Page 137 - GTM-4-1

Global Translational Medicine MRgFUS sonication parameters prediction

Figure 2. Neural network training

Table 2. Parameters of a multilayer neural network Figure 4 presents a density plot of residuals,
(Model: “sequential_3”) for temperature prediction during demonstrating that the neural network model has errors
sonication closer to zero. For practicality, if access to a computer with
R and libraries is not available, the linear model coefficients
Layer (type) (Output, shape) Parameter can be entered into Microsoft Excel, as shown in Figure 5.
dense_15 (Dense) (None, 64) 448
re_lu_3 (ReLU) (None, 64) 0 4. Discussion
dense_14 (Dense) (None, 128) 8,320 MRgFUS treatment is a promising innovative treatment
dropout_3 (Dropout) (None, 128) 0 for movement disorders, such as essential tremor and
dense_13 (Dense) (None, 64) 8,256 advanced PD. However, skull barriers can prevent sufficient
dense_12 (Dense) (None, 1) 65 temperature elevation in the target area. In a retrospective
39
Total parameters 17,089 analysis of 25 patients, Chang et al. found that the skull
Trainable parameters 17,089 volume and the SDR were associated with the maximum
temperature achieved during MRgFUS treatment. These
Non-trainable parameters 0 findings may help identify eligible candidates for MRgFUS.
Abbreviation: ReLU: Rectified linear unit.
D’Souza et al. investigated the impact of cranial bone
40
characteristics on the efficacy and safety of procedures
represents the predicted temperature. The training matrix employing an integrated 3T MRgFUS system. The analysis
is shown in the left column, and the testing matrix is shown included data from 189 patients diagnosed with essential
in the right column. The linear model is displayed at the tremor, with particular emphasis on a subgroup exhibiting
top (Figure 3A), and the neural network model is displayed low SDR. The results indicated a significant difference in
at the bottom (Figure 3B). The linear model in the training the probability of reaching the target temperature of 54°C
sample has an MAE of 2.32, an RMSE of 3.1, an error rate between patients with SDR values below 0.45 and those
of 6.3%, and an R value of 0.71. In the test sample, the with SDR values of 0.45 or higher.
2
MAE is 2.2, the RMSE is 3.12, and the error rate is 6.0%.
In a study of 270 patients, researchers compared
41
The graphs indicate that the neural network model outcomes based on SDR. Patients were categorized into low
has less residual dispersion than the linear model. This is SDR (<0.40, n = 31) or high SDR (≥0.40, n = 239) groups.
supported by the RMSE parameter, which was higher for A matched-case analysis (n = 56) showed that patients
the linear (3.12) than the neural network model (2.56). The with low SDR needed greater sonication power and
maximum prediction error was 17.8°C for the linear model energy, leading to smaller lesion volumes. Despite these
and 12.1°C for the neural network model, indicating that differences, both groups achieved similar tremor control.
the neural network performed better. While patients with low SDR experienced slightly more

Volume 4 Issue 1 (2025) 129 doi: 10.36922/gtm.5419