Generalization assessment of non-invasive black-box temperature estimators from therapeutic ultrasound

Abstract

The objective of this work is the generalisation performance assessment, in terms of intensity, of non-invasive temperature models based on radial basis functions neural networks. The models were built considering data collected at three therapeutic ultrasound intensities, (among 0.5, 1.0, 1.5 and 2.0 W/cm2) and then were validated in fresh data, which contain information from the trained intensities and form the untrained intensity. The models were built to estimate the temperature evolution (during 35 min) in a gel-based phantom, heated by physiotherapeutic ultrasound at four different intensities. It was found that the best models built without data from the intermediate intensities (0.5, 1.0 and 1.5 W/cm2) perform well in validation at all the intensities. On the other hand, the models built without data from the extrapolated intensity (2,0 W/cm2) presented unsatisfactory results in validation. This is because the models parameters were found considering a space bounded by the data used in their construction, and then the application of data outside this space resulted in poor performance. The models build without the intermediate data, for the three considered points, presented a maximum absolute error inferior to 0.5 ºC (which is accepted for therapeutic applications). The best models also presented a low computational complexity, as desired for real-time applications.