A Hybrid Model for Multimodal Brain Tumor Segmentation

We present a fully automatic segmentation method for multi-modal brain tumor segmentation. The proposed generative-discriminative hybrid model generates initial tissue probabilities, which are used subsequently for enhancing the classi�cation and spatial regularization. The model has been evaluated on the BRATS2013 training set, which includes multimodal MRI images from patients with high- and low-grade gliomas. Our method is capable of segmenting the image into healthy (GM, WM, CSF) and pathological tissue (necrotic, enhancing and non-enhancing tumor, edema). We achieved state-of-the-art performance (Dice mean values of 0.69 and 0.8 for tumor subcompartments and complete tumor respectively) within a reasonable timeframe (4 to 15 minutes).

Survey estimation and marginal effects in Stata

Stata is a general purpose software package that has become popular among various disciplines such as epidemiology, economics, or social sciences. Users like Stata for its scientific approach, its robustness and reliability, and the ease with which its functionality can be extended by user written programs. In this talk I will first give a brief overview of the functionality of Stata and then discuss two specific features: survey estimation and predictive margins/marginal effects. Most surveys are based on complex samples that contain multiple sampling stages, are stratified or clustered, and feature unequal selection probabilities. Standard estimators can produce misleading results in such samples unless the peculiarities of the sampling plan are taken into account. Stata offers survey statistics for complex samples for a wide variety of estimators and supports several variance estimation procedures such as linearization, jackknife, and balanced repeated replication (see Kreuter and Valliant, 2007, Stata Journal 7: 1-21). In the talk I will illustrate these features using applied examples and I will also show how user written commands can be adapted to support complex samples. Complex can also be the models we fit to our data, making it difficult to interpret them, especially in case of nonlinear or non-additive models (Mood, 2010, European Sociological Review 26: 67-82). Stata provides a number of highly useful commands to make results of such models accessible by computing and displaying predictive margins and marginal effects. In my talk I will discuss these commands provide various examples demonstrating their use.

Three essays on prices and selection

Competitive Market Segmentation Abstract In a two-firm model where each firm sells a high-quality and a low-quality version of a product, customers differ with respect to their brand preferences and their attitudes towards quality. We show that the standard result of quality-independent markups crucially depends on the assumption that the customers' valuation of quality is identical across firms. Once we relax this assumption, competition across qualities leads to second-degree price discrimination. We find that markups on low-quality products are higher if consuming a low-quality product involves a firm-specific disutility. Likewise, markups on high-quality products are higher if consuming a high-quality product creates a firm-specific surplus. Selection upon Wage Posting Abstract We discuss a model of a job market where firms announce salaries. Thereupon, they decide through the evaluation of a productivity test whether to hire applicants. Candidates for a job are locked in once they have applied at a given employer. Hence, such a market exhibits a specific form of the bargain-then-ripoff principle. With a single firm, the outcome is efficient. Under competition, what might be called "positive selection" leads to market failure. Thus our model provides a rationale for very small employment probabilities in some sectors. Exclusivity Clauses: Enhancing Competition, Raising Prices Abstract In a setting where retailers and suppliers compete for each other by offering binding contracts, exclusivity clauses serve as a competitive device. As a result of these clauses, firms addressed by contracts only accept the most favorable deal. Thus the contract-issuing parties have to squeeze their final customers and transfer the surplus within the vertical supply chain. We elaborate to what extent the resulting allocation depends on the sequence of play and discuss the implications of a ban on exclusivity clauses.

Sonographic patterns of lung consolidation in mechanically ventilated patients with and without ventilator-associated pneumonia: A prospective cohort study.

PURPOSE Thoracic ultrasound (TUS) has been successfully used in the diagnosis of community-acquired pneumonia. Little is known about its diagnostic potential in ventilator-associated pneumonia (VAP). The purpose of this study was to systematically describe the morphology and temporal changes of sonographic patterns in mechanically ventilated patients and to evaluate the diagnostic performance characteristics of TUS-based VAP diagnoses. MATERIALS AND METHODS Patients who were placed on invasive ventilation for reasons other than pneumonia and who were considered at risk for the development of VAP received daily TUS examinations while being closely monitored for the development of pneumonia. RESULTS Fifty-seven patients were studied. The incidence of VAP was 21.1%. Sonographic patterns of reduced or absent lung aeration were found in 64.2% of examinations. The sonographic pattern of lung consolidation with either dynamic or static air bronchograms was 100% sensitive and 60% specific for VAP in those patients who developed clinical signs and symptoms compatible with pneumonia. The pretest and posttest probabilities were 0.38 and 0.6, respectively. CONCLUSIONS Sonographic patterns of abnormal aeration are frequently observed in mechanically ventilated patients. If sonographic lung consolidation with either static or dynamic air bronchograms is absent, VAP is highly unlikely. The presence of these sonographic patterns in patients with signs and symptoms suggestive of pneumonia significantly increases the probability of VAP.

