Applicability test for reducing noise on PET dynamic images using phantom applying deep image prior

The complexity of imaging techniques such as Positron Emission Tomography (PET) offers a wealth of information. However, this data is often compromised by the presence of noise, which can distort our understanding and analysis of these dynamic images. In an attempt to mitigate this issue, we propose the utilization of a technique titled deep image prior (DIP). This strategy employs the advantages of artificial intelligence to enhance the quality of PET dynamic images.

Decoding Deep Image Prior

The concept of DIP was initially formulated by Ulyanov et al and makes use of an architecture commonly referred to as 'U-net'. What makes DIP particularly advantageous is its early convergence of structural shapes in images before noise, thereby achieving an improvement in noise properties without the necessity of an extensive volume of supervisory images. As such, this method proves highly practical for fields such as nuclear medicine that often only have access to one image.

Assessing the Application for DIP

While previous studies have evaluated DIP's impact on noise properties in PET and Single Photon Emission Tomography (SPECT) based images, none have specifically explored the question of the appropriate 'epoch' number for the image generating procedure. The denoising of all images involved in a PET dynamic image necessitate efficient and time-saving strategies for the process to be feasible. In this study, we introduce an index that will permit the selection of one optimal generated image during the DIP process.

Experimentation and Results

Our testing procedure made use of a previously reported phantom and PET scan data [9]. The phantom was filled with solutions containing different decay rate isotopes, ¹⁸F and ¹¹C, and the images were generated using Filtered Back Projection (FBP) and Ordered-Subsets Expectation Maximization (OSEM) algorithms. Quantitative accuracy and quality were gauged in the resultant rate constant images.

Our results showed that the decay rates on our images were not significantly different from reference values. In multiple domains, we saw lower coefficients of variances in DIP-based images than those drawn from original images. This applied to both reconstructed and decay rate images in both the ¹⁸F and ¹¹C filled regions.

Conclusions

In conclusion, our proposed method facilitates the selection of an optimal image during the DIP procedure, and is therefore a viable solution for noise reduction in dynamic PET images. This approach results in rate constant images that are less affected by noise. Further studies will focus on adjusting the level mixture around the boundary portions of the image, which hold promise for even greater accuracy and quality in PET imaging.