With a recent increase in deep learning and image processing applications in medical diagnosis, it is necessary to estimate the amount of data required to adequately train these algorithms. This is especially difficult when the ideal data does not exist in an accessible database.
We are designing a study to collect images to train a deep learning algorithm to detect severe burns on human skin. We will generate our own training database and estimate the sample size necessary for the algorithm to achieve an appropriate level of accuracy when validated on an independent dataset. One method for sample size estimation is a learning curve (Figueroa, et al. 2012, Equation 1), dependent on minimum achievable error, learning rate, and decay rate parameters.
Using bootstrap, we apply this learning curve to our current small dataset. This method estimates the algorithm would require a dataset of approximately 108 subjects to achieve 88% accuracy in detecting burned tissue. Limitations of included outliers, non-convergent parameter estimates, and unrepresentative validation sets. Despite the limitations, we found this method provides a practical estimate of training database sample size.
|