Noninvasive in situ temperature measurement (thermometry) is an attractive means of mapping the region of thermal damage during thermal ablation treatments. Existing methods of ultrasound thermometry are ineffective beyond 50 °C due to multiple physical limitations, including a non-monotonic relationship between temperature and sound speed that reaches a plateau around 60 °C, tissue phase transitions and deformation. An ultrasound-based technology that can monitor treatment over the entire therapeutic temperature range is desirable clinically.

 

This paper describes a deep learning-based approach that uses thermometry data from the periphery of the heating zone (where temperatures are less than 50 °C) to infer temperature throughout the treatment zone. Spatiotemporal 2D temperature maps from 3–12 s HIFU heating exposures (in 0.5 s increments) were generated (using COMSOL) with a subset used for training the network and the rest for testing. Peripheral temperature values (excluding the first 5 mm closest to the axial focus), scalar time values, and a binary flag indicating heating/cooling were inputs to the network, and the temperature profile axially through the HIFU focus was predicted.

 

The temperature prediction accuracy was better than 0.5 °C during heating and cooling. This paper will also address robustness to noise in the input temperature measurements and discuss future directions including experiments with ultrasound backscatter data and strategies to explicitly incorporate heat diffusion physics into the learning paradigm.