Unit:
Κατεύθυνση Ιατρική Φυσική-ΑκτινοφυσικήLibrary of the School of Health Sciences
Author:
Tomazinaki Mina Ermioni
Supervisors info:
Ευστάθιος Στυλιάρης, Αναπληρωτής Καθηγητής, Τμήμα Φυσικής, ΕΚΠΑ
Παντελής Καραΐσκος, Καθηγητής, Ιατρική Σχολή, ΕΚΠΑ
Κωνσταντίνος Λουκάς, Επίκουρος Καθηγητής, Ιατρική Σχολή, ΕΚΠΑ
Original Title:
Artificial Neural Networks in Single Photon Emission Tomography (SPECT)
Translated title:
Artificial Neural Networks in Single Photon Emission Tomography (SPECT)
Summary:
The recent advances in Artificial Intelligence (AI) and Machine Learning (ML) have affected Medical Physics and tend to set new frontiers in Image Reconstruction in Medicine. Nowadays modalities of tomography, such as SPECT, make use of software novelties since the current hardware have reach an upper limit. The aim of this study focuses on new approaches in Tomographic Image Reconstruction exploiting Radon Transform symmetries and novel training structures based on Artificial Neural Networks (ANNs). Having in mind that the mathematical transform of the tomographic process must be preserved, by keeping the simplicity and the strictly labeled service that each layer could provide in the network, the following conditions should be fulfilled: The input data always has to be characterized by the sinogram of the image while the output has to reflect information about the image’s representation. Instead of using the conventional sinogram technique, a new approach has to be introduced. A simple remapping of the well-known sinogram free of dependencies of the images’ features could lead to reconstructed images of diagnostic power and low time cost. The efficiency of this altered sinogram is tested as input of various ANN architectures in the current work in order to predict pixel values which guide the full image reconstruction. The proposed prototype is implemented in combined techniques with well-established algorithms, such as Algebraic Reconstruction Technique (ART), leading to accurate results characterized by low chi- square values, even with limited available data. The results of this work as a prototype study could serve further research with ANN involvement for accurate Image Reconstructions of high clinical value.
Main subject category:
Health Sciences
Keywords:
Artificial Neural Networks, Tomographic Image Reconstruction, Radon Transform, SPECT
