To pretrain or not? A systematic analysis of the benefits of pretraining in diabetic retinopathy.

There is an increasing number of medical use cases where classification algorithms based on deep neural networks reach performance levels that are competitive with human medical experts. To alleviate the challenges of small dataset sizes, these systems often rely on pretraining. In this work, we aim to assess the broader implications of these approaches in order to better understand what type of pretraining works reliably (with respect to performance, robustness, learned representation etc.) in practice and what type of pretraining dataset is best suited to achieve good performance in small target dataset size scenarios. Considering diabetic retinopathy grading as an exemplary use case, we compare the impact of different training procedures including recently established self-supervised pretraining methods based on contrastive learning. To this end, we investigate different aspects such as quantitative performance, statistics of the learned feature representations, interpretability and robustness to image distortions. Our results indicate that models initialized from ImageNet pretraining report a significant increase in performance, generalization and robustness to image distortions. In particular, self-supervised models show further benefits to supervised models. Self-supervised models with initialization from ImageNet pretraining not only report higher performance, they also reduce overfitting to large lesions along with improvements in taking into account minute lesions indicative of the progression of the disease. Understanding the effects of pretraining in a broader sense that goes beyond simple performance comparisons is of crucial importance for the broader medical imaging community beyond the use case considered in this work.

Variation in CFHR3 determines susceptibility to meningococcal disease by controlling factor H concentrations.

Neisseria meningitidis protects itself from complement-mediated killing by binding complement factor H (FH). Previous studies associated susceptibility to meningococcal disease (MD) with variation in CFH, but the causal variants and underlying mechanism remained unknown. Here we attempted to define the association more accurately by sequencing the CFH-CFHR locus and imputing missing genotypes in previously obtained GWAS datasets of MD-affected individuals of European ancestry and matched controls. We identified a CFHR3 SNP that provides protection from MD (rs75703017, p value = 1.1 × 10-16) by decreasing the concentration of FH in the blood (p value = 1.4 × 10-11). We subsequently used dual-luciferase studies and CRISPR gene editing to establish that deletion of rs75703017 increased FH expression in hepatocyte by preventing promotor inhibition. Our data suggest that reduced concentrations of FH in the blood confer protection from MD; with reduced access to FH, N. meningitidis is less able to shield itself from complement-mediated killing.

Toward Scalable and Unified Example-Based Explanation and Outlier Detection.

When neural networks are employed for high-stakes decision-making, it is desirable that they provide explanations for their prediction in order for us to understand the features that have contributed to the decision. At the same time, it is important to flag potential outliers for in-depth verification by domain experts. In this work we propose to unify two differing aspects of explainability with outlier detection. We argue for a broader adoption of prototype-based student networks capable of providing an example-based explanation for their prediction and at the same time identify regions of similarity between the predicted sample and the examples. The examples are real prototypical cases sampled from the training set via a novel iterative prototype replacement algorithm. Furthermore, we propose to use the prototype similarity scores for identifying outliers. We compare performance in terms of the classification, explanation quality and outlier detection of our proposed network with baselines. We show that our prototype-based networks extending beyond similarity kernels deliver meaningful explanations and promising outlier detection results without compromising classification accuracy.

Resolving challenges in deep learning-based analyses of histopathological images using explanation methods.

Deep learning has recently gained popularity in digital pathology due to its high prediction quality. However, the medical domain requires explanation and insight for a better understanding beyond standard quantitative performance evaluation. Recently, many explanation methods have emerged. This work shows how heatmaps generated by these explanation methods allow to resolve common challenges encountered in deep learning-based digital histopathology analyses. We elaborate on biases which are typically inherent in histopathological image data. In the binary classification task of tumour tissue discrimination in publicly available haematoxylin-eosin-stained images of various tumour entities, we investigate three types of biases: (1) biases which affect the entire dataset, (2) biases which are by chance correlated with class labels and (3) sampling biases. While standard analyses focus on patch-level evaluation, we advocate pixel-wise heatmaps, which offer a more precise and versatile diagnostic instrument. This insight is shown to not only be helpful to detect but also to remove the effects of common hidden biases, which improves generalisation within and across datasets. For example, we could see a trend of improved area under the receiver operating characteristic (ROC) curve by 5% when reducing a labelling bias. Explanation techniques are thus demonstrated to be a helpful and highly relevant tool for the development and the deployment phases within the life cycle of real-world applications in digital pathology.

