Robustness of machine learning based image analysis is one of the biggest challenges for bringing AI technology into clinical practice. Predictive performance can be impacted by changes in the data distribution due to scanner differences, data mismatch and population biases. In this project we develop methods for robust and reliable machine learning including automated quality control of image analysis tools.
EMMA – Ensembles of Multiple Models and Architectures
EMMA is our award winning approach for robust brain lesion segmentation which achieved the first place at the 2017 Multimodal Brain Tumor Segmentation Challenge (BraTS).
Automated Quality Control
When integrating computational tools such as automatic segmentation into clinical practice, it is of utmost importance to be able to assess the level of accuracy on new data, and in particular, to detect when an automatic method fails. However, this is difficult to achieve due to absence of ground truth. Segmentation accuracy on clinical data might be different from what is found through cross-validation because validation data is often used during incremental method development, which can lead to overfitting and unrealistic performance expectations.
Reverse Classification Accuracy is a framework that we developed to automatically predict the performance of segmentation algorithms. A illustrative overview of the framework is shown below, and more details can be found in our related publications, for example, on predicting segmentation accuracy in cardiac MRI.