22455 research outputs found
Sort by
Die Erkennung von Zahnpathologien auf apikalen Röntgenaufnahmen
Get PDFThe goal of the presented study was to illustrate the inter- and intra-examiner reliability of pathologies on apical radiographs and compare them with similar studies. Specifically, the differences between diagnostic categories, evaluation rounds and the professional experience of the examiner were studied to check for statistical relevance. For the detection of pathologies, variating Cohens Kappa (CK) values were recorded in the moderate to perfect range, indicating a significant difference in the assessment of the diagnostic categories and the experience of the examiner. In contrast, CK values were consistently in the perfect range regardless of the professional experience of the examiner and categories for the detection of restorative measures on periapical radiographs, as evidenced in the study by Wülk et al. (2023). Endo-perio-lesions had the best CK values of 0,840 CK and were thus rated most reliable, whereas periapical lesions and periodontal bone loss scored the least with CK values of 0,435 and 0,454 respectively. Additionally, the odds ratios (OR) indicate a large disparity between the experienced and less experienced examiners, while showing almost identical values across the evaluation rounds. The hypothesis that there is no difference in the reliability of assessment between diagnostic categories and the experience of examiners was therefore rejected in this study, while it was confirmed in the research by Wülk et al. (2023). A main driver for this discrepancy could be that pathologies are generally more difficult to assess in radiographs requiring a higher level of experience in the examiner than restorative measures. Differences between evaluation rounds were not detected in both studies, ensuring the replicability of the studies. The results suggest that enhanced university teaching is recommended to improve education and training programs for the detection of pathologies on apical radiographs, especially for periapical lesions and periodontal bone loss. Additional research on the quality of detection for additional diagnostic categories on apical radiographs, OPGs and bitewings is recommended to aid the identification and elimination of further uncertainties. The additional research should also address all potential weaknesses in radiological diagnostics to support automated and standardized results. With the rapid advancements in artificial intelligence (AI) use-cases in dentistry, the generation of high-quality data for AI models must be ensured
Reducing the effort for data annotation
Get PDFThe rise and success of modern supervised machine and deep learning models, which have become parts of our everyday lives, are partially fueled by the increasing availability of large datasets with high-quality annotations.
However, the availability of such annotations, also referred to as labels, remains a critical bottleneck for many machine learning applications as they are a prerequisite for supervised model training.
This is particularly evident in domains where the data annotation process is ambiguous and cumbersome or where it requires the knowledge of scarce domain experts.
In application domains such as medical imaging or industrial manufacturing, this often leads to the situation where, despite the availability of a large amount of non-annotated data, only a fraction of this data is annotated with appropriate labels.
This problem of model training with limited labeled data is the focus of this thesis which covers methods to leverage unlabeled data, samples without annotation, and weakly labeled data, samples with a low-information annotation, for model training.
Concretely, it includes contributions to the areas of semi-supervised learning, positive unlabeled learning, constrained clustering, and transfer learning.
First, this thesis introduces the concept of deep semi-supervised learning and provides an overview of recent research on self-training, entropy regularization, consistency regularization, and hybrid approaches.
The goal of semi-supervised learning is to train machine learning models on a small dataset of annotated training data while simultaneously using a larger dataset of completely unlabeled data.
Since the main developments in this area are driven by the computer vision community, many of these methods have been developed mainly for image data.
This motivated one contribution to investigate their application in a time series classification scenario.
Another contribution investigates the applicability of semi-supervised learning in a medical imaging context to reduce the data annotation effort in this domain.
Positive unlabeled learning is another exciting sub-field of low-supervised learning.
Here, the training data contains only positive or unlabeled samples, while the goal is to learn a binary classifier that can distinguish unseen positive and negative samples.
Despite the absence of negative samples during model training, recent positive unlabeled learning methods that use weighted loss functions enable successful model training in this challenging data regime.
One contribution to this topic presents a framework that uses explicit estimates of predictive uncertainty to enable self-training in such positive unlabeled settings.
The next section introduces the concept of weakly supervised learning with pairwise binary constraint annotations for constrained clustering.
One contribution in this area proposes a method that combines it with concepts from semi-supervised learning to train these models in a semi-constrained manner.
