Deep learning for automated image analysis in clinical drug development trials

Purpose and goal

Antaros Medical is pioneering imaging methodologies to design and deliver clinical imaging studies for improved, evidence-based decision making and differentiation in drug development. The goal of the Vinnova project ´Start your AI journey´ was to streamline certain parts of the image analysis of MR images on liver for medical trials of non-alcoholic fatty liver disease (NAFLD/NASH). The project has produced very promising results on our test data and has reached the main goals of efficient and less operator-dependent image analysis of liver.

Expected results and effects

The project demonstrates that automatic segmentation of liver has great potential for producing efficient and more operator-independent measurements. Evaluation on three datasets yields high similarity with manual reference segmentations and shows that reproducibility is improved compared to a human operator. Further validation will be made before the methods can be applied in operational activities. The project has resulted in new hires and produced new ideas on how image analysis can be improved with our AI methods.

Planned approach and implementation

Antaros Medical has utilized expertise from Uppsala University. The deep learning architecture that was applied in the project was a U-net convolutional neural network (U-net CNN). The network was trained on data of manual liver segmentations from two previous NASH trials, with images from two different MR scanners. Extensive experiments were made before two final networks were established: One network for segmentation for liver volume measurements and one network for segmentation in liver fat fraction images. Evaluation has been made on three test datasets.