BEGIN:VCALENDAR VERSION:2.0 PRODID:researchseminars.org CALSCALE:GREGORIAN X-WR-CALNAME:researchseminars.org BEGIN:VEVENT SUMMARY:Samuli Siltanen (University of Helsinki) DTSTART:20210304T170000Z DTEND:20210304T180000Z DTSTAMP:20260423T035637Z UID:Inverse/36 DESCRIPTION:Title: Learning from electric X-ray images: the new EIT\nby Samuli Siltanen (University of Helsinki) as part of International Zoom Inverse Problems Se minar\, UC Irvine\n\n\nAbstract\nA fundamental connection between Electric al Impedance Tomography (EIT) and classical X-ray tomography was found in [Greenleaf et al 2018]. There it was shown that a one-dimensional Fourier transform applied to the spectral parameter of Complex Geometric Optics (C GO) solutions to a Beltrami equation is a useful technique. Microlocal ana lysis of the involved complex principal type operators reveals singulariti es propagating in curious ways. They enable a novel filtered back-projecti on type nonlinear reconstruction algorithm for EIT. This approach is calle d Virtual Hybrid Edge Detection (VHED). \n\nOne of the medically most prom ising applications of EIT is stroke imaging. There are two main types of s troke: (1) brain hemorrhage and (2) ischemic stroke caused by a blood clot . The symptoms for those two conditions are the same\, but the treatments are completely the opposite. There are two main uses for EIT here: (a) cla ssifying the type of stroke already in the ambulance with a cost-effective portable device\, and (b) monitoring the state of recovering stroke patie nts in the intensive care unit. \n\nThe main difficulty in using EIT for h ead imaging is the resistive skull. Because of that\, the relevant signal from the brain is weak and almost buried in noise. Given the extreme ill-p osedness of the inverse conductivity problem\, it is quite a challenge to design a robust EIT algorithm for either (a) or (b). \n\nVHED offers a way to divide the information in EIT measurements into geometrically understo od pieces. One could wish that those pieces are less sensitive to noise th an a full reconstructed image of the conductivity. This presentation shows how machine learning can be used for classifying stroke (problem (a)) abo ve based on VHED profiles. Examined are fully connected neural networks (F CNN)\, convolutional neural networks (CNN) and recurrent neural networks ( RNN). Perhaps surprisingly\, CNNs offer the worst performance\, while RNNs are slightly better than FCNNs.\n LOCATION:https://researchseminars.org/talk/Inverse/36/ END:VEVENT END:VCALENDAR