Semantic Segmentation with Multi-Source Domain Adaptation for Radiological Images
Hugo Oliveira, Arnaldo de Albuquerque Araújo, Jefersson A. dos Santos · 2021
Differences in digitization equipment and techniques in radiology may hamper the use of data-driven deep learning approaches. In order to mitigate this limitation, in this work we merge generative image translation networks with supervised semantic segmentation architectures, yielding two semisupervised methods for domain adaptation in medical images. We compare our methods with traditional baselines in the literature using 3 image domains, 16 datasets and 8 segmentation tasks organized into three sets of experiments. Analysis of the results showed that the proposed methods for Domain Adaptation often reached Jaccard scores of 0.9 or higher in unsupervised or semi-supervised settings. We observe that unsupervised domain adaptation performance is close to the performance of fully supervised adaptation in most cases, bridging an important gap in the efficacy of neural networks between labeled and unlabeled datasets.