Computer Vision News - August 2020

Research 28 Best of MIDL 2020 Every month, Computer Vision News reviews a research paper from our field. This too comes from MIDL: “3D-RADNet: Extracting labels from DICOM metadata for training general medical domain deep 3D convolutional neural networks”. We are indebted to the authors Richard Du and Varut Vardhanbhuti for allowing us to use their images to illustrate this review. You can find their paper here and the presentation held at MIDL 2020 at the video at the end of the article. 3 D - R A D N e t by Marica Muffoletto Introduction Although Deep learning seems like the perfect solution for many tasks right now, it still suffers from a well-known and worth exploring limitation: DL networks are usually strongly data-dependent, they are not generalizable and they require a vast specific dataset for tasks similar to the ones that might be already solved. For these reasons, transfer learning aims at transferring the knowledge, developed by a pre-trained network architecture on an available big dataset, to another task by fine-tuning on different data. This is particularly relevant in the field of biomedical engineering,wheremedicaldata isoftenhardtocollectandtoannotate.Unfortunately, the majority of available networks are also trained on 2D slices to avoid over-fitting on too many parameters, whereas most of medical data are volumetric, offering useful