A neural network based estimation of tumour parameters from a breast thermogram

This paper proposes the estimation of the position and size of a spherical tumour in a human breast using the temperatures obtained on the surface of the breast through a breast thermogram in conjunction with artificial neural networks. The first part of the work involves the numerical simulation of heat conduction in a cancerous breast by employing the Pennes bio-heat transfer equation using a finite element based commercial solver COMSOL. The surface temperatures thus obtained from these simulations are trained against the tumour parameters by using artificial neural networks (ANN). An optimal neural network can then be utilized to retrieve parameters with good accuracy. The forward heat transfer problem was solved for two cases: (i) constant heat generation rate inside the tumour and (ii) variable heat generation rate inside the tumour, thereby generating 447 temperature data vectors in each case. In this study, the “measured” temperatures correspond to those numerically generated for a known size and position of the tumour. For the case of no noise in the “measured” data, accuracies of 90% in position and 95% in the radius for the constant heat generation rate were estimated using ANN, while another network retrieved the same with accuracies of 88% and 98% respectively for the case of varying heat generation rate. In practice, depending upon the resolution of an IR camera, the number of data points at which temperatures are recorded can be conveniently chosen from a breast thermogram. Hence, sensitivity studies were conducted to investigate the effect of the size of the data vector (“measured” temperatures) on the retrievals, and it is seen that an optimum indeed exists.