digital sunday – Symposien
Symposien beleuchten die wissenschaftliche Fragestellung der Digitalisierung, während Workshops und Referate die praktischen, rechtlichen und technischen Vorträge thematisieren.
|Nairobi 3&4||12:00 - 13:15||30.09.2018|
|Deep learning in Ophthalmology - Technical approaches|
Deep learning represents a new technology, which will make decision support systems and automatic imaging analysis possible. Currently, the focus is on OCT imaging modalities. In this symposium, computer scientists will present their current projects in the field of Ophthalmology in a clinical context. The symposium shall provide a forum for exchange between different working groups.
Karsten Kortüm (München)
Pearse Keane (London)
Markus Rohm (München)
It could be shown in earlier work, that it is possible to predict the visual acuity in AMD patients with an root mean squared error of 0.14 logMAR using only classical machine learning approaches. In this talk, we will discuss how more accurate and robust predictions can created using deep learning approaches. Another aspect of the talk will be how to estimate the reliability of the model in the light of real world data, where it is possible that a productive system may recieve heavily differing examples (outliers).
Hrvoje Bogunovic (Wien)
Ophthalmology is ideally suited for machine and deep learning applications due to availability of fast and standardized high-resolution non-invasive imaging of the retina. In this presentation we will cover the application of deep learning for imaging biomarker segmentation in optical coherence tomography (OCT). In addition, we will present the potential of deep learning for individualized prognosis of early/intermediate AMD and treatment management of patients with wet-AMD patients.
Raphael Sznitman (Bern)
We investigate the use of a Deep Learning approach for automatic grading of OCT Bscans with a variety of potential biomarkers. Using only Bscan grading information to train our approach, we show that our method produces high-end performances, and allows localization of biomarkers in a fully automated way for free, whereby avoiding the tedious need to aggregate pixel-wise annotations of biomarkers.
|Applying Deep Learning methods to OCT image analysis in the context of Age-related macula degeneration||DS05-04|
Holger Langner (Mittweida)
In our project "Therapieprädiktion durch OCT und Patientendemographie in der Ophthalmologie" (TOPOs) we analyze a large data set of macular OCT images and textual data from patients who have been diagnosed with Age-related Macula Degeneration (AMD). Our work touches a broad spectrum of OCT image analysis problems, starting with automatic diagnosis classification, detecting the presence of AMD-related indicators and their changes over time, and finally predicting the individual long-term outcome of a VGEF-based therapy. Our talk intorduces our approach to apply Deep Learning methods to these tasks, and presents first results that shall convey a brief impression of what can be expected from Deep Learning methods in this domain.
Thomas Schultz (Bonn)
We propose an automated method for detection and quantification of Drusen from Optical Coherence Tomography (OCT). We segment Bruch´s membrane and the Retinal Pigment Epithelium (RPE) layer using a Convolutional Neural Network (CNN) and shortest path finding. Drusen are recognized as deviations from an idealized RPE, and false positives eliminated from an en face projection. Our approach takes less than a second per B-scan, and achieves higher accuracy than the previous state of the art.