In oncologic imaging, hybrid Positron-Emission-Tomography/Magnetic Resonance (PET/MR) scanners offer a great potential for improving diagnostic accuracy. However, a proper patient treatment demands a high PET image quality in terms of signal-to-noise ratio and sharpness, reflecting accurate lesion detection and quantification, resulting in long PET examination times in the range of several minutes. Hence, respiratory or cardiac induced motion artifacts are inevitable. The simultaneously acquiring MR side now offers the possibility to detect and correct these induced artifacts^1-3. For this, it is essential to acquire a 4D (3D + time) MR motion model under free movement conditions as fast and accurately as possible. Furthermore, it is desired to keep the additional workload of the MR motion sequence as short as possible to keep the flexibility of acquiring further diagnostic MR sequences, which are usually performed during a PET scan. Generation of a reliable motion model, besides the MR imaging sequences, requires a simultaneous acquired surrogate marker (e.g. MR navigator or electrocardiography (ECG)) to track the true underlying motion. In order to enable a smooth processing workflow, all motion correction (MC) steps should be performed online on the scanner without the need of manual interaction. We proposed a 4D dynamic MR sequence⁴ which is able to meet the mentioned requirements. In previous studies we showed the feasibility of this method and the improvements of the PET image quality for respiratory MC^5,6. In this work, we will show the feasibility of the proposed method to perform respiratory and cardiac MC simultaneously with a reduced scan time of only 60s in the scope of a clinical setup.

The complete MC system is shown in Fig. 1 and implemented into Gadgetron⁷ for online processing.
Acquisition: MR and PET images as well as surrogate signals (MR self-navigation signal, ECG) are acquired simultaneously within the first minute (Fig. 2). For the remaining 4 minutes of the PET scan, arbitrary diagnostic MR sequences are being run, hence the MR self-navigation signal is no longer available, whereas the ECG signal is still acquired. For MR imaging, we apply a 3D spoiled gradient-echo sequence with a random variable-density Gaussian ESPReSSo subsampling⁴ (TE=1.23ms, TR=2.6ms, FOV=500x500x360mm) and for each $$$T_\text{NAV}$$$=200ms and $$$T_\text{ECG}$$$=3ms the respective surrogate signals are captured.
Reconstruction: All reconstructions steps are carried out in Gadgetron. The ECG signal is processed by means of a kernel principal component analysis⁸ to extract an ECG-derived respiration signal (EDR) which covers the complete PET examination time. With the help of the EDR signal, the missing gap of the respiratory surrogate signal can be filled in terms of a sensor fusion approach: In the first minute for which both surrogate signals (EDR and MR self-navigation) are acquired, a structure of Wiener filters is trained to learn a joint respiratory representative, which better reflects the underlying respiratory motion. These filters then estimate a continuous respiratory surrogate signal from the EDR signal for the whole examination time. Separate cardiac and respiratory gating is performed with adaptable respiratory and cardiac view sharing amongst neighbouring gates. The cardiac gates are determined via a modified QRS complex detector⁹ with arrhythmia control. Within a Compressed Sensing reconstruction of the gated subsampled MR data⁵, we perform a non-rigid multilevel cubic B-Spline registration¹⁰ for the respiratory and cardiac phases separately. This has the advantage of fully utilizing all acquired samples by e.g. combining all cardiac gates for the respiratory registration and vice versa, yielding a more accurate deformation field. The respiratory and cardiac motion models are then weighted and linearly combined. This motion model is applied to the gated PET images together with an MR-based motion-corrected attenuation map¹¹ to reconstruct a motion-corrected (end-expiratory transformed) PET image by means of a 3D-OSEM with 2 iterations, 21 subsets and 4 mm Gaussian filter.
Coronal in-vivo patient data were acquired for 25 patients (14 female, age 60.5 +/- 9.8) with suspected lung or liver metastasis and myocardial FDG uptake on a 3T PET/MR (Biograph mMR, Siemens). ROIs and lines were placed on moving lesions of the corrected, uncorrected and end-expiratory gated PET image.

A proper 4D gated (4 respiratory and 8 cardiac gates) MR image can be reconstructed in a short scan time, resulting in a reliable motion model which clearly improves the obtained PET image quality (Fig. 3). These results are supported by the extracted PET metric values of the moving lesions as percentage improvements (Table 1). In conclusion, a clinical feasible respiratory and cardiac PET/MR motion correction system for improved diagnostic accuracy is presented.