This allows the use of large amounts of completely unlabeled data to guide model training on a smaller dataset with pairwise binary constraint annotations.
Another contribution in this area leverages the cluster detection capabilities of these models to recognize dynamically changing categories.
The final section includes a description of transfer learning approaches as well as an application of transfer learning with learning tasks of varying granularity in a medical context.Der Erfolg moderner überwachter Machine- und Deep-Learning-Modelle, die mittlerweile Teil unseres Alltags geworden sind, fußt teilweise auf der zunehmenden Verfügbarkeit großer Datensätze mit hochwertigen Annotationen.
Die Verfügbarkeit solcher Annotationen, auch als Labels bezeichnet, bleibt jedoch ein kritisches Bottleneck für viele Anwendungen, da sie eine Voraussetzung für das Training von überwachten Modellen darstellen.
Dies ist insbesondere in Domänen ein Problem, in welchen der Prozess der Datenannotation unklar und aufwändig ist oder das Wissen von Fachexperten erfordert.
In Anwendungsbereichen wie der medizinischen Bildgebung oder der industriellen Fertigung führt dies oft dazu, dass trotz der Verfügbarkeit einer großen Menge nicht annotierter Daten nur ein Bruchteil dieser Daten mit geeigneten Labels versehen werden kann.
Das Modelltraining mit begrenzt annotierten Daten ist der Fokus dieser Arbeit, welche Methoden behandelt, um auch nicht annotierte Daten und schwach annotierte Daten, deren Annotationen geringen Informationsgehalt besitzen, für das Modelltraining zu nutzen.
Die vorliegende Arbeit enthält Beiträge zu Semi-supervised Learning, Positive-unlabeled Learning, Constrained Clustering und Transfer Learning.
Zu Beginn wird das Konzept des Semi-supervised Learning vorgestellt und es wird ein Überblick über die aktuelle Forschung zu Self-Training, Entropy-Regularisation, Consistency-Regularisation und hybriden Ansätzen gegeben.
Das Ziel von Semi-supervised Learning besteht darin, Modelle auf einem kleinen Datensatz mit annotierten Trainingsdaten zu trainieren, wobei zusätzlich ein größerer Datensatz von nicht annotierten Daten in das Modelltraining mit einbezogen wird.
Da die wichtigsten Entwicklungen in diesem Bereich aus dem Bereich Computer Vision getrieben werden, wurden viele dieser Methoden hauptsächlich für Bilddaten entwickelt.
Dies motivierte einen Beitrag zur Untersuchung ihrer Anwendung in einem Szenario zur Klassifizierung von Zeitreihen.
Ein weiterer Beitrag untersucht die Anwendbarkeit von Semi-supervised Learning in einem medizinischen Bildgebungskontext, um den Aufwand für die Datenannotation zu reduzieren.
Positive Unlabeled Learning ist ein weiteres Teilgebiet von Semi-supervised Learning.
Dabei enthält der Trainingsdatensatz nur positive oder nicht annotierte Datenpunkte, während das Ziel darin besteht, einen binären Klassifikator zu lernen, der zwischen positiven und negativen Datenpunkten unterscheiden kann.
Trotz des Fehlens annotierter negativer Datenpunkte während des Modelltrainings ermöglichen Positive Unlabeled Learning Methoden ein erfolgreiches Modelltraining in dieser schwierigen Datensituation.
Ein Beitrag zu diesem Thema stellt ein Framework vor, welches explizite Schätzungen der Vorhersageunsicherheit verwendet, um Self-Training in solch einem Kontext zu ermöglichen.
Ein weiterer Abschnitt stellt das Konzept des Weakly-supervised Learning mit paarweisen binären Constraint Annotationen für Constrained Clustering vor.
Ein Beitrag in diesem Bereich schlägt eine Methode vor, die Constrained Clustering mit Konzepten aus dem Semi-supervised Learning kombiniert, um auch nicht annotierte Daten für das Training dieser Modelle zu verwenden.
Dies ermöglicht die Verwendung großer Mengen nicht-annotierter Daten, um das Modelltraining auf einem kleineren Datensatz mit paarweisen binären Constraint Annotationen zu verbessern.
Ein weiterer Beitrag in diesem Bereich nutzt die Fähigkeiten dieser Modelle zur Clustererkennung, um dynamisch wechselnde Kategorien in den Daten zu erkennen.
