Computer Vision News - August 2020

3D-RADNet 29 Best of MIDL 2020 structural information that gets lost with 2D networks and hasn’t really been explored yet. Another important characteristic of medical images lies in a set of information called digital imaging and communications in medicine metadata (DICOM) . These come together with the scans and contain various classifications of each image under specific fields. The method by Richard Du and Varut Vardhanbhuti proposes to take advantage of the information in the DICOMs to semi-automatically label a large public MRI and CT dataset available from The Cancer Imaging Archive (TCIA) and train a 3D Convolution neural network to classify the images. The chosen dataset covers different parts of the body, summarized in the figure below, but to demonstrate the effectiveness of transfer learning their proposed network is applied on a liver segmentation task. The figure below shows the parts of the body covered by the single scans, and, on the left, the DICOM labels selected for this study. From the TCIA online database (available here ) , the authors took the collections of MRI and CT scans totaling 17667 data for 4453 unique subjects. They excluded all the ones with less than 16 slices and linearly resized all of them to 48x192x192, normalized by min-max normalization and discretized to 256 grey levels. The labels extracted from the DICOMs are: • Presence of contrast agent, recorded in the field Contrast/Bolus Agent . • Type of scan view, found in the field Patient Image Orientation describing the • Modality, including CT and different MRI sequences. These can be identified by direction cosine of the first row and the first column with respect to the patient. the Scanning Sequence field , which distinguishes between the main types (spin-echo (SE), inversion recovery (IR) and gradient echo (GR)) and the fields Repetition time and Echo time , which define if the sequence is T1-weighted or T2-weighted.