Good just isn't Good Enough: Humean Chances and Boltzmannian Statistical Physics

Statistical physicists assume a probability distribution over micro-states to explain thermodynamic behavior. The question of this paper is whether these probabilities are part of a best system and can thus be interpreted as Humean chances. I consider two Boltzmannian accounts of the Second Law, viz.\ a globalist and a localist one. In both cases, the probabilities fail to be chances because they have rivals that are roughly equally good. I conclude with the diagnosis that well-defined micro-probabilities under-estimate the robust character of explanations in statistical physics.

Voting Power and Probability

Voting power is commonly measured using a probability. But what kind of probability is this? Is it a degree of belief or an objective chance or some other sort of probability? The aim of this paper is to answer this question. The answer depends on the use to which a measure of voting power is put. Some objectivist interpretations of probabilities are appropriate when we employ such a measure for descriptive purposes. By contrast, when voting power is used to normatively assess voting rules, the probabilities are best understood as classical probabilities, which count possibilities. This is so because, from a normative stance, voting power is most plausibly taken to concern rights and thus possibilities. The classical interpretation also underwrites the use of the Bernoulli model upon which the Penrose/Banzhaf measure is based.

Deep molecular response is reached by the majority of patients treated with imatinib, predicts survival, and is achieved more quickly by optimized high-dose imatinib: results from the randomized CML-study IV

PURPOSE Deep molecular response (MR(4.5)) defines a subgroup of patients with chronic myeloid leukemia (CML) who may stay in unmaintained remission after treatment discontinuation. It is unclear how many patients achieve MR(4.5) under different treatment modalities and whether MR(4.5) predicts survival. PATIENTS AND METHODS Patients from the randomized CML-Study IV were analyzed for confirmed MR(4.5) which was defined as ≥ 4.5 log reduction of BCR-ABL on the international scale (IS) and determined by reverse transcriptase polymerase chain reaction in two consecutive analyses. Landmark analyses were performed to assess the impact of MR(4.5) on survival. RESULTS Of 1,551 randomly assigned patients, 1,524 were assessable. After a median observation time of 67.5 months, 5-year overall survival (OS) was 90%, 5-year progression-free-survival was 87.5%, and 8-year OS was 86%. The cumulative incidence of MR(4.5) after 9 years was 70% (median, 4.9 years); confirmed MR(4.5) was 54%. MR(4.5) was reached more quickly with optimized high-dose imatinib than with imatinib 400 mg/day (P = .016). Independent of treatment approach, confirmed MR(4.5) at 4 years predicted significantly higher survival probabilities than 0.1% to 1% IS, which corresponds to complete cytogenetic remission (8-year OS, 92% v 83%; P = .047). High-dose imatinib and early major molecular remission predicted MR(4.5). No patient with confirmed MR(4.5) has experienced progression. CONCLUSION MR(4.5) is a new molecular predictor of long-term outcome, is reached by a majority of patients treated with imatinib, and is achieved more quickly with optimized high-dose imatinib, which may provide an improved therapeutic basis for treatment discontinuation in CML.

Safety and efficacy of imatinib in CML over a period of 10 years: data from the randomized CML-study IV.

Tyrosine kinase inhibitors (TKI) have changed the natural course of chronic myeloid leukemia (CML). With the advent of second-generation TKI safety and efficacy issues have gained interest. The randomized CML - Study IV was used for a long-term evaluation of imatinib (IM). 1503 patients have received IM, 1379 IM monotherapy. After a median observation of 7.1 years, 965 patients (64%) still received IM. At 10 years, progression-free survival was 82%, overall survival 84%, 59% achieved MR(5), 72% MR(4.5), 81% MR(4), 89% major molecular remission and 92% MR(2) (molecular equivalent to complete cytogenetic remission). All response levels were reached faster with IM800 mg except MR(5). Eight-year probabilities of adverse drug reactions (ADR) were 76%, of grades 3-4 22%, of non-hematologic 73%, and of hematologic 28%. More ADR were observed with IM800 mg and IM400 mg plus interferon α (IFN). Most patients had their first ADR early with decreasing frequency later on. No new late toxicity was observed. ADR to IM are frequent, but mostly mild and manageable, also with IM 800 mg and IM 400 mg+IFN. The deep molecular response rates indicate that most patients are candidates for IM discontinuation. After 10 years, IM continues to be an excellent initial choice for most patients with CML.Leukemia advance online publication, 13 March 2015; doi:10.1038/leu.2015.36.

