University of Tübingen

About: University of Tübingen is a education organization based out in Tübingen, Germany. It is known for research contribution in the topics: Population & Immune system. The organization has 40555 authors who have published 84108 publications receiving 3015320 citations. The organization is also known as: Eberhard Karls University & Eberhard-Karls-Universität Tübingen.

Leveraging uncertainty information from deep neural networks for disease detection.

Abstract: Deep learning (DL) has revolutionized the field of computer vision and image processing. In medical imaging, algorithmic solutions based on DL have been shown to achieve high performance on tasks that previously required medical experts. However, DL-based solutions for disease detection have been proposed without methods to quantify and control their uncertainty in a decision. In contrast, a physician knows whether she is uncertain about a case and will consult more experienced colleagues if needed. Here we evaluate drop-out based Bayesian uncertainty measures for DL in diagnosing diabetic retinopathy (DR) from fundus images and show that it captures uncertainty better than straightforward alternatives. Furthermore, we show that uncertainty informed decision referral can improve diagnostic performance. Experiments across different networks, tasks and datasets show robust generalization. Depending on network capacity and task/dataset difficulty, we surpass 85% sensitivity and 80% specificity as recommended by the NHS when referring 0-20% of the most uncertain decisions for further inspection. We analyse causes of uncertainty by relating intuitions from 2D visualizations to the high-dimensional image space. While uncertainty is sensitive to clinically relevant cases, sensitivity to unfamiliar data samples is task dependent, but can be rendered more robust.

Why ReLU Networks Yield High-Confidence Predictions Far Away From the Training Data and How to Mitigate the Problem

Abstract: Classifiers used in the wild, in particular for safety-critical systems, should not only have good generalization properties but also should know when they don't know, in particular make low confidence predictions far away from the training data. We show that ReLU type neural networks which yield a piecewise linear classifier function fail in this regard as they produce almost always high confidence predictions far away from the training data. For bounded domains like images we propose a new robust optimization technique similar to adversarial training which enforces low confidence predictions far away from the training data. We show that this technique is surprisingly effective in reducing the confidence of predictions far away from the training data while maintaining high confidence predictions and test error on the original classification task compared to standard training.

Association of the T-G Polymorphism in Adiponectin (Exon 2) With Obesity and Insulin Sensitivity: Interaction With Family History of Type 2 Diabetes

Abstract: The adipocyte-derived hormone adiponectin seems to protect from insulin resistance, a key factor in the pathogenesis of type 2 diabetes. Genome-wide scans have mapped a susceptibility locus for type 2 diabetes and the metabolic syndrome to chromosome 3q27, where the adiponectin gene is located. A common silent T-G exchange in nucleotide 94 (exon 2) of the adiponectin gene has been associated with increased circulating adiponectin levels. Metabolic abnormalities associated with the G allele have not been reported. We therefore assessed whether this polymorphism alters insulin sensitivity and/or measures of obesity using the Tubingen Family Study database (prevalence of the G allele, 28%). In 371 nondiabetic individuals, we found a significantly greater BMI in GG + GT (25.5 +/- 0.7 kg/m(2)) compared with TT (24.1 +/- 0.3 kg/m(2); P = 0.02). Insulin sensitivity (determined by euglycemic clamp, n = 209) was significantly lower in GG + GT (0.089 +/- 0.007 units) compared with TT (0.112 +/- 0.005 units; P = 0.02). This difference disappeared completely on adjustment for BMI. Because our population contains a relatively high proportion of first-degree relatives of patients with type 2 diabetes, we stratified by family history (FHD). Much to our surprise, the genotype differences in BMI and insulin sensitivity in the whole population were attributable entirely to differences in the subgroup without FHD, whereas in the subgroup with FHD, the G allele had absolutely no effect. Moreover, individuals without FHD had a significantly lower BMI than individuals with FHD (25.2 +/- 0.4 vs. 26.2 +/- 0.5 kg/m(2); P = 0.01), which was not the case for the GG + GT subgroup without FHD (27.0 +/- 0.9 kg/m(2); NS). This suggests that in individuals without familial predisposition for type 2 diabetes, the adiponectin polymorphism may mildly increase the obesity risk (and secondarily insulin resistance). In contrast, in individuals who are already burdened by other genetic factors, this small effect may be very hard to detect.

Cognitive functions correlate with white matter architecture in a normal pediatric population: A diffusion tensor MRI study

Abstract: A possible relationship between cognitive abilities and white matter structure as assessed by magnetic resonance diffusion tensor imaging (DTI) was investigated in the pediatric population. DTI was performed on 47 normal children ages 5-18. Using a voxelwise analysis technique, the fractional anisotropy (FA) and mean diffusivity (MD) were tested for significant correlations with Wechsler full-scale IQ scores, with subject age and gender used as covariates. Regions displaying significant positive correlations of IQ scores with FA were found bilaterally in white matter association areas, including frontal and occipito-parietal areas. No regions were found exhibiting correlations of IQ with MD except for one frontal area significantly overlapping a region containing a significant correlation with FA. The positive direction of the correlation with FA is the same as that found previously with age, and indicates a positive relationship between fiber organization and/or density with cognitive function. The results are consistent with the hypothesis that regionally specific increased fiber organization is a mechanism responsible for the normal development of white matter tracts.