Neural network reconstruction of single-photon emission computed tomography images

Area of Science:

• Medical Imaging
• Artificial Intelligence in Medicine
• Computational Science

Background:

• Traditional image reconstruction techniques for tomography can be computationally expensive.
• Artificial neural networks (ANNs) show potential for accurate image reconstruction from tomographic data.
• Previous work demonstrated ANN capability for reconstructing sections of single photon emission computed tomography (SPECT) images.

Purpose of the Study:

• To develop and present a method for deriving ANN activation functions tailored for full SPECT image reconstruction.
• To compare the trainability and generalization ability of a statistically tailored ANN with standard methods.
• To achieve accurate tomographic image reconstruction with reduced computational cost.

Main Methods:

• Developed a backpropagation ANN utilizing activation functions derived from the probability density functions (PDFs) of training data.
• Trained and evaluated the statistically tailored ANN against a standard sigmoidal backpropagation ANN.
• Assessed performance based on trainability, generalization ability, and reconstructed image quality.

Main Results:

• The statistically tailored ANN demonstrated comparable image reconstruction quality to the training data.
• The method showed high precision in learning the planar data-to-tomographic image relationship.
• The ANN achieved accurate reconstruction without the high computational cost of some traditional methods.

Conclusions:

• Statistically tailored activation functions enable effective training of ANNs for high-fidelity SPECT image reconstruction.
• This ANN approach offers a computationally efficient alternative for accurate tomographic image reconstruction.
• An adequately trained ANN can compensate for physical effects like photon transport, noise, and artifacts.