An investigation of the diversity of outcomes along with machine learning in the prediction of ischemia using PET cardiac perfusion imaging

Aug 6, 2026

Positron emission tomography (PET) provides precise, non-invasive visualization of cardiovascular function, and ischemia—insufficient blood and oxygen supply—requires accurate detection to reduce serious risk. This thesis investigates how segmentation and analysis of PET cardiac perfusion images on Carimas vary across medical and non-medical backgrounds, and how those differences influence myocardial blood flow (MBF)-based ischemia prediction. Polar maps were generated from 55 stress perfusion images, and MBF comparisons used statistical testing while a pre-trained CNN was validated via k-fold to assess classification performance.

LanguageEnglish

CategoryResearch & Report

FormatPDF

Pages68

Reading Time2 h 51 min

AccessFree to download

Literature Overview

Introduction

Thesis Structure

Motivation

Previous studies

Aims and Research Questions

Aims of the study

Research questions

Background

Positron Emission Tomography

Myocardial Ischemia

Carimas

PET imaging for the diagnosis of myocardial ischemia

AI for Cardiovascular diseases

Image classification and ML models

Materials and Methods

Patient selection

Image acquisition

Data acquisition protocols

Workflow

Statistical tests

Coding

Results

Medical vs Non-Medical Expert

Classification Results

K-fold results

Discussion

Future studies

Conclusion

Acknowledgement

References

An investigation of the diversity of outcomes along with machine learning in the prediction of ischemia using PET cardiac perfusion imaging Naipunya Guruprasad Master’s thesis University of Turku, Turku PET Centre Faculty of Medicine, Institute of Biomedicine 06.05.2024 Master´s degree in Biomedical Imaging Research field: PET imaging and image analysis using machine learning. Credits: 40 ECTS Supervisors: 1. Dr. Jarmo Teuho Adjunct Professor, Turku PET Centre, University of Turku 2. Prof. Riku Klén Assistant Professor (Imaging instrumentation and detection technologies), Turku PET Centre, University of Turku Abstract University of Turku, Turku PET Centre Faculty of Medicine, Institute of Biomedicine Naipunya Guruprasad: “An investigation of the diversity of outcomes along with machine learning in the prediction of ischemia using PET cardiac perfusion imaging”. Master’s thesis, 62pp. Master’s Level May 2024 Positron emission tomography (PET) is an advanced, non-invasive medical imaging technology that allows for precise internal imaging. The recent progress in PET technology, particularly in cardiovascular applications, has significantly improved our knowledge of heart-related conditions. PET scans offer clear visuals of blood circulation and cardiac functions, playing an important role in detecting various conditions. One such condition is ischemia, a state where apart of the body lacks sufficient blood and oxygen supply, posing serious risks. This can be fatal and hence the detection of this is incredibly essential. For accurate predictions of such conditions, healthcare professionals rely on specialised software dedicated to segmentation and visualization of PET cardiac perfusion images, and one such software is Carimas. The integration of predictive machine learning models with Carimas not only enhances but also confirms the accuracy of diagnostic findings, revolutionizing the way we approach cardiac care. The objective was to have individuals from medical and non-medical backgrounds perform segmentation/analysis on Carimas and to then compare them with one another. Additionally, a pre-trained machine learning model uses the data created on the software to predict whether the patient is ischemic or not. This strategy offers a thorough insight of how segmentation techniques used by people with various backgrounds affect the prediction of ischemia, as well as machine learning models. 55 PET cardiac stress perfusion images from ischemic and non-ischemic patients have been used to create polar maps with radiowater [ 15O-H2O] labelling. The Myocardial Blood Flow (MBF) data is compared using statistical tests like the Wilcoxon test, Jaccard Index and Dice Coefficient. The pre-trained The Convolutional Neural Network (CNN) model undergoes K-fold validation to verify its performance. Our study aimed to evaluate the accuracy of ischemia prediction derived from manual segmentation performed by individuals with diverse backgrounds indeed gives different MBF data which in turn affects the classification of patients. The CNN model, on the other hand, shows a high area under the characteristic curve value indicating good performance that is independent of the level of expertise of the individuals who created the polar maps. Keywords: Ischemia, PET Imaging, Carimas, Convolutional neural networks. Table of Contents 1. Literature Overview…………………………………………………………………………. 1 1.1 Introduction………………………………………………………………………………. 1 1.2 Thesis Structure………………………………………………………………………….. 2 1.3 Motivation………………………………………………………………………………... 3 1.4 Previous studies…………………………………………………………………………. 4 2. Aimsand Research Questions………………………………………………………………. 6 2.1 Aims ofthe study…………………………………………………………………………6 2.2 Research questions……………………………………………………………………….. 6 3. Background…………………………………………………………………………………. 7 3.1 Positron Emission Tomography………………………………………………………… 7 3.1.1 15O-water asa tracer……………………………………

FAQs

What is the main objective of the thesis?

To assess whether manual segmentation performed by individuals with diverse backgrounds leads to different MBF data and, consequently, impacts ischemia prediction. It also evaluates whether a CNN classification model performs robustly regardless of the experts’ segmentation background.

How were PET cardiac perfusion images used in the study?

Fifty-five PET cardiac stress perfusion images from ischemic and non-ischemic patients were processed to create polar maps using 15O-water labeling. MBF data were then compared using statistical tests such as Wilcoxon, Jaccard Index, and Dice Coefficient.

What method validated the machine learning model’s performance?

A pre-trained convolutional neural network (CNN) underwent k-fold validation to verify classification performance for ischemia prediction.

Aug 6, 2026

Upload by Emma Wilson