Machine learning for wavelets to enhance PET reconstruction

In the field of Nuclear Medicine, positron emission tomography (PET) plays an important role as one of the most common diagnostic tools in the area of medical imaging. However, various physical degradation factors occur when the data is recorded, which leads to a low signal-to-noise ratio. This makes the quality of the PET images less than optimal. The proposed method for solving this problem is to use a machine learning approach to find a sparse representation of the sinograms. Where a suitable sparse representation should retain as much of the signal and as little of the noise as possible. To accomplish this a sparse autoencoder was trained on wavelet decompositions of sinograms simulated from medical images in order to learn underlying structures. Three different wavelet families were tested, Daubechies 4, biorthogonal 4.4, and Haar. The trained model was able to find sparse representations of the input sinograms in the wavelet domain. Although the sparse autoencoder managed to learn the basic structures of the sinograms, it struggled with the more complex details. Compared to a conventional denoising method using hard thresholding the sparse autoencoder did not manage to produce as good of a result in terms of the reconstructed PET image quality.