The Fréchet inception distance (FID) is a widely used metric to quantify differences between image sets, most commonly to compare real images with synthetically generated ones. In this paper, we examine the robustness of the FID metric at quantifying this similarity by applying it to sets of real-images from the same distribution. With the use of 3 benchmark datasets and 3 real world datasets, we evaluate FID scores between distinct sets of real images by partitioning a superset into subsets of balanced class instances, and computing FID across a range of set sizes for both in- and out of distribution data. Here we show that FID values decrease approximately linearly with the inverse of the number of images in each subset. We observe sub-linear gradients for real world images, compared to linear and super-linear for simpler benchmark datasets. In general we observe a clear step change in conditioning when transitioning from an under-determined problem, however this does not appear to correspond to a measurable difference in the FID value, its variance, or the Rank of the covariance matrix used to compute the FID. Extrapolating FID values to extremely large set sizes (10$^{9}$ and 10$^{10}$) of real images from the same superset yields appreciable FID values. This questions the effectiveness of FID at measuring the similarity of real and generated images for practical image set sizes in real world applications.
All individuals on planet earth are sensitive to the environment, but some more than others. These individual differences in sensitivity to environments are seen across many animal species including humans, and can influence personalities as well as vulnerability and resilience to mental disorders. Yet, little is known about the underlying brain mechanisms. Key genes that contribute to individual differences in environmental sensitivity are the serotonin transporter, dopamine D4 receptor and brain-derived neurotrophic factor genes. By synthesizing neurodevelopmental findings of these genetic factors, and discussing them through the lens of mechanisms related to sensitive periods, which are phases of heightened neuronal plasticity during which a certain network is being finetuned by experiences, we propose that these genetic factors delay but extend postnatal sensitive periods. This may explain why sensitive individuals show behavioral features that are characteristic of a young brain state at the level of sensory information processing, such as reduced filtering or blockade of irrelevant information, resulting in a sensory processing system that ‘keeps all options open’.
Objective.We propose a nonparametric figure of merit, the contrast equivalent distance CED, to measure contrast directly from clinical images.Approach.A relative brightness distanceδis calculated by making use of the order statistic of the pixel values. By multiplyingδwith the grey value rangeR, the mean brightness distance MBD is obtained. From the MBD, the CED and the distance-to-noise ratio DNR can be derived. The latter is the ratio of the MBD and a previously suggested nonparametric measureτfor the noise. Since the order statistic is independent of the spatial arrangement of the pixel values, the measures can be obtained directly from clinical images. We apply the new measures to mammography images of an anthropomorphic phantom and of a phantom with a step wedge as well as to CT images of a head phantom.Main results.For low-noise images of a step wedge, the MBD is equivalent to the conventional grey value distance. While this measure permits the evaluation of clinical images, it is sensitive to noise. Therefore, noise has to be quantified at the same time. When the ratioσ/τof the noise standard deviationσtoτis available, validity limits for the CED as a measure of contrast can be established. The new figures of merit can be calculated for entire images as well as on regions of interest (ROI) with an edge length not smaller than 32 px.Significance.The new figures of merit are suited to quantify the quality of clinical images without relying on the assumption of a linear, shift-invariant system. They can be used for any kind of greyscale image, provided the ratioσ/τcan be estimated. This will hopefully help to achieve the optimisation of image quality vs dose required by radioprotection laws.
AbstractAimsPatients with suspected non-ST-segment elevation acute coronary syndrome (NSTE-ACS) are routinely transferred to the emergency department (ED). A clinical risk score with point-of-care (POC) troponin measurement might enable ambulance paramedics to identify low-risk patients in whom ED evaluation is unnecessary. The aim was to assess safety and healthcare costs of a pre-hospital rule-out strategy using a POC troponin measurement in low-risk suspected NSTE-ACS patients.Methods and resultsThis investigator-initiated, randomized clinical trial was conducted in five ambulance regions in the Netherlands. Suspected NSTE-ACS patients with HEAR (History, ECG, Age, Risk factors) score ≤3 were randomized to pre-hospital rule-out with POC troponin measurement or direct transfer to the ED. The sample size calculation was based on the primary outcome of 30-day healthcare costs. Secondary outcome was safety, defined as 30-day major adverse cardiac events (MACE), consisting of ACS, unplanned revascularization or all-cause death. : A total of 863 participants were randomized. Healthcare costs were significantly lower in the pre-hospital strategy (€1349 ± €2051 vs. €1960 ± €1808) with a mean difference of €611 [95% confidence interval (CI): 353–869; P < 0.001]. In the total population, MACE were comparable between groups [3.9% (17/434) in pre-hospital strategy vs. 3.7% (16/429) in ED strategy; P = 0.89]. In the ruled-out ACS population, MACE were very low [0.5% (2/419) vs. 1.0% (4/417)], with a risk difference of −0.5% (95% CI −1.6%–0.7%; P = 0.41) in favour of the pre-hospital strategy.ConclusionPre-hospital rule-out of ACS with a POC troponin measurement in low-risk patients significantly reduces healthcare costs while incidence of MACE was low in both strategies.Trial registrationClinicaltrials.gov identifier NCT05466591 and International Clinical Trials Registry Platform id NTR 7346.
One method of assessing the image quality of a mammography unit is to estimate a contrast-detail-curve (CDC) that is obtained from images of a technical phantom. It has been proposed to estimate this CDC by using an end-to-end neural network (NN) which only needs one image to determine the CDC. That approach, however, has been developed on the basis of images of one single mammography unit. In this work, we train NNs on synthetic images of contrast-detail phantoms for mammography and test the so-trained NNs on images that are obtained from real mammography units. The goal of this paper is to demonstrate that such a deep learning approach is capable to generalize to predict CDCs for various real mammography units. Our experiments cover various manufacturers and the proposed approach is shown to work across different NN architectures and preprocessing methods which highlights its generalizability.