Memory snapshot dataset of a compromised host with malware using obfuscation evasion techniques.

This article presents a dataset for studying the detection of obfuscated malware in volatile computer memory. Several obfuscated reverse remote shells were generated using Metasploit-Framework, Hyperion, and PEScrambler tools. After compromising the host, Memory snapshots of a Windows 10 virtual machine were acquired using the open-source Rekall's WinPmem acquisition tool. The dataset is complemented by memory snapshots of uncompromised virtual machines. The data includes a reference for all running processes as well as a mapping for the designated malware running inside the memory. The datasets are available in the article, for advancing research towards the detection of obfuscated malware from volatile computer memory during a forensic analysis.

Identification of regulatory variants associated with genetic susceptibility to meningococcal disease.

Non-coding genetic variants play an important role in driving susceptibility to complex diseases but their characterization remains challenging. Here, we employed a novel approach to interrogate the genetic risk of such polymorphisms in a more systematic way by targeting specific regulatory regions relevant for the phenotype studied. We applied this method to meningococcal disease susceptibility, using the DNA binding pattern of RELA - a NF-kB subunit, master regulator of the response to infection - under bacterial stimuli in nasopharyngeal epithelial cells. We designed a custom panel to cover these RELA binding sites and used it for targeted sequencing in cases and controls. Variant calling and association analysis were performed followed by validation of candidate polymorphisms by genotyping in three independent cohorts. We identified two new polymorphisms, rs4823231 and rs11913168, showing signs of association with meningococcal disease susceptibility. In addition, using our genomic data as well as publicly available resources, we found evidences for these SNPs to have potential regulatory effects on ATXN10 and LIF genes respectively. The variants and related candidate genes are relevant for infectious diseases and may have important contribution for meningococcal disease pathology. Finally, we described a novel genetic association approach that could be applied to other phenotypes.

Unmasking Clever Hans predictors and assessing what machines really learn.

Current learning machines have successfully solved hard application problems, reaching high accuracy and displaying seemingly intelligent behavior. Here we apply recent techniques for explaining decisions of state-of-the-art learning machines and analyze various tasks from computer vision and arcade games. This showcases a spectrum of problem-solving behaviors ranging from naive and short-sighted, to well-informed and strategic. We observe that standard performance evaluation metrics can be oblivious to distinguishing these diverse problem solving behaviors. Furthermore, we propose our semi-automated Spectral Relevance Analysis that provides a practically effective way of characterizing and validating the behavior of nonlinear learning machines. This helps to assess whether a learned model indeed delivers reliably for the problem that it was conceived for. Furthermore, our work intends to add a voice of caution to the ongoing excitement about machine intelligence and pledges to evaluate and judge some of these recent successes in a more nuanced manner.

Evaluating the Visualization of What a Deep Neural Network Has Learned.

Deep neural networks (DNNs) have demonstrated impressive performance in complex machine learning tasks such as image classification or speech recognition. However, due to their multilayer nonlinear structure, they are not transparent, i.e., it is hard to grasp what makes them arrive at a particular classification or recognition decision, given a new unseen data sample. Recently, several approaches have been proposed enabling one to understand and interpret the reasoning embodied in a DNN for a single test image. These methods quantify the "importance" of individual pixels with respect to the classification decision and allow a visualization in terms of a heatmap in pixel/input space. While the usefulness of heatmaps can be judged subjectively by a human, an objective quality measure is missing. In this paper, we present a general methodology based on region perturbation for evaluating ordered collections of pixels such as heatmaps. We compare heatmaps computed by three different methods on the SUN397, ILSVRC2012, and MIT Places data sets. Our main result is that the recently proposed layer-wise relevance propagation algorithm qualitatively and quantitatively provides a better explanation of what made a DNN arrive at a particular classification decision than the sensitivity-based approach or the deconvolution method. We provide theoretical arguments to explain this result and discuss its practical implications. Finally, we investigate the use of heatmaps for unsupervised assessment of the neural network performance.Deep neural networks (DNNs) have demonstrated impressive performance in complex machine learning tasks such as image classification or speech recognition. However, due to their multilayer nonlinear structure, they are not transparent, i.e., it is hard to grasp what makes them arrive at a particular classification or recognition decision, given a new unseen data sample. Recently, several approaches have been proposed enabling one to understand and interpret the reasoning embodied in a DNN for a single test image. These methods quantify the "importance" of individual pixels with respect to the classification decision and allow a visualization in terms of a heatmap in pixel/input space. While the usefulness of heatmaps can be judged subjectively by a human, an objective quality measure is missing. In this paper, we present a general methodology based on region perturbation for evaluating ordered collections of pixels such as heatmaps. We compare heatmaps computed by three different methods on the SUN397, ILSVRC2012, and MIT Places data sets. Our main result is that the recently proposed layer-wise relevance propagation algorithm qualitatively and quantitatively provides a better explanation of what made a DNN arrive at a particular classification decision than the sensitivity-based approach or the deconvolution method. We provide theoretical arguments to explain this result and discuss its practical implications. Finally, we investigate the use of heatmaps for unsupervised assessment of the neural network performance.