Schließlich enthält ein weiterer Abschnitt eine Beschreibung von Transfer Learning-Ansätzen sowie eine Anwendung von Transfer Learning mit Machine Learning-Problemen unterschiedlicher Granularität in einem medizinischen Kontext
Disentangling multi sensory contribution to navigation in a rat using a novel freely moving virtual reality-enabled path integration task
Get PDFThe ability to assess distance and direction traveled plays an existential role within the animal kingdom. Animals use many kinds of sensory systems in order to be able to achieve this feat. These sensory systems combine information into two frames of reference which animals use to retrieve a previously taken path.
Animals are mainly able to use both an allocentric reference frame, which helps to associate objects to each other, and an egocentric reference frame, which ties a connection between ourselves and a given object. The use of the egocentric frame is highlighted when outside cues are limited or absent. Under such conditions, like in the darkness, we rely on our ability to integrate our past path in order to retrieve a starting location. Understanding this process of path integration also
means helping to uncover the evolutionary basis for memory; the mechanism for planning and integrating information in our mind has evolved from a physical need to navigate in the environment (György Buzsáki and E. I. Moser 2013). There is much evidence to support the similarity between physical and mental travel (György Buzsáki and James J Chrobak 2005; Eichenbaum et al. 1999).
Researchers have tried to understand this process of integrating egocentric cues using many animal behavioral assays. These assays have been hard to combine with recordings from the brain due to their inability to be repeated over multiple trials within the same animal, or a lack of precision in accurately measuring animal behavior. Many of the assays don’t have a full control over visual or vestibular perturbation in order to study their respective contributions.
Here I developed a novel path integration task and used a freely moving virtual reality system (Del Grosso and Sirota 2019) to enable variable homing locations within an arena and implement perturbations of the vestibular input and an optic flow. This task structure can be split into path integration and goal diected segments within a trial, is suitable for long-term extracellular recordings, and results in a high throughput of trials. I recorded over 50 000 trials and over 100 000 rears which were essential in correctly characterizing the task, quantifying performance, and testing for the use of allocentric information. Flexible homing location and binary response of the rat (rearing) allow for objective measures of the accuracy in the path integration task. With freely moving virtual reality I could create conflicting information to study the contribution of each idiothetic sensory input to the path-integration. I show that transient visual or vestibular perturbations during the task segment associated with path integration strongly deteriorate the performance of the animal on the task. Such perturbations are expected to disrupt the head-direction system, which is necessary for successful path integration, and hence strongly suggests the developed task is indeed purely path integration-dependent. I present a preliminary analysis of the multiple place cells’ recordings from the CA1 layer of the hippocampus which provides the first glimpse on the spatial coding of the same cells during different modes of navigation associated with distinct segments of the task. Finally, I discuss how our task can be adapted easily in the future to further investigate this complicated mechanism
Renal CD64+ MPs protect kidney in both cisplatin and ischemia reperfusion induced acute kidney injury
Get PDFMononuclear phagocytes (MPs) consist of dendritic cells (DCs), macrophages and monocytes, they have important functions in mediating immune responses in tissues. It is reported in kidney contains diverse MP subsets, however, due to insufficient subset specific depletion model, the functions of these MPs are unclear in acute kidney injury (AKI). C-Type Lectin Domain Containing 9A (Clec9a) is a receptor sensing polymeric F-actin exposed in necrotic cells, and it is highly expressed on common dendritic cell progenitor (CDP) that give rise to dendritic cells (DCs), by genetic tracing of Clec9a expression histroy will specifically mark DCs but not other immune cells. In steady state the kidney contains at least 4 subsets with Clec9a expressing history, cDC1, cDC2 and a large CD64 expressing group which can be subdivided into F4/80hi and CD11bhi sub-populations. By using cDC1 depletion model (XCR1-venus-DTR), we found cDC1 depletion did not increase kidney injury severity after cisplatin treatment. The similar situation also happened when we depleted cDC2 in Clec9acreIRF4f/f mice treated with cisplatin. However, cisplatin induced injury increased significantly when we used newly generated Clec9acreCD64iDTR model to deplete CD64+ MPs in the kidney, this result indicated CD64+ MPs can protect kidney in cisplatin induced AKI. By constructing ischemia reperfusion acute kidney injury, we found CD64+ MPs also have protective role in both steady state and inflammation state, the mechanism could be attributed to regulation of anti/pro-inflammatory cytokines produced by CD64+ MPs. Thus, we conclude CD64+ MPs protect kidney both in cisplatin induced AKI and also in ischemia reperfusion induced AKI, which may shed lights to therapeutic targeting of this population for curing AKI in the future
Ionoacoustics for three-dimensional particle bunch monitoring
Get PDFIonoacoustics refers to the measurement of acoustic waves excited by ions as they slow down in matter. One particularly interesting application is the monitoring of individual laser-accelerated ion bunches, which challenges many established online detection systems. The ionoacoustic approach offers advantages in this context. It is resistant to the strong electromagnetic pulse emitted during the laser-plasma interaction, and it does not exhibit saturation even at very high particle fluxes. Recovering the Ion-Bunch Energy distribution from a recorded Acoustic Trace (I-BEAT) has conceptually been proven. This work introduces two new detector designs, the I-BEAT 3D and the TI-BEAT detector, and develops a fast analysis routine based on an analytical model. The result is a novel and versatile online detector for short and intense ion bunches that enables immediate feedback.
