A prerequisite of training a neural network is that all samples are consistent. For example, the MICCAI BraTS example described above expects four input MR images, each from a particular type of MR sequence, and reference segments identifying three tissue types for segmentation. Adding a sample with only two input MR images, or reference segments with five labels provokes a failure. Likewise, prediction using the trained neural network will fail with data of a different structure.
To ensure consistency, PAI uses a manifest file (JSON format) to store information about the training process and history. The consistency check includes the number and type of datasets included, their descriptions, as well as the pre-processing procedures. The manifest file is based on the first valid sample (“leading sample”) when the first training occurs.
The following checks are included;
•Proper loading of the associated images – checking whether all associated images are properly loaded.
•Number of dimensions – the number of dimensions of the input images must be consistent with the number of dimensions of the images used for the original training. It must also be consistent with the number of dimensions supported by the chosen model.
•Number of associated images – the number of associated input images must be the same as the number of images associated in the samples used for the original training.
•“Project” field, serving as image description – the description of the associated input images stored in the “Project” database field must be the same as in the images used for the original training
•Modalities – the modalities of the associated input images must be the same as for the images used for the original training.
If one of the above requirements is not met, PAI will:
•Prediction: Cancel processing and inform the user
•Re-training: Skip the sample and print the information in the log
Model Configuration Check
The final control checks the settings in the user interface. Every time the user starts a training or saves the learning set, the content of the selected manifest file will be compared to the current settings in the user interface. This avoids using training parameters different from previous training sessions.
The data and model consistency checks are performed in the background during the workflows. Issues will be automatically detected and the user guided to correct them.