On Pixel-Wise Explanations for Non-Linear Classifier Decisions by Layer-Wise Relevance Propagation.

Understanding and interpreting classification decisions of automated image classification systems is of high value in many applications, as it allows to verify the reasoning of the system and provides additional information to the human expert. Although machine learning methods are solving very successfully a plethora of tasks, they have in most cases the disadvantage of acting as a black box, not providing any information about what made them arrive at a particular decision. This work proposes a general solution to the problem of understanding classification decisions by pixel-wise decomposition of nonlinear classifiers. We introduce a methodology that allows to visualize the contributions of single pixels to predictions for kernel-based classifiers over Bag of Words features and for multilayered neural networks. These pixel contributions can be visualized as heatmaps and are provided to a human expert who can intuitively not only verify the validity of the classification decision, but also focus further analysis on regions of potential interest. We evaluate our method for classifiers trained on PASCAL VOC 2009 images, synthetic image data containing geometric shapes, the MNIST handwritten digits data set and for the pre-trained ImageNet model available as part of the Caffe open source package.

Extracting latent brain states--Towards true labels in cognitive neuroscience experiments.

Neuroscientific data is typically analyzed based on the behavioral response of the participant. However, the errors made may or may not be in line with the neural processing. In particular in experiments with time pressure or studies where the threshold of perception is measured, the error distribution deviates from uniformity due to the structure in the underlying experimental set-up. When we base our analysis on the behavioral labels as usually done, then we ignore this problem of systematic and structured (non-uniform) label noise and are likely to arrive at wrong conclusions in our data analysis. This paper contributes a remedy to this important scenario: we present a novel approach for a) measuring label noise and b) removing structured label noise. We demonstrate its usefulness for EEG data analysis using a standard d2 test for visual attention (N=20 participants).

Multiple kernel learning for brain-computer interfacing.

Combining information from different sources is a common way to improve classification accuracy in Brain-Computer Interfacing (BCI). For instance, in small sample settings it is useful to integrate data from other subjects or sessions in order to improve the estimation quality of the spatial filters or the classifier. Since data from different subjects may show large variability, it is crucial to weight the contributions according to importance. Many multi-subject learning algorithms determine the optimal weighting in a separate step by using heuristics, however, without ensuring that the selected weights are optimal with respect to classification. In this work we apply Multiple Kernel Learning (MKL) to this problem. MKL has been widely used for feature fusion in computer vision and allows to simultaneously learn the classifier and the optimal weighting. We compare the MKL method to two baseline approaches and investigate the reasons for performance improvement.

A new scoring system derived from base excess and platelet count at presentation predicts mortality in paediatric meningococcal sepsis.

