A Deep-Learning Approach for Non-Invasive Temperature Measurements Using Ultrasound Images

This paper proposes a deep-learning temperature measurement method using ultrasound images that is based on the thermal dependance of local changes in the speed of sound. In this method, the temperature distribution is measured using a non-invasive image analysis technique. In a previous study, we found a temperature measurement accuracy of 1.0 °C or less. However, our previous method has some disadvantages. First, the image analysis parameters (e.g., the size of the template and the cross-correlation threshold) are empirically determined. Second, it is necessary to obtain the thermal constant ktissue according to the type of tissue and the analysis parameters. To overcome these problems, we propose a new method using deep-learning. This new method is divided into three steps. The first step is to determine the image analysis parameters from the ultrasound images using a convolutional neural network (CNN). The second step is to analyze the image using the estimated analysis parameters to obtain a normalized temperature distribution. The third step is to determine the thermal constant ktissue to calibrate the temperature increase using multi-layered perceptron (MLP). In this paper, first, we propose three types of image fusion methods to input the ultrasound images into the CNN. Comparing the results of the three methods, we determine the optimal CNN structure. Second, we determine the optimal MLP structure by changing the number of hidden layers and neurons. Finally, as described above, we obtain the temperature distribution. Our results indicate that the proposed deep-learning method can effectively provide non-invasive temperature measurements.