Computer Vision News - January 2022

61 Yingjing Feng and we did not have the ground-truth atrial activities associated with each ECG. That’s why I looked into building “digital twins” for AF patients using computer models. These models were built on detailed atrial geometries made from patient meshes, with realistic electrophysiological parameters and computational rules.We generated a large and diverse set of AF episodes, caused by different AF mechanisms. We then trained machine learning classifiers on the synthetic dataset to distinguish the mechanismnon-invasively, and applied these classifiers on patient signals. With this, we successfully predicted a patient group that was more likely to benefit from the current ablation procedure. Extraction of spatiotemporal features without an image scanner The most important contribution we made was the development of spatiotemporal machine learning algorithms without using an image scanner, by exploiting common structures between ECG channels. One example is shown below, where we used a second-order blind source separation algorithm based on Belouchrani et al. to extract 10 atrial activities (shown as S 1 to S 10 , at the bottom right corner of the figure). The figure also shows 252-lead ECG signals recorded from a patient, colored by the contribution (S 1 -to-lead contribution) received from the principal atrial activity (S 1 ). It can be seen that signals in dark red, which received the highest contribution from S 1 , had a similar morphology with S 1 . The S 1 -to-lead contribution, which was computed from ECGs, demonstrated consistency in encoding the location of S 1 , regardless of the difference in geometries and electrical properties between patients. Therefore, it was used as a spatiotemporal feature for classification. With our scanner-free algorithms on imaging the atria, AF patients are no longer subject to any imaging process (MRI or CT) to obtain a non-invasive diagnosis of the underlying mechanisms. This means improved accessibility for both clinical and home uses. Acknowledgements: The research is supported by European Union Horizon 2020 Research and Innovation programme “Personalised In-silico Cardiology (PIC)” under theMarie Skłodowska-Curie grant agreement no. 764738 and GENCI computing resources, allocation A0080310517. Lead placement of 252- lead ECGs (in colors) Transmembrane voltage (mV) -80 30 2520ms 2540ms 2560ms 2580ms N 2600ms 2620ms 2640ms 2660ms (a) 2520ms 2540ms 2560ms 2580ms N 2600ms 2620ms 2640ms 2660ms (b) Transmembrane voltage (mV) -80 30 2000ms 2100ms 2200ms 2300ms 2400ms 2500ms 2600ms 2700ms (a) 2000ms 2100ms 2200ms 2300ms 2400ms 2500ms 2600ms 2700ms (b) Simulated AF with different mechanisms (prediction target) Regional heterogeneity and fiber assignment in our atrial bi-layer models Simulated ECGs (prediction input) Acetylcholine effects Focal source Reentry Endocardial layer Epicardial layer Apply classifiers on patient ECGs for pre-operative patient screening Train random forest classifiers from the synthetic dataset Feature 1 < value 1 ? Feature 2 < value 2 ? Feature 3 < value 3 ? Feature 4 < value 4 ? Feature 5 < value 5 ? Yes No Yes No Computer models (“digital twins”) Synthetic dataset Machine learning ECGs recorded from the front ECGs recorded from the back