In the recent publication in the journal Physics in Medicine and Biology, MIMRTL team members Xue Li and Alan McMillan reported a deep learning approach to perform attenuation correction (AC) in breast PET/MR imaging. This method aims to overcome significant challenges associated with traditional PET/MR imaging, such as anatomical truncation and the absence of bone information, which can compromise the accuracy of attenuation correction and, subsequently, the diagnostic utility of PET/MR in breast cancer care. A link to the paper is available here: https://doi.org/10.1088/1361-6560/ad2126
The core of the research involves the development and application of deep learning models designed to generate synthetic CT (sCT) images from uncorrected nonattenuation corrected or NAC PET/MR data. The study employed a deep learning mean absolute error model, a deep learning mean squared error (MSE) model, and a U-Net architecture with perceptual loss to predict sCT images. These models were trained and validated using data from 23 subjects diagnosed with invasive breast cancer. The results demonstrated that the sCT images from the MAE, MSE, and perceptual loss models were similar in mean absolute error (MAE), peak signal-to-noise ratio, and normalized cross-correlation, indicating no significant difference in SUV accuracy between the PET images reconstructed using the DLMSE and DLPerceptual sCTs compared to those from the reference CT for attenuation correction in all tissue regions. Most notably, all deep learning methods outperformed the traditional PET/MR Dixon-based method according to SUV analysis.
Other co-authors on the paper include: Jacob M Johnson, Roberta M Strigel, Leah C Henze Bancroft, Samuel A Hurley, S Iman Zare Estakhraji, Manoj Kumar, and Amy M Fowler. This research showcases the applicability of deep learning in medical imaging but also opens new avenues for improving the precision and effectiveness of PET/MR imaging, ultimately contributing to better outcomes for patients undergoing breast cancer care.