Non-invasive modelling of ultrasound-induced temperature in tissues: a b-splines neural network solution

Abstract Efficient hyperthermia therapy session requires knowledge of the exact amount of heating needed at a particular tissue location and how it propagates around the area. Until now, ultrasound heating treatments are being monitored by Magnetic Resonance Imaging (MRI) which, besides raising the treatment instrumental cost, requires the presence of a team of clinicians and limits the hyperthermia ultrasound treatment area due to the space restrictions of an MRI examination procedure. This paper introduces a novel non-invasive modelling approach of ultrasound-induced temperature in tissue. This comes as a cost effective alternative to MRI techniques, capable of achieving a maximum temperature resolution of 0.26 °C, clearly inferior to the MRI gold standard resolution of 0.5 °C/cm 3 . Furthermore, we propose an innovative modelling methodology, where various similar models are built and are further combined through an optimization procedure, that we call neural ensemble optimization (NEO). This combination mechanism is shown to be superior to more simple schemes such as simple averages or evolutionary strategy based techniques.