INTRODUCTION: The aim of this study was to derive a novel prognostic score for mortality in paediatric meningococcal sepsis (MS) based on readily available laboratory markers. METHODS: A multicentre retrospective cohort study for the consortium set and a single centre retrospective study for replication set. The consortium set were 1,073 children (age 1 week to 17.9 years) referred over a 15-year period (1996 to 2011), who had an admission diagnosis of MS, referred to paediatric intensive care units (PICUs) in six different European centres. The consortium set was split into a development set and validation set to derive the score. The replication set were 134 children with MS (age 2 weeks to 16 years) referred over a 4-year period (2007 to 2011) to PICUs via the Children's Acute Transport Service (CATS), London. RESULTS: A total of 85/1,073 (7.9%) children in the consortium set died. A total of 16/134 (11.9%) children in the replication set died. Children dying in the consortium set had significantly lower base excess, C-reactive protein (CRP), platelet and white cell count, more deranged coagulation and higher lactate than survivors. Paediatric risk of mortality (PRISM) score, Glasgow meningococcal septicaemia prognosis score (GMSPS) and Rotterdam score were also higher. Using the consortium set, a new scoring system using base excess and platelet count at presentation, termed the BEP score, was mathematically developed and validated. BEP predicted mortality with high sensitivity and specificity scores (area under the curve (AUC) in the validation set=0.86 and in the replication set=0.96). In the validation set, BEP score performance (AUC=0.86, confidence interval (CI): 0.80 to 0.91) was better than GMSPS (AUC=0.77, CI: 0.68, 0.85), similar to Rotterdam (AUC=0.87, CI: 0.81 to 0.93) and not as good as PRISM (AUC=0.93, CI: 0.85 to 0.97). CONCLUSIONS: The BEP score, relying on only two variables that are quickly and objectively measurable and readily available at presentation, is highly sensitive and specific in predicting death from MS in childhood.

Insights from classifying visual concepts with multiple kernel learning.

Combining information from various image features has become a standard technique in concept recognition tasks. However, the optimal way of fusing the resulting kernel functions is usually unknown in practical applications. Multiple kernel learning (MKL) techniques allow to determine an optimal linear combination of such similarity matrices. Classical approaches to MKL promote sparse mixtures. Unfortunately, 1-norm regularized MKL variants are often observed to be outperformed by an unweighted sum kernel. The main contributions of this paper are the following: we apply a recently developed non-sparse MKL variant to state-of-the-art concept recognition tasks from the application domain of computer vision. We provide insights on benefits and limits of non-sparse MKL and compare it against its direct competitors, the sum-kernel SVM and sparse MKL. We report empirical results for the PASCAL VOC 2009 Classification and ImageCLEF2010 Photo Annotation challenge data sets. Data sets (kernel matrices) as well as further information are available at http://doc.ml.tu-berlin.de/image_mkl/(Accessed 2012 Jun 25).

Genome-wide association study identifies variants in the CFH region associated with host susceptibility to meningococcal disease.

Meningococcal disease is an infection caused by Neisseria meningitidis. Genetic factors contribute to host susceptibility and progression to disease, but the genes responsible for disease development are largely unknown. We report here a genome-wide association study for host susceptibility to meningococcal disease using 475 individuals with meningococcal disease (cases) and 4,703 population controls from the UK. We performed, in Western European and South European cohorts (consisting of 968 cases and 1,376 controls), two replication studies for the most significant SNPs. A cluster of complement factor SNPs replicated independently in both cohorts, including SNPs within complement factor H (CFH) (rs1065489 (p.936D

Protein C promoter polymorphisms associate with sepsis in children with systemic meningococcemia.

Meningococcal disease may present as sepsis, meningitis or a combination of both. Protein C (PC) is an important regulator of thrombin activity. Two polymorphisms in the promoter region of PC (C-1654T, A-1641G) have been shown to affect PC levels. In patients with meningococcal sepsis, low PC levels have been correlated with increased severity and poor outcome. We established a multicenter case-control study to determine whether PC promoter polymorphisms are associated with occurrence and outcome of meningococcal disease and sepsis. 288 previously healthy children with meningococcal infection from 97 pediatric hospitals in Germany, Switzerland, Italy, and Austria and 309 healthy controls were included in the study. A strong age-dependant effect was found. Patients younger than 1 year carried significantly more often the CG-CG genotype than healthy controls (28.6% vs. 17.8%, P = 0.04). Carriers of the CG allele showed a 3.43-fold increased odds ratio (OR) to develop sepsis (95% CI: 1.05-11.20; 85.7% vs. 63.6%, P = 0.036). The TA-TA genotype conferred a protective role for the development of sepsis (P = 0.017) with a Haldane OR of 0.09 (95% CI: 0.01-0.94). Systolic blood pressure values were significantly decreased in patients carrying the CG-CG genotype (70 vs. 86 mmHg, P = 0.005), and the need for adrenergic support significantly higher (70% vs. 26%, P = 0.018), resulting in an OR of 6.61 (95% CI: 1.28-34.14). These findings show that in young children PC promoter genotype is associated with susceptibility for meningococcal disease, the development of meningococcal sepsis, lower blood pressure, and need for adrenergic support.

