with classification accuracies of 95.0% for the dataset from the same center the model was developed and 82.5% for the other. The performances found were similar to that of nuclear medicine physicians who achieved accuracies of 95.0% and 81.3%, respectively. Using the derived SVM, accurate discrimination of PD from non-PD can be achieved that is equivalent to standard visual assessment by expert nuclear medicine physicians. Furthermore, we can assume that the model is generalizable towards centers using comparable acquisition and image processing methods. Hence, implementation of this SVM model as diagnostic aid in clinical practice is encouraged. In the second study, two Gaussian SVM classifiers were developed to identify patients with myocardial ischemia and patients at risk of MACE. The derived models were subsequently validated using a previously unseen dataset. Input features included various clinical parameters and coronary artery calcium score (CAC) for both models and for the MACE model, left ventricular ejection fraction and myocardial perfusion imaging (MPI) SPECT scan outcome were added. The ischemia model was further evaluated by comparing its prediction performance to that of absent CAC indicative of the absence of ischemia. Validation of the ischemia model led to a sensitivity and specificity of 89.7% and 31.8%, respectively. A comparable prediction performance was found for predicting ischemia using absent CAC, obtaining a sensitivity of 84.1% and specificity of 37.1%. We observed that the MACE model was not generalizable towards previously unseen data as specificity decreased substantially from 16.5% to 3.1%. The higher amount of input features needed to predict ischemia in comparison to standalone CAC scoring and the low specificity of the MACE model impede clinical implementation of these models. Further evaluation of both models is therefore needed to be able to provide a more individualized risk assessment of ischemia and MACE using ML.
Item Type:Essay (Master)
Isala, Zwolle
Faculty:TNW: Science and Technology
Subject:44 medicine
Programme:Technical Medicine MSc (60033)
Link to this item:
Export this item as:BibTeX
HTML Citation
Reference Manager


Repository Staff Only: item control page