Unifying Global and Local Statistical Measures for Anatomy-Guided Emission Tomography Reconstruction

摘要

Some tumours and lesions do not have a boundary in the anatomical image that matches their functional boundaries. Therefore, most anatomical priors yield little to no added value for reconstructing these features compared to conventional priors. In this work, we propose a new joint classification and reconstruction framework to capture the underlying functional and structural information and exploit it to enhance the signal-to-noise ratio (SNR) in these features during emission tomography (ET) reconstruction. As a proof of concept, a lesion with 50% reduced activity was inserted in the gray matter (GM) of a realistic 3D PET brain phantom. The activity was reconstructed with the proposed algorithm, as well as with the earlier validated asymmetrical Bowsher prior, using a perfectly aligned, simulated MR as anatomical information. Next, twenty subtle hypointense lesions were introduced in the GM, again invisible in the MR. The optimized reconstructions obtained with the new method were compared to those obtained with asymmetrical Bowsher and A-MAP. The SNR in the lesions was plotted versus the bias on the lesion signal. With the new algorithm sharper lesion boundaries were observed compared to the other two methods. In addition, it outperformed the asymmetrical Bowsher prior in terms of SNR in the lesions at the same (low) bias level. However, higher SNR was obtained with A-MAP at all bias levels and similar SNR was reached by the asymmetrical Bowsher prior if higher bias on the signal is allowed. This new anatomy-guided reconstruction algorithm looks promising for improving the SNR and lesion detection in e.g. PET brain imaging compared to other anatomical priors, but needs further investigation. It has the additional advantage of estimating the underlying tissue classes jointly from the functional and anatomical information, such that errors in the a priori segmentation are expected to cause less artifacts than methods relying on a fixed predefined segmentation.