Tumor budding in colorectal cancer revisited: results of a multicenter interobserver study.

Tumor budding in colorectal cancer (CRC) is recognized as a valuable prognostic factor but its translation into daily histopathology practice has been delayed by lack of agreement on the optimal method of assessment. Within the context of the Swiss Association of Gastrointestinal Pathology (SAGIP), we performed a multicenter interobserver study on tumor budding, comparing hematoxylin and eosin (H&E) with pan-cytokeratin staining using a 10 high power field (10HPF) and hotspot (1HPF) method. Two serial sections of 50 TNM stage II-IV surgically treated CRC were stained for H&E and pan-cytokeratin. Tumor buds were scored by independent observers at six participating centers in Switzerland and Austria using the 10HPF and 1HPF method on a digital pathology platform. Pearson correlation (r) and intra-class correlation coefficients (ICC) comparing scores between centers were calculated. Three to four times more tumor buds were detected in pan-cytokeratin compared to H&E slides. Correlation coefficients for tumor budding counts between centers ranged from r = 0.46 to r = 0.91 for H&E and from r = 0.73 to r = 0.95 for pan-cytokeratin slides. Interobserver agreement across all centers was excellent for pan-cytokeratin [10HPF: ICC = 0.83 and 1HPF: ICC = 0.8]. In contrast, assessment of tumor budding on H&E slides reached only moderate agreement [10HPF: ICC = 0.58 and 1HPF: ICC = 0.49]. Based on previous literature and our findings, we recommend (1) pan-cytokeratin staining whenever possible, (2) 10HPF method for resection specimens, and (3) 1HPF method for limited material (preoperative biopsy or pT1). Since tumor budding counts can be used to determine probabilities of relevant outcomes and as such more optimally complement clinical decision making, we advocate the avoidance of cutoff scores.

Monitoring and Switching of First-line Antiretroviral Therapy in sub-Saharan Africa: Collaborative Analysis of Adult Treatment Cohorts.

BACKGROUND HIV-1 viral load (VL) testing is recommended to monitor antiretroviral therapy (ART) but not universally available. We examined monitoring of first-line and switching to second-line ART in sub-Saharan Africa, 2004-2013. METHODS Adult HIV-1 infected patients starting combination ART in 16 countries were included. Switching was defined as a change from a non-nucleoside reverse-transcriptase inhibitor (NNRTI)-based regimen to a protease inhibitor (PI)-based regimen, with a change of ≥1 NRTI. Virological and immunological failures were defined per World Health Organization criteria. We calculated cumulative probabilities of switching and hazard ratios with 95% confidence intervals (CI) comparing routine VL monitoring, targeted VL monitoring, CD4 cell monitoring and clinical monitoring, adjusted for programme and individual characteristics. FINDINGS Of 297,825 eligible patients, 10,352 patients (3·5%) switched during 782,412 person-years of follow-up. Compared to CD4 monitoring hazard ratios for switching were 3·15 (95% CI 2·92-3·40) for routine VL, 1·21 (1·13-1·30) for targeted VL and 0·49 (0·43-0·56) for clinical monitoring. Overall 58.0% of patients with confirmed virological and 19·3% of patients with confirmed immunological failure switched within 2 years. Among patients who switched the percentage with evidence of treatment failure based on a single CD4 or VL measurement ranged from 32·1% with clinical to 84.3% with targeted VL monitoring. Median CD4 counts at switching were 215 cells/µl under routine VL monitoring but lower with other monitoring (114-133 cells/µl). INTERPRETATION Overall few patients switched to second-line ART and switching occurred late in the absence of routine viral load monitoring. Switching was more common and occurred earlier with targeted or routine viral load testing.

