Page 139 - GTM-4-1

Global Translational Medicine MRgFUS sonication parameters prediction

was not an issue. For the entire cohort, the MAE was 2.29,
the RMSE was 3.13, and the error rate was 6.2%.
In addition, we chose to use a deep neural network
for its ability to perform optimal computations based on
the input data, regardless of regression complexity. Using
a CNN with hidden layers, the model showed promising
results with an MAE of 1.93, an RMSE of 2.69, and an error
rate of 5.37%. Thus, the consistency of the neural network
was validated using the training dataset. Overall, with
Figure 5. An Excel spreadsheet for automatic temperature calculation an MAE of 1.93°C and an R² of 0.76, the neural network
Note: ALIGN refers to the first sonication.
outperformed the linear models in temperature prediction.
demographic factors. Boutet et al. concluded that while Our study further supports previous findings,
SDR impacts the energy needed during sonication, it demonstrating that the temperature at the target during
does not accurately predict treatment success in the study ultrasonic heating is influenced not only by power, energy,
population. They proposed that a broader assessment, and exposure time but also by the patient’s body mass
including local variations in bone density, is necessary index, sex, and age, which are likely related to the age and
to determine which patients are good candidates for sex characteristics of the tissues.
MRgFUS.
However, our study has several limitations. First, it is
Yuen et al. explored how different skull measurements, based on data obtained from 152 patients who underwent
43
especially the SDR, relate to the success of MRgFUS the MRgFUS procedure. These patients had different
treatment for essential tremor. They reanalyzed the records conditions, such as essential tremor, PD, and various
of 62 patients treated at the Mayo Clinic from 2017 to forms of dystonia. Moreover, the distribution of patients
2021, examining the association between skull metrics and by nosology may not account for other potential factors
treatment details, such as the highest power and energy affecting accurate temperature calculation. We did not
used. The findings show that while SDR is a key predictor explore factors such as race, population demographics, or
of treatment success, other skull features also play a the impact of certain medications on MRgFUS treatment.
significant role in how the treatment is delivered. Machine While the study identified several parameters that
learning methods were used to improve predictions of significantly affect treatment, there may be other important
treatment outcomes, with some success. The study results variables not included in the models that could improve
indicated that while SDR was a strong indicator, combining the prediction accuracy.
it with other skull metrics could further improve predictive
accuracy. Yuen et al. concluded that patient selection 5. Conclusion
for MRgFUS should consider multiple skull features,
acknowledging that factors beyond just SDR influence Sex, age, SDR, and the initial tissue response to the first
treatment outcomes. sonication—specifically the energy delivered during this
initial treatment and the corresponding temperature—
In our study, we used three linear models and a CNN are critical factors in optimizing subsequent sonication
to effectively compare and determine the most accurate parameters, such as power, energy, and duration.
model for temperature prediction. In the first model, the
RMSE was 3.68, the R² was 0.62, and the MAE was 2.78. Acknowledgments
On the training set, the RMSE was 3.55, with an R² value We thank the Research Center of Neurology, Moscow,
of 0.64 and an MAE value of 2.67. In the second model, we Russia, the National Society of Movement Disorders and
identified power sonication duration, early termination, Parkinson Disease Research, and Bashkir State Medical
bone tissue coefficient, initial sonication parameters, as well University for academic support.
as results, age, and sex, as the key factors that significantly
affect temperature within the model (P < 0.05). The third Funding
model, created by eliminating redundant parameters,
tended to underestimate temperature. In the test sample, None.
the MAE was 2.2, the RMSE was 3.12, and the error rate Conflict of interest
was 6%. In the training sample, the MAE was 2.32, and
the RMSE was 3.14. The low RMSE differences between Naufal Zagidullin is an Editorial Board Member of this
training and test samples indicate that model overfitting journal and Guest Editor of this special issue but was not in

Volume 4 Issue 1 (2025) 131 doi: 10.36922/gtm.5419