Adversarial attacks and the development of (deep) neural networks robust against them are currently two widely researched topics. The robustness of Learning Vector Quantization (LVQ) models against adversarial attacks has however not yet been studied to the same extent. We therefore present an extensive evaluation of three LVQ models: Generalized LVQ, Generalized Matrix LVQ and Generalized Tangent LVQ. The evaluation suggests that both Generalized LVQ and Generalized Tangent LVQ have a high base robustness, on par with the current state-of-the-art in robust neural network methods. In contrast to this, Generalized Matrix LVQ shows a high susceptibility to adversarial attacks, scoring consistently behind all other models. Additionally, our numerical evaluation indicates that increasing the number of prototypes per class improves the robustness of the models.

No abstract

Cardiovascular diseases account for 17 million deaths per year worldwide. Of these, 25% are categorized as sudden cardiac death, which can be related to ventricular tachycardia (VT). This kind of arrhythmia can be caused by focal activation sources outside the sinus node. A curative treatment is catheter ablation of these foci in order to inactivate the abnormal triggering activity. However, the localization procedure is usually time-consuming and requires an invasive procedure in the catheter lab. To facilitate and expedite the treatment, we present two novel localization support techniques based on convolutional neural networks (CNNs) addressing these clinical needs. In contrast to existing methods, our approaches were designed to be independent of the patient-specific geometry and directly applicable to surface ECG signals, while also delivering a binary transmural position. Furthermore, one of the method's outputs could be interpreted as several ranked solutions. The CNNs were trained on a data set containing only simulated data and evaluated both on simulated and clinical test data. On simulated data, the median test error was below 3 mm. The median localization error on the unseen clinical (test) data was between 32 mm and 41 mm without optimizing the pre-processing and CNN on the clinical (test) data. Interpreting the output of one of the approaches as ranked solutions, the best median error of the top-3 solutions dropped to 20 mm on the clinical test set. The transmural position was correctly detected in up to 82% of all clinical cases. These results demonstrate a proof of principle to utilize CNNs to localize the activation source without the intrinsic need of patient-specific geometrical information. Furthermore, delivering multiple solutions can help the physician to find the real activation source amongst more than one possible locations. With a further optimization to clinical data, these methods have a high potential to speed up clinical interventions, replace certain steps within those and therefore consequently decrease procedural risk and improve VT patients' outcomes.