A decade has passed since the groundbreaking work by Defrise et al. (2012), which demonstrated that TOF PET imaging is self-correcting for a variety of physical degradation factors, most notably attenuation correction, which is currently based on CT imaging in PET-CT scans. At that time, the acronym for Maximum-Likelihood reconstruction of Attenuation and Activity, MLAA, became widely used in application papers, and other sophisticated optimization algorithms began to emerge. Nevertheless, nowadays an attentive observer will find that self-attenuation correction has become a secondary topic compared to the generation of attenuation maps using Deep Learning from non-attenuation-corrected PET images, aided by any other available information. The synthesis of attenuation maps has become an image-to-image task, well-suited for convolutional neural networks (CNN), generative adversarial networks (GAN), or even diffusion models and benefit from wide use of PET-CT imaging for generating vast amounts of training data, although given the sequential nature of PET and CT misalignments are possible.The key question now is: what level of detail from attenuation information can be recovered from suboptimal PET image? This certainly showcases the "power" of deep learning. But does it mean that human AI-developed over centuries of observation and modeling of nature-will remain in the shadow of more generic, powerful machine learning methods? We will attempt to shed light on this issue, though perhaps only the coming years will offer a clear answer.
Teimoorisichani et al. (Thu,) studied this question.