A review of the genetics of essential hypertension.

PURPOSE OF REVIEW: Essential hypertension affects more than 20% of the adult population, and has a multifactorial origin arising from an interaction between susceptibility genes and environmental factors. Several strategies have been used to identify hypertension susceptibility genes. This review highlights recent efforts in genetic dissection of essential hypertension. RECENT FINDINGS: Recently, further chromosomal regions harboring blood pressure loci have emerged in genome-wide linkage studies. Findings from a new systematic two-dimensional genome scan are presented, as well as sex-specific loci linked to hypertension in inbred rodent models. Many case-control association studies have been carried out, but results so far have been equivocal. This review discusses some interesting studies combining linkage and association strategies using gene-gene interactions, and studies the use of haplotypes instead of SNPs. Two novel hypertension susceptibility genes are presented, and a short summary on new insights into genes of the renin-angiotensin and adrenergic systems is given. SUMMARY: To date, linkage and association studies have not been convincing. Genome-wide association studies may prove to be an effective approach to the problems posed by complex traits. Combined with candidate gene approaches, it is hoped this strategy will yield convincing evidence for genes associated with essential hypertension.

Polymorphisms in the interleukin-1 gene cluster in children and young adults with systemic meningococcemia.

BACKGROUND: An association has been described between mortality in children with meningococcal disease and functional polymorphisms in the interleukin-1 (IL1) cluster. We undertook a multicenter study to evaluate associations of these polymorphisms in a Central European population. PATIENTS AND METHODS: The study involved 95 Middle European pediatric hospitals. We collected blood samples from, and clinical information about, 285 previously healthy children with meningococcal infection. We used a newly developed multiplexed mutagenic separated PCR assay to analyze 6 polymorphisms within the IL1 cluster: IL1A (-889)C/T, IL1A (+4845)G/T, IL1B (-511)C/T, IL1B (-31)C/T, IL1B (+3954), and IL1RA (+2018)C/T. We studied the same polymorphisms in a comparison group of 481 healthy newborns. RESULTS: Genotype frequencies between patients and the comparison group differed significantly only for the IL1RA (+2018)C/T variant: The CC genotype was more frequent in patients (11%) than in healthy controls (5%; P = 0.008). In the patient group, the C allele was significantly more prevalent (67%) in nonsurvivors than in survivors (42%; P = 0.02). CONCLUSION: The IL1RA (+2018)C/T polymorphism is associated with the risk of meningococcal disease and with its outcome.

Haplotypes of the beta-2 adrenergic receptor associate with high diastolic blood pressure in the Caerphilly prospective study.

OBJECTIVES: Current evidence demonstrates that both genetic and environmental factors influence blood pressure. The sympathetic nervous system is a key player in blood pressure control and functional genetic variants of the beta-2 adrenergic receptor (B2AR) have been identified and implicated in the pathogenesis of hypertension. The present study aimed to determine the effects of common haplotypes of the B2AR gene upon blood pressure in the Caerphilly Prospective Study. DESIGN: Two thousand five hundred and twelve men (aged 45-59 years) participated in the study. We selected individuals in the upper (n = 347) and lower (n = 279) quintiles of the diastolic blood pressure distribution fixed at two time points [phase 2 (1984-88) or phase 3 (1989-93)] as cases and controls. METHODS: We analysed two functional polymorphisms (Arg16Gly and Gln27Glu) of B2AR and their haplotypes. RESULTS: We found a higher risk of hypertension in individuals homozygous for the Gln27 compared to those individuals homozygous for Glu27 [odds ratio (OR) = 1.94; 95% confidence interval (CI) = 1.34-2.81; P = 0.001]. Three haplotypes (Gly16Gln27, Gly16Glu27 and Arg16Gln27) were present in both quintile groups. Logistic regression analysis showed that haplotypes with a Gln27 allele (Gly16Gln27 and Arg16Gln27) conferred a significantly higher risk for hypertension than the Gly16Glu27 haplotype (OR = 1.55; 95% CI = 1.11-2.17, OR = 1.37; 95% CI = 1.04-1.81; P = 0.009 and P = 0.027, respectively). However, there was no evidence to support a statistically significant difference in odds ratios for the Gly16Gln27 and Arg16Gln27 haplotypes (P = 0.477), suggesting that it is the Gln27 allele alone, rather than any haplotype, which best explains the association. CONCLUSIONS: In a prospectively studied Caucasian male cohort, high diastolic blood pressure was associated with B2AR haplotypes containing the pro-downregulatory Gln27 variant.

