Comparison of the internal noise models for channelized Hotelling observer

The channelized Hotelling observer (CHO) has become a widely used approach for evaluating medical image quality, acting as a surrogate for human observers in early-stage research on assessment and optimization of imaging devices and algorithms. Its popularity stems from experiments showing that the CHOs detection performance can be tuned to correlate well with human observers detection performance. Typically, this tuning is achieved introducing an internal-noise model tuned on example data obtained from human observers. Thus, it can be argued that internal-noise model tuning step is essentially a model training exercise; therefore, just as in supervised learning, it is essential to test the CHO model on a set of data that is distinct from that used to tune the model. Furthermore, the test data should be significantly different from the training data if the CHO is to provide useful insight about new imaging algorithms or devices. Motivated by these considerations, in this work a train-test approach was proposed, with new models selection criterions, and used to evaluate ten established internal noise models utilizing four (here presenting only one) different channel models. I also propose a new internal noise model (although not the main intention of this work) which I, incidentally, find outperforms the ten established models. The results show that the proposed train-test approach is necessary, with the newly proposed model selection criterions, for avoiding spurious conclusions. It was also demonstrate that, in some CHO models, the optimal internal-noise parameter is very sensitive to the choice of training data; therefore, these models are prone to overfitting, and will not likely generalize well to new data.