Computer Vision News - April 2020

2 Summary Research 4 by Marica Muffoletto Segmentation is one of the most common problems in medical imaging. Recently, with the newly renovated AI rush, Convolutional Neural Networks (CNNs) have become a standard approach for these tasks. Within the specific application of lung cancer, the most fatal type across population, old feature-design approaches such as Support Vector machines are not sufficient to define non-solid and sub-solid nodules due to their irregular shape and confused boundaries. The problem of most of these deep learning methods (including the 3D U-Net architectures with an optimal performance in both segmentation and malignancy prediction tasks) lies in the lack of explainability of the models and the impossibility to recreate the discrepancies in annotations among different studies and observers. Many solutions have been proposed, including a combination of conditional variational auto-encoder (cVAE) with a U-Net to produce a change in the sampled set of features that can output a plausible range of segmentations. Another one uses a pre-trained segmentation model with an expectation-maximization (EM) approach and obtains state-of-the-art results in organ segmentation. Unfortunately, it can’t perform equally well in lung nodules given the difficult nature of the task. This is instead the scope of the paper we are reviewing. iW-Net Every month, Computer Vision News reviews a research paper from our field. For April, we chose to talk about a medical imaging paper: iW-Net - an automatic and minimalistic interactive lung nodule segmentation deep network. We are indebted to the authors (Guilherme Aresta, Colin Jacobs, Teresa Araújo, António Cunha, Isabel Ramos, Bram van Ginneken and Aurélio Campilho), for allowing us to use their images to illustrate this review. You can find their paper at this link. "The problem of most of these deep learning methods lies in the lack of explainability of the models and the impossibility to recreate the discrepancies in annotations among different studies and observers." Introduction