In this dissertation, a new reconstruction algorithm to estimate the complete temperature field during hyperthermia is developed which relies upon a limited amount of transient measured temperature data. The predictive capabilities of this new algorithm are then systematically studied; first using one-dimensional simulated treatments, then using three-dimensional simulated treatments, and finally applying it to hyperthermia treatments of normal canine thighs. It was found that this new algorithm predicts the complete temperature fields more accurately and robustly than the steady-state approach. In particular, it can better predict the complete temperature fields in situations where the number of unknown blood perfusion parameters are greater than the number of available temperature sensors. It was also found that the steady-state temperature field could be estimated to within 1°C if there was no measurement noise, no model mismatch, and as few as three measurement locations for seven perfusion zones. The addition of measurement noise degraded the performance of this estimation algorithm especially when the number of measurement locations was small. It was found that use of Tikhonov regularization of order zero significantly improved the performance of the algorithm and that there was an optimal choice for the regularization parameter. For the animal experiments, normal canine thighs were instrumented with one-hundred twelve thermocouples and heated to steady-state using a 6 cm planar ultrasound transducer operating at 0.5 MHz: then the power was turned off and the transient cool down temperature data was stored for later use by the reconstruction algorithm. Only a subset of the one-hundred twelve measurements was used as input to the reconstruction algorithm. The remaining measurements were used to compare the results of the reconstruction algorithm with the true temperatures. The results showed that in general the predicted perfusion and reconstructed temperature field did not change significantly as sensors were removed. However, the error was quite large for some of the situations studied particularly when only twenty-seven piecewise constant regions of perfusion were used. Increasing the number of perfusion regions reduced this error suggesting that model mismatch had contributed significantly to the error.