Introduction

Time-efficiency and radiofrequency safety are crucial for MRI protocols. Although MR Thermometry techniques exist, including for example T1 relaxation and proton resonance frequency shift, their low-sensitivity increase the acquisition time¹. We suggest a personalized, non-invasive approach: knowing the tissue properties, distance to several surface temperature sensors and the surface temperature, we seek to accurately predict the internal body temperature. An approach through deep learning is proposed. In a brain slice, we consider N points where we want to know the temperature and we consider N_s surface temperature sensors. For each point N_p in N, we consider an additional set of i equidistant points placed on an imaginary line between one surface sensor and N_p. Using MRI, we can acquire and segment the images to attribute to each point its tissue properties. Considering this as a classification problem with a defined precision, for example 0.1°C, we train a neural network model on a numerical brain slice with multiple points using their attributes (tissue properties, distance to four surface sensors and known temperatures acquired through simulation). We then test on two other slices and compare with the simulation values to estimate the accuracy.

Methods

An electromagnetic/thermal co-simulation was performed using CST (Dassault Systèmes, France). A 16-rung 3T birdcage head coil was modeled. A human numerical model (“TOM”, adult male from the CST Voxel Family) was imported into the head coil model with a 2 mm isotropic resolution (Figure 1 - Left). A time-domain electromagnetic simulation was performed with the coil tuned to 123 MHz. Then, the thermal losses were used to perform a 100 seconds-long thermal simulation. The temperature for three different axial slices of the brain, chosen at z=20 mm, z=23 mm and z=35 mm (to have the same tissue types but structural differences), was saved at 100 seconds (Figure 1 - Right). The slices were segmented in CST with five tissue types (Fat, Gray Matter, Bone, CerebroSpinal Fluid, Muscle) and imported in MATLAB to obtain the properties maps for each slice: conductivity, permittivity, density (Figure 2). The thermal maps were also imported in MATLAB for deep learning implementations. The neural network, implemented with TensorFlow², was trained using 6690 points from the brain slice that had the highest temperature range among the three slices we selected (37.1°C – 39.8°C), i.e. the z = 20 mm slice. The neural network had 3 layers, 2048 nodes, a ReLU (Rectified Linear Unit) activation function, a learning rate of 0.0001, 6000 epochs. The cost after the final epoch was 0.39. The input was a matrix with 35 features for each point: 33 tissue properties (3 properties for each of the 11 segments), 1 norm distance to one of the four sensors, 1 temperature value. Each point was represented 4 times to consider the distance and properties to the 4 sensors. Random Gaussian noise with standard deviation 0.05 was added to provide for data augmentation. The hyper-parameters were optimized to improve the accuracy while avoiding over-fitting the model. The estimated temperature maps for the test slices were plotted, compared to the simulated maps and the correlation between the prediction and the CST temperature was assessed.

Results & Discussion

Figure 3 shows the correlation plots between the simulated and predicted temperature (3a,3b) and the cost versus number of iterations (3c). Figure 4 shows the temperature maps obtained with CST and the maps predicted by NITE. An accuracy of 86% was obtained for the training by adjusting the neural network parameters and using a random dropout method. We observe a positive, linear correlation. However, given the structural differences between the training slice and the test slices, the linearity is not yet optimal: indeed, the further away from the training slice, the less accurate the prediction becomes. Figure 5 shows a comparison of time performance between NITE and CST: once trained, the network can generate the maps several orders of magnitude faster. We plan to continue adjusting our algorithm, e.g. by tuning more finely the random dropout, optimizing the time or adjusting the number of nodes, and afterwards testing on brain slices with even bigger structural differences. Alternatively, we plan to train a brain volume (TOM’s) and test it on other models such as DUKE. Then, an in-vitro study will be performed on a phantom with surface and internal temperature sensors to validate our method experimentally.

Conclusion

The NITE approach shows potential to improve MRI safety, when optimized and experimentally validated: this could be a first step towards a precision approach for every patient, to provide them with a safe and efficient scanning protocol.

Acknowledgements

No acknowledgement found.