Laser Thermal Ablation: Model and Parameter Estimates to Predict Cell Death from MR Thermometry Images

To correlate the model predicted regions of cell death with the tissue response, we created a thermal lesion in seven in vivo rabbit brains. We used a clinical 0.5 T MR imaging system with an extremity coil to guide a laser fiber into each brain, and continuously acquire gradient-echo (GE) MR images (TR = 77.2 msec; TE = 38.9 msec; flip angle = 30°; 256 x 128 matrix; 16 x 16 cm field of view (FOV); one 3.0 mm thick section, 10 sec acq. time) before, during, and after heating. At four hours post-ablation, we acquired T2-weighted spin-echo MR images (TR = 4000 msec; TE = 115 msec; 512 x 256 matrix; 16 x 16 cm FOV; one 2.0 mm thick section) in the same orientation as the GE MR images. Lesion formation was achieved with an Nd:YAG laser. Heating durations varied between 30 and 581 sec. To process the temperature maps from the GE MR images, we subtracted pre-ablation baseline phase maps from remaining phase maps. We used a proton resonance frequency thermal coefficient of 0.01 ppm/°C. We removed noise in the temperature maps with a temporal filter. To compare model predicted regions of cell death with the tissue response, we aligned the post-ablation MR image to the GE MR images used for temperature maps with a rigid-body registration that aligned fiducials near the lesion. Results In Figure 1, the post-ablation T2-weighted MR lesion images show a distinct circular hyperintense rim surrounding a central hypointense core. It was previously shown that the outer boundary of the hyperintense rim corresponds to the boundary of cell death as seen in registered histology images on the order of one MR voxel (0.70 mm) (1). We manually segmented the boundary of cell death in the registered post-ablation MR image, and created a binary image of the cell death region to compare with the modeled tissue damage region on a voxel-by-voxel basis. Model parameters were simultaneously estimated with an iterative optimization using every interesting voxel (4375 voxels) in 328 temperature images from the seven experiments. In Figure 2, we plotted as a function of lesion, the number of false positives (FP) and false negatives (FN). The number of FP and FN were small as compared to the size of the actual cell death region. For a necrotic region of 766 voxels across all lesions, the model provided a voxel specificity and sensitivity of 98.1% and 78.5%, respectively. Mislabeled voxels were typically within one voxel from the segmented necrotic boundary with median distances of 0.77 mm and 0.22 mm for FP and FN, respectively. We compared our model to the critical temperature model that assumes cell death is not observable below a critical temperature and occurs rapidly and completely above the critical temperature (2). Across all lesions, our model predicted 13 fewer FP voxels (1.8 million cells) and 57 fewer FN voxels (8.2 million cells), with the number of cells in a voxel based on a cubic cell with a 20 m edge length. Discussion We can compare our model and data analysis technique to previously reported ones. Arrhenius-based models with parameters from other experiments not surprisingly did not always work (2). A critical temperature model typically performed better (2-3). However, this model neglects the heating duration and is sensitive to transient noise in the temperature data. We use a model that considers the temperature history, and had fewer errors than the critical temperature model. This model in principal will be able to predict cell death for a wider range of temperature histories. Our analysis method uses all interesting voxels unlike previous reports which only analyzed voxels along the segmented cell death boundary (2-4). Hence, in theory, we can more precisely assess the model fit error. We conclude that our model coupled with a sequence of MR temperature maps can be used to accurately predict the tissue response. Features such as accurate image registration, filtering, and parameter optimization are important steps to accurately fit the model to the segmented region of cell death. Results show that for rabbit brain, the estimated region of necrosis closely corresponds to the actual cell death region. This is good evidence that our model can predict the therapeutic region.