Real-Time Non Subtraction Thermometry Using Artificial Neural Networks

Manivannan Jayapalan¹

Thermal monitoring in focused ultrasound applications is crucial step where MR is most widely used as it provides better thermal monitoring capability than others. Regular PRF shift technique involves, some form of image subtraction using a baseline pre-treatment images. Subject motion and tissue deformation due to coagulation can severely distort these techniques. Self-referenced methods require a large area of tissue around the ablation for polynomial fitting and cant be used when tissue cooling is applied to sensitive structures. Here a new method of thermal monitoring using Radial Basis Function Neural Network (RBFNN) trained by orthogonal least square algorithm is proposed. This method eliminates the need for baseline subtraction and also tolerates subject motion to a great extent. A feed forward, radial basis neural network is used with 2 input, 1 output and a hidden layer where the number of units in that layer is obtained using orthogonal least square algorithm learning method. Gaussian function is used as kernel whose centers are obtained through network learning. 2-D surface co-ordinates of phase image in a selected ROI is used as inputs while its corresponding phase value are used as output to train the network. Then the network is tested, where, the phase values obtained from the network and the actual values are compared. It was observed that the network output matches very well with the actual values which clearly proves that the neural networks approximates the phase distribution function very well.