viseg

The viseg package is our tool for the segmentation of MRI data.

In order to build the Virdx product we need to segment the anatomic structures of the prostate in the MRI images as well as outline potential lesion candidates. Currently, our segmentation pipeline is based on the nnUNet, which is developed by the DKFZ in Heidelberg. The nnUNet is a self-configuring neural network that has won numerous medical image segmentation challenges and is still a very hard baseline to beat. We have implemented our own version of the nnUNet, and benchmarked it against the original implementation to show that we obtain identical performance for 3D lesion and anatomy segmentation. Having our own, trimmed back version of nnUNet ensures that it is easier to implement new features and makes it easier to write tests for all functionality. Only few features have not been implemented as we did not considered those to be useful at this time. A full list is given below.

Key features and goals of viseg:

• Robust training pipeline for segmentation tasks using different input channels and target classes.
• Out of the box inference for Vicom images and Virdx Studies.
• Estimation of model uncertainty to enable active learning approaches and quantify the model uncertainty to increase the confidence in the predictions.
• Integration of viseg and clearml to log outputs from training runs, the trained models, and the datasets that were used to create the models to have the full reproducibility.