4G/5G promoter polymorphism in the plasminogen-activator-inhibitor-1 gene in children with systemic meningococcaemia.

UNLABELLED: Meningococcal disease may present as sepsis, meningitis or a combination of both. Impaired fibrinolysis and massive elevation of the plasminogen activator inhibitor-1 (PAI-1) is a characteristic feature of meningococcal sepsis. Previously, an association between mortality and the functional 4G/5G promoter polymorphism of the PAI-1gene in a cohort of UK and Dutch children with meningococcal sepsis was reported. We carried out a prospective, multicentre study to investigate the association of the 4G/5G PAI-1 polymorphism, diagnosis, and outcome in meningococcal disease in a Central European and UK population. Blood samples and clinical information of 347 previously healthy children with meningococcal infection were collected from 95 paediatric hospitals in Germany, Switzerland, Italy, the United Kingdom, and Austria from 2000 until 2002. Mortality was significantly associated with the 4G/4G genotype (12 of 90 (13%) vs. 15 of 240 (6%), P = 0.037), resulting in an odds ratio of 2.31. The diagnosis of sepsis (independent of symptoms of meningitis) was significantly more frequent in carriers of the 4G/4G genotype (P = 0.01), resulting in an odds ratio of 2.21 to develop sepsis. Meningitis was not associated with the PAI-1 4G/5G polymorphism, and allele frequencies were similar in patient and control groups. CONCLUSION: Our data show a correlation between the 4G/4G genotype in the plasminogen activator inhibitor-1 gene and poor outcome in children with meningococcal infection. In addition, 4G homozygous patients were prone to develop sepsis. We found no influence of the plasminogen activator inhibitor-1 polymorphism on the susceptibility to invasive meningococcal infection.

Clinical relevance of CD95 (Fas/Apo-1) on T cells of patients with B-cell chronic lymphocytic leukemia.

OBJECTIVE: Apoptosis mediated via CD95 (Fas/Apo-1) is a key regulator for the biology of normal and malignant lymphocytes. Although the function of CD95 on B-cell chronic lymphocytic leukemia cells (B-CLL cells) has been studied intensively, the clinical importance of CD95 expression on normal T cells in B-CLL has not been clarified. This study aimed to investigate whether expression of CD95 on peripheral blood T cells correlates with clinically relevant parameters of B-CLL disease. MATERIALS AND METHODS: Expression of CD95 (Fas/Apo-1) on peripheral blood T lymphocytes of patients with B-CLL was determined using flow cytometry and was correlated with expression of activation markers, sensitivity to apoptosis by anti-CD95, and clinical data, such as blood count, Binet stage, therapy, progression-free probability, and survival probability. RESULTS: Differential CD95 expression did not correlate with activation markers or with levels of apoptosis through anti-CD95. However, high levels of CD95 on T cells from B-CLL patients correlated significantly with low lymphocyte doubling time, increased Binet stages, and requirement for chemotherapeutic treatment. Furthermore, increased cell-surface CD95 on T cells was associated with reduced progression-free probability and poorer survival. CONCLUSIONS: CD95 levels on T cells correlate with the clinical course of B-CLL. Prospective studies appear warranted to investigate whether CD95 on T cells has a direct influence on B-CLL disease progression.