The consortium has been developing deep learning architectures and models to tackle the task of automatic segmentation of liver and lung lesions. As is the case with all machine learning architectures, there are many choices to be made, including input normalisation, variants of loss functions, and types of regularisation. We have been systematically varying these on a contrast-enhanced CT dataset and reporting on which have the most impact on the performance of the algorithm.
Also the consortium worked on fully automated image segmentation to remove the bottleneck of most radiomic analyses involving manual or semi-automated tools for the tumor detection and delineation step before its characterization. Finally the consortium made a contribution to the identification of robust and reliable features. ComBat has been modified to include bootstrap and Monte Carlo estimates (higher robustness) as well as the ability to choose a reference amongst the centers, to avoid arbitrary transforms leading to impossible values for some features. Results obtained in a multicentric cohort of patients with cervical cancer and multiple image modalities (FDG PET and MRI sequences) demonstrates the added value of the method. We have been developing a Radiomics platform for HCC based on contrast-enhanced CT. This has been motivated in part by the availability of a rich data source, and in part by the fact that at least half of the LIRADS guidelines are based on CT. U-Nets have been trained on data from the publicly available Medical Decathlon Challenge (MDC) dataset which consists of 131 annotated portal venous phase CT scans. The performance of our models will next be assessed for a multi-timestep CT dataset (TACE2). For the curation of the TACE2 dataset, we first converted the files from DICOM to NIfTI format, then identified the arterial and portal venous phase scans, and manually delineated liver lesions using ITKsnap and MITK tools. In order to a further improve segmentation accuracies for challenging lesion appearances, baseline U-Net architectures were extended with attention modules. A visualisation of the network’s attention coefficients suggested that attention mechanisms have the potential to suppress or accentuate regions during decision-making, thereby increasing the focus on informative image regions. Currently, transfer learning techniques are implemented to apply models (that were trained on single timestep CT data) to the prediction of multiphasic CT data. As the TACE2 dataset has not yet been annotated by experts, unsupervised and weakly supervised domain adaptation will be exploited to learn domain-independent cancer features. All ESRs participated to the workshop organized 12/2018 in Maastricht on AI for Imaging. Next to this the individual ESRs all participated in multiple workshops and conferences at their own institutions.
