Thermochemical ablation (TCA) is a novel minimally-invasive ablation approach¹. Acetic acid and sodium hydroxide are injected simultaneously, mixing and reacting directly before entering the tissue. The exothermal reaction releases heat that is used for thermal ablation. Additionally, the reaction product sodium acetate (NaOAc) creates a hyperosmolar area at the target site. Recently, ²³Na MRI was used to assess sodium acetate distributions after injection². However, for more detailed characterization of TCA injection, 4D data with reasonable temporal resolution are required. In this work, a compressed sensing³ approach was applied to acquire 4D data of injections with high spatial and good temporal resolution.

Experiments were performed on a 7T whole-body MR system (Magnetom 7T, Siemens Healthcare, Erlangen, Germany). A double-resonant (1H: 297.2 MHz; 23Na: 78.6 MHz) quadrature birdcage coil (Rapid Biomed, Rimpar, Germany) was used. The data were acquired with a 3D density adapted radial sampling pattern⁴ with a golden angle distribution of the projections⁵.

An ex vivo bovine liver was embedded into agarose gel next to a reference tube with NaOAc solution (2.5 M). During the experiment, NaOAc solution (3 ml, 2.5 M) was injected into the liver through a 20 G angio catheter (B. Braun Melsungen, Melsungen, Germany) using an MR contrast agent injector (Accutron Injector MR, Medtron, Saarbrücken, Germany).

During and after the injection experiment, two datasets (DS_injection, DS_{post-injection}) with a total number of 50,000 projections each were acquired. Subsets of DS_injection with 2500 projections were used for the reconstruction of each time frame, resulting in a temporal resolution of T_aq=1.04min and an undersampling factor USF≈28. The data subsets were interleaved so that a new timeframe begins every 1250 projections (31.25sec).

The data were reconstructed with Nonuniform Fast Fourier Transform (NUFFT)⁶ and iterative 3D-Dictionary-Learning Compressed Sensing reconstruction (3D-DLCS)⁷. In the 3D-DLCS reconstruction, an adaptive dictionary consisting of D=300 three-dimensional blocks of size B=3×3×3 pixels was used; the dictionary is initialized with an overcomplete discrete cosine transform and updated in each iteration step with the k-singular-value-decompositions algorithm^8,9.

The reconstruction was evaluated with simulated data DS_simulated obtained from DS_{post-injection}. A dataset containing 2500 projections was created with added Gaussian noise to achieve a SNR similar to the measurements. Mean normalized signal intensity relative to the reference tube within a region-of-interest (ROI) and peak signal-to-noise ratio (PSNR) were computed for the NUFFT and the 3D-DLCS reconstructions. The reconstructions were performed on a standalone desktop PC (Intel Core i7-2600 CPU, 3.4 GHz, 16 GB memory)

The 3D-DLCS reconstructions of both the simulated data (DS_simulated) and acquired (DS_injection) data show a strong reduction of noise when compared with the corresponding NUFFT reconstructions. Most notably, low intensity regions that are not recovered in the NUFFT reconstruction become discernable in the 3D-DLCS reconstruction (arrows in Fig. 2). However, slight blurring can be seen when compared to the ground truth. For the 3D-DLCS reconstruction, the PSNR is improved by 18.6dB and the mean intensity within the ROI delineated in Fig. 1a is much closer to the value from the ground truth (see Table 1). The 3D-DLCS reconstruction of each time frame in the dynamic dataset (DS_injection) took less than 10min, resulting in a total reconstruction time of 6.5h for all 39 time frames.The very sparse nature of the sodium images from the TCA experiment make it well suited for Compressed Sensing based reconstructions such as the 3D-DLCS algorithm. Because the ²³Na concentrations to be measured are much higher compared to physiological concentrations in tissue, relatively high undersampling can be used without having to acquire multiple averages. The strong reduction of artifacts in the 3D-DLCS reconstruction compared to NUFFT with such a high undersampling factor (USF = 28) is feasible because few entries from the dictionary are needed in order to represent each block of the image. Our work indicates that 4D sodium imaging can be employed to characterize TCA procedures retrospectively.