Computer Vision News - April 2022

44 Congrats, Doctor! Computer-aided diagnostic systems powered by deep learning models need to not only perform wellonlimitedknowndatasetsbutalsogeneralize to unseen samples and be robust to challenges such as outliers and artefacts and threats like adversarial attacks. To this end our research aimed at developing methods that improve and thoroughly evaluate the robustness of machine learning models for medical diagnosis. Improve Model Robustness Training deep learning models on limited data can prevent models from generalizing to unseen samples. Data augmentation is an established way of combatting overfitting and improving model generalization. However, applying random affine transformations to training samples is limited to exploring the immediate vicinity of training samples [ see figure following ]. To solve this problem, we introduced a novel data augmentation technique that utilizes manifold-exploring affine geometric transformations that create samples that lie on the border of the manifolds between two-classes, maximizing the variance the network is exposed to during training [see fig below]. Our method improved model robustness against affine and projective transformations and increased model accuracy on fine-grained skin lesion and breast tumor classification. Finally, we proposed a metric based on geodesic distance that quantified the robustness of classifiers by measuring the distance of the augmented samples to the model decision boundaries. Magda (Magdalini) Paschali recently completed her PhD at the Technical University of Munich at the Chair for Computer AidedMedical Procedures. Her research during her PhD focused on improving and evaluating the robustness of deep neural networks (DNNs) for medical imaging applications. Magda continues her research as a Postdoctoral Scholar in Stanford University at the Computational Neuroimage Science Laboratory (CNSLAB), where she focuses onmachine learning models that can improve the understanding, diagnosis, and treatment of neuropsychiatric disorders. Congrats, Doctor Magda!