As a new form of magnetic resonance imaging (MRI) contrast enhancement scheme, temporal chemical exchange saturation transfer (CEST) can provide rich pathological and pharmacokinetic (PK) information. However, the detailed analysis of temporal CEST data is significantly hampered by low signal-to-noise ratio (SNR) conditions. Here, pruned exact linear time (PELT)-based change point detection (CPD) was innovatively applied to simulated and experimental temporal CEST data to obtain change point distributions. Excitingly, different simulated PK scenarios (i.e. curve types) gave rise to distinct change point distribution patterns and similar change point distribution patterns were observed for dynamic glucose-enhanced (DGE) CEST data, validating the applicability of PELT-based CPD to temporal CEST data. The application of PELT- based CPD to extremely noisy temporal thiol-CEST data provided interesting findings similar to our previous unpublished results using a different time series analysis method, thanks to its great noise tolerance also demonstrated on simulated data and experimental data with added pseudo-white noise. The application of CPD to temporal CEST data (i.e. CEST time series) essentially reveals clustering patterns of the time series analysed. Principal component analysis (PCA)-K-means based clustering analysis (a commonly used unsupervised machine learning (ML) technique) was also applied to CEST time series and their corresponding change point "spectra". The clustering analysis of DGE CEST time series allowed the differentiation between positive and negative area under curve (AUC) regions in an DGE CEST image while the analysis of the corresponding change point "spectra" gave rise to a more complicated clustering pattern that needs iv further analysis. To further explore the applicability of ML in the analysis of CEST time series, a bidirectional long short-term memory (Bi-LSTM) deep neural network was trained on simulated CEST time series of different curve types. The trained Bi-LSTM model performed well on the test dataset containing only simulated time series. However, discrepancies were observed between the normalised experimental time series of each curve type determined by Bi-LSTM and the representative simulated time series of each curve type from the training data set. This illustrates the potential misclassification of the experimental CEST time series, leading to the consideration of the including experimental time series in the training and test datasets.
Change point detection and machine learning based analysis of temporal CEST MRI data
Pan, C. (Author). 2025
Western Sydney University thesis: Master's thesis