$\mu$-Map-Guided Attenuation Correction of PET Images Based on Deep Learning

Attenuation Correction (AC) for PET images is typically achieved through the of attenuation maps (<tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\mu$</tex> maps) to improve the quality of generated PET images. However, existing deep learning-based methods for PET attenuation correction often directly generate the corrected images, but this generation process lacks the attenuation maps' constraint. As a result, the quality of the generated images is suboptimal. To address this issue, we design a generation model that can perform reverse verification by generating the corresponding <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\mu$</tex>-map, thereby improving the accuracy of the attenuation correction. Additionally, a local spatial attention mechanism is incorporated into the model to adaptively focus on key regions, enhancing the details of generated PET images. Specifically, we first design a PET-to- <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\mu$</tex>-map generation model, which compares the generated <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\mu$</tex>-maps from PET images with the corresponding real <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\mu$</tex>-maps, thereby optimizing the model's generation process. Next, we further incorporate a local spatial attention mechanism into the model, enabling it to dynamically adjust attention weights and focus on critical areas of the PET images, improving the details and structural information of low-quality PET images. Finally, we conduct experiments on the test set and compare the results using relevant evaluation metrics. The experimental results demonstrate that our model generates images with higher similarity to the real PET images.