Looking too Old? How First Impressions of Older-looking Candidates Contribute to Unfavorable Selection Decisions

Building on theories of impression formation based on faces, this research investigates the impact of job candidates’ facial age appearance on hiring as well as the underlying mechanism. In an experiment, participants decided whether to hire a fictitious candidate aged 50 years, 30 years or without age information. The candidate’s age was signaled either via chronological information (varied by date of birth) or via facial age appearance (varied by a photograph on the résumé). Findings showed that candidates with older-appearing faces—but not chronologically older candidates—triggered impressions of low health and fitness, compared to younger-appearing candidates. These impressions reduced perceptions of person-job fit, which lowered hiring probabilities for older-appearing candidates. These findings provide the first evidence that trait impressions from faces are a determinant of age discrimination in personnel selection. They call for an extension of current models of age discrimination by integrating the effects of face-based trait impressions, particularly with respect to health and fitness.

Saddlepoint Approximations to the Probability of Ruin in Finite Time for the Compound Poisson Risk Process Perturbed by Diffusion

A large deviations type approximation to the probability of ruin within a finite time for the compound Poisson risk process perturbed by diffusion is derived. This approximation is based on the saddlepoint method and generalizes the approximation for the non-perturbed risk process by Barndorff-Nielsen and Schmidli (Scand Actuar J 1995(2):169–186, 1995). An importance sampling approximation to this probability of ruin is also provided. Numerical illustrations assess the accuracy of the saddlepoint approximation using importance sampling as a benchmark. The relative deviations between saddlepoint approximation and importance sampling are very small, even for extremely small probabilities of ruin. The saddlepoint approximation is however substantially faster to compute.

Low Risk of Death, but Substantial Program Attrition, in Pediatric HIV Treatment Cohorts in Sub-Saharan Africa

BACKGROUND: To date, an estimated 10% of children eligible for antiretroviral treatment (ART) receive it, and the frequency of retention in programs is unknown. We evaluated the 2-year risks of death and loss to follow-up (LTFU) of children after ART initiation in a multicenter study in sub-Saharan Africa. METHODS: Pooled analysis of routine individual data from 16 participating clinics produced overall Kaplan-Meier estimates of the probabilities of death or LTFU after ART initiation. Risk factors analysis used Weibull regression, accounting for between-cohort heterogeneity. RESULTS: The median age of 2405 children at ART initiation was 4.9 years (12%, younger than 12 months), 52% were male, 70% had severe immunodeficiency, and 59% started ART with a nonnucleoside reverse transcriptase inhibitor. The 2-year risk of death after ART initiation was 6.9% (95% confidence interval [CI]: 5.9 to 8.1), independently associated with baseline severe anemia (adjusted hazard ratio [aHR]: 4.10 [CI: 2.36 to 7.13]), immunodeficiency (adjusted aHR: 2.95 [CI: 1.49 to 5.82]), and severe clinical status (adjusted aHR: 3.64 [CI: 1.95 to 6.81]); the 2-year risk of LTFU was 10.3% (CI: 8.9 to 11.9), higher in children with severe clinical status. CONCLUSIONS: Once on treatment, the 2-year risk of death is low but the LTFU risk is substantial. ART is still mainly initiated at advanced disease stage in African children, reinforcing the need for early HIV diagnosis, early initiation of ART, and procedures to increase program retention.

Long distance dispersal shaped patterns of human genetic diversity in Eurasia

Most previous attempts at reconstructing the past history of human populations did not explicitly take geography into account, or considered very simple scenarios of migration and ignored environmental information. However, it is likely that the Last Glacial Maximum (LGM) affected the demography and the range of many species, including our own. Moreover, long-distance dispersal (LDD) may have been an important component of human migrations, allowing fast colonization of new territories and preserving high levels of genetic diversity. Here, we use a high-quality microsatellite dataset genotyped in 22 populations to estimate the posterior probabilities of several scenarios for the settlement of the Old World by modern humans. We considered models ranging from a simple spatial expansion to others including LDD and a LGM-induced range contraction, as well as Neolithic demographic expansions. We find that scenarios with LDD are much better supported by data than models without LDD. Nevertheless, we show evidence that LDD events to empty habitats were strongly prevented during the settlement of Eurasia. This unexpected absence of LDD ahead of the colonization wave front could have been caused by an Allee effect, either due to intrinsic causes such as an inbreeding depression built during the expansion, or to extrinsic causes such as direct competition with archaic humans. Overall, our results suggest only a relatively limited effect of the LGM-contraction on current patterns of human diversity. This is in clear contrast with the major role of LDD migrations, which have potentially contributed to the intermingled genetic structure of Eurasian populations.