Computer Vision News - August 2020

Research 32 Best of MIDL 2020 Conclusion This paper shows promising results in the field of transfer learning; significantly higher performance is found on the same task compared to training from scratch in 3D networks. The pipeline proposed, which includes taking advantage of labels from DICOM metadata, could solve the aforementioned issues of generalizable networks and volumetric data in medical imaging. Further improvements could be applied to the architecture and to the training process, but overall 3DRADNet represents one of the few attempts at creating a backbone network for transfer learning specifically for medical images, which is now largely desired, allowing to extract features from large-scale unlabeled DICOM datasets.