Computer Vision News - September 2020

Adversarial Data Augmentation ... 9 Two experiments were performed in this work. The first supervised and semi- supervised scenarios were performed using only 1 or 3 labelled subjects, splitting the dataset into 4 subsets: a labelled set, an unlabelled training stet, a validation and a test set. The network was compared with different data augmentation approaches by optimising their hyper-parameters if so needed. The network proposed outperforms the other methods when used in a supervised scenario . For the rest of the scenarios, the network achieves competitive results although it performs lesser in the case of semi-supervised learning , in comparison to the cGAN approach for 1 labelled subject. The second experiment was focused on learning using a limited population scenario . Normal healthy subjects used for training and the evaluation included pathological cases, which is a very typical scenario in healthcare. The results showed that the performance of the conventional method dropped in pathological cases while the proposed method outperformed the others, with improvements across five different populations. Fig. 3. shows that the variety of image styles is increased by encapsulating the shape information . Different styles may aid the network in learning high-level shape-based representation instead of texture- based. This could lead to improved network robustness on new, unknown classes. Figure 3. Visualisation of the generated GANs (including before and after bias field attack). This is a very promising work and as discussed previously, it offers interesting results for the usual healthcare scenario case of data scarcity. Feel free to read the whole paper for a deeper understanding and to get more details. Stay safe until next time! I would like to thank the authors of this paper Chen Chen and Daniel Rueckert for their permission to present their paper and to use the figures and data presented!