Publication

Abstract : Deep convolutional neural networks were used to successfully segment several important neural tissue classes in MRI brain images, and approaches for integrating prior information into the networks to increase their performance on this task were investigated. Regrettably, only the first of them is addressed in this paper. To make the implementation of nonstandard architectures, which was expected to be required for the second goal, it was determined to provide a framework for defining and training networks by only using fundamental components. While this was an educational experience, the amount of progress accomplished was far less than if a conventional network package had been utilized instead. The requirement to deal with all of the lowest level aspects of network construction, from initialization schemes to adaptive learning rates and all the other components of the optimizer pipeline has left no time for utilizing this infrastructure to do something which has not been accomplished by the existing frameworks, and of course in all other respects it is far more limited than they are. It would in-stead be focused on a clear and detailed analysis of the full pipeline, which is required to build a network for solving the first problem. Despite several difficulties tracking down bugs in the optimizer, GPU memory allocation, and the last-minute accidental deletion of a large portion of the experimental results, the software implementation made available at: achieves DICE results of 0:8, which, while not class leading, would still place well in many benchmarks.