Describing the dose deposition of protons in matter, particularly in the entrance window and the water reservoir of the detector, connects the most important proton bunch parameters to the resulting ionoacoustic signals and hence serves for the development of a simple and fast data analysis method for particle bunches with Gaussian energy spectra and lateral distributions. This is key to the analysis and interpretation of the acoustic traces obtained with the detector designs.
The I-BEAT 3D detector uses four transducers and is studied in experiments at two laser-driven proton beamlines located at CALA and the HZDR. Via the new analytical model, it provides the mean bunch energy, energy spread and lateral bunch position of individual bunches in less than one second. Additionally, the detector allows the monitoring of relative changes in lateral bunch size. Experiments with proton bunches at kinetic energies between 10 MeV and 30 MeV reveal sub-MeV and sub-mm resolution. Cross-calibration using radiochromic films proves that the new methodology can determine the absolute number of particles contained in a single bunch with uncertainties of only 10%.
The TI-BEAT detector uses a short water reservoir to allow ions with sufficient energy to pass through. It is developed to prove that ionoacoustics can also function as transmission monitor for particle bunches. The detector concept is evaluated at the GSI’s SIS-18 synchrotron with xenon ions at kinetic energies of 385 MeV/u. The precision of determining the absolute lateral position and relative changes in the lateral size of ion bunches is 0.15 mm and less than 0.1 mm, respectively. The particle number is extracted with a precision between 4% and 8%.
The new detector designs and the data evaluation method provide a fast, compact, cost-effective and electromagnetic pulse-resistant online tool for monitoring important ion bunch parameters at application sites. This provides the foundation for irradiation studies and can serve as feedback for automated optimization of laser-driven proton sources
Investigations on oxygen-rich materials with various functional groups and oxygen-containing bridged azoles
Get PDF
Statistical methods for sparse functional object data: elastic curves, shapes and densities
Get PDFMany applications naturally yield data that can be viewed as elements in non-linear spaces. Consequently, there is a need for non-standard statistical methods capable of handling such data. The work presented here deals with the analysis of data in complex spaces derived from functional L2-spaces as quotient spaces (or subsets of such spaces). These data types include elastic curves represented as d-dimensional functions modulo re-parametrization, planar shapes represented as 2-dimensional functions modulo rotation, scaling and translation, and elastic planar shapes combining all of these invariances. Moreover, also probability densities can be thought of as non-negative functions modulo scaling. Since these functional object data spaces lack a natural Hilbert space structure, this work proposes specialized methods that integrate techniques from functional data analysis with those for metric and manifold data. In particular, but not exclusively, novel regression methods for specific metric quotient spaces are discussed. Special attention is given to handling discrete observations, since in practice curves and shapes are typically observed only as a discrete (often sparse or irregular) set of points. Similarly, density functions are usually not directly observed, but a (small) sample from the corresponding probability distribution is available. Overall, this work comprises six contributions that propose new methods for sparse functional object data and apply them to relevant real-world datasets, predominantly in a biomedical context