Convolutional neural networks for myocardial perfusion SPECT imaging classification: a full and low-dose study

• Main Author: Gonçalves, Gonçalo Cravo de Jesus
• Publication Date: 2023
• Format: Master thesis
• Language: eng
• Source: Repositórios Científicos de Acesso Aberto de Portugal (RCAAP)
• Download full: http://hdl.handle.net/10400.21/17628
• topic: Deep learning; Convolutional neural networks; Image classification SPECT; Myocardial perfusion imaging; Nuclear medicine; Aprendizagem profunda; Redes neuronais convolucionais; Classificação de imagens SPECT; Cintigrafia de perfusão do miocárdio; Medicina nuclear; MEB

ABSTRACT - Myocardial perfusion imaging (MPI) by single-photon emission computed tomography (SPECT) plays a crucial role in the diagnosis of coronary artery disease. Moreover, the quantification of these images typically involves the extraction of quantitative parameters obtained from the rest stress perfusion. However, the acquisition systems have some limitations such as spatial blurring and low-count data, which may introduce bias in the classification. Additionally, these processes are time-consuming and user-dependent, leading to significant intra and inter-operator variability. Furthermore, over the years there has been a constant effort to reduce the dose of MPI. In this sense, both the variability classification issues and the dose reduction concerns can impact the true assessment of SPECT-MPI. In recent years, with the rise of artificial intelligence algorithms, several studies have proposed automatic Deep Learning techniques for the classification of MPI, moreover regarding low-count data. In this project, we ran 5 Convolutional Neural Network models with pre-trained weights: one trained on stress real full-time data (100%, as 100R), three individual models with synthetic 75%, 50%, and 25% count settings, and another one with all datasets combined (ALL). Thus, we compared their performance when tested on full-time and low-time studies and assessed the application of synthetic subsampled data from the Poisson Resampling technique in SPECT-MPI classification tasks. In conclusion, both 100R and ALL models achieved good and similar results when tested in real full-time (the 100R model achieved an accuracy of 0.70 and the ALL model achieved an accuracy of 0.65) and real low-time at 75% (both models achieved an accuracy of 0.71). Bellow this percentage, the models’ accuracy began to drop, possibly due to the limited information these images contain. Thus, subsampled data from a Poisson resampling method may be a possible solution to conduct further studies regarding the classification of low-time SPECT-MPI.