@article{BaumhoerDietzKneiseletal.2019, author = {Baumhoer, Celia A. and Dietz, Andreas J. and Kneisel, C. and Kuenzer, C.}, title = {Automated Extraction of Antarctic Glacier and Ice Shelf Fronts from Sentinel-1 Imagery Using Deep Learning}, series = {Remote Sensing}, volume = {11}, journal = {Remote Sensing}, number = {21}, issn = {2072-4292}, doi = {10.3390/rs11212529}, url = {http://nbn-resolving.de/urn:nbn:de:bvb:20-opus-193150}, pages = {2529}, year = {2019}, abstract = {Sea level rise contribution from the Antarctic ice sheet is influenced by changes in glacier and ice shelf front position. Still, little is known about seasonal glacier and ice shelf front fluctuations as the manual delineation of calving fronts from remote sensing imagery is very time-consuming. The major challenge of automatic calving front extraction is the low contrast between floating glacier and ice shelf fronts and the surrounding sea ice. Additionally, in previous decades, remote sensing imagery over the often cloud-covered Antarctic coastline was limited. Nowadays, an abundance of Sentinel-1 imagery over the Antarctic coastline exists and could be used for tracking glacier and ice shelf front movement. To exploit the available Sentinel-1 data, we developed a processing chain allowing automatic extraction of the Antarctic coastline from Seninel-1 imagery and the creation of dense time series to assess calving front change. The core of the proposed workflow is a modified version of the deep learning architecture U-Net. This convolutional neural network (CNN) performs a semantic segmentation on dual-pol Sentinel-1 data and the Antarctic TanDEM-X digital elevation model (DEM). The proposed method is tested for four training and test areas along the Antarctic coastline. The automatically extracted fronts deviate on average 78 m in training and 108 m test areas. Spatial and temporal transferability is demonstrated on an automatically extracted 15-month time series along the Getz Ice Shelf. Between May 2017 and July 2018, the fronts along the Getz Ice Shelf show mostly an advancing tendency with the fastest moving front of DeVicq Glacier with 726 ± 20 m/yr.}, language = {en} } @article{PookFreudenthalKorteetal.2020, author = {Pook, Torsten and Freudenthal, Jan and Korte, Arthur and Simianer, Henner}, title = {Using Local Convolutional Neural Networks for Genomic Prediction}, series = {Frontiers in Genetics}, volume = {11}, journal = {Frontiers in Genetics}, doi = {10.3389/fgene.2020.561497}, url = {http://nbn-resolving.de/urn:nbn:de:bvb:20-opus-216436}, year = {2020}, abstract = {The prediction of breeding values and phenotypes is of central importance for both livestock and crop breeding. In this study, we analyze the use of artificial neural networks (ANN) and, in particular, local convolutional neural networks (LCNN) for genomic prediction, as a region-specific filter corresponds much better with our prior genetic knowledge on the genetic architecture of traits than traditional convolutional neural networks. Model performances are evaluated on a simulated maize data panel (n = 10,000; p = 34,595) and real Arabidopsis data (n = 2,039; p = 180,000) for a variety of traits based on their predictive ability. The baseline LCNN, containing one local convolutional layer (kernel size: 10) and two fully connected layers with 64 nodes each, is outperforming commonly proposed ANNs (multi layer perceptrons and convolutional neural networks) for basically all considered traits. For traits with high heritability and large training population as present in the simulated data, LCNN are even outperforming state-of-the-art methods like genomic best linear unbiased prediction (GBLUP), Bayesian models and extended GBLUP, indicated by an increase in predictive ability of up to 24\%. However, for small training populations, these state-of-the-art methods outperform all considered ANNs. Nevertheless, the LCNN still outperforms all other considered ANNs by around 10\%. Minor improvements to the tested baseline network architecture of the LCNN were obtained by increasing the kernel size and of reducing the stride, whereas the number of subsequent fully connected layers and their node sizes had neglectable impact. Although gains in predictive ability were obtained for large scale data sets by using LCNNs, the practical use of ANNs comes with additional problems, such as the need of genotyping all considered individuals, the lack of estimation of heritability and reliability. Furthermore, breeding values are additive by design, whereas ANN-based estimates are not. However, ANNs also comes with new opportunities, as networks can easily be extended to account for additional inputs (omics, weather etc.) and outputs (multi-trait models), and computing time increases linearly with the number of individuals. With advances in high-throughput phenotyping and cheaper genotyping, ANNs can become a valid alternative for genomic prediction.}, language = {en} } @article{HoeserBachoferKuenzer2020, author = {Hoeser, Thorsten and Bachofer, Felix and Kuenzer, Claudia}, title = {Object detection and image segmentation with deep learning on Earth Observation data: a review — part II: applications}, series = {Remote Sensing}, volume = {12}, journal = {Remote Sensing}, number = {18}, issn = {2072-4292}, doi = {10.3390/rs12183053}, url = {http://nbn-resolving.de/urn:nbn:de:bvb:20-opus-213152}, year = {2020}, abstract = {In Earth observation (EO), large-scale land-surface dynamics are traditionally analyzed by investigating aggregated classes. The increase in data with a very high spatial resolution enables investigations on a fine-grained feature level which can help us to better understand the dynamics of land surfaces by taking object dynamics into account. To extract fine-grained features and objects, the most popular deep-learning model for image analysis is commonly used: the convolutional neural network (CNN). In this review, we provide a comprehensive overview of the impact of deep learning on EO applications by reviewing 429 studies on image segmentation and object detection with CNNs. We extensively examine the spatial distribution of study sites, employed sensors, used datasets and CNN architectures, and give a thorough overview of applications in EO which used CNNs. Our main finding is that CNNs are in an advanced transition phase from computer vision to EO. Upon this, we argue that in the near future, investigations which analyze object dynamics with CNNs will have a significant impact on EO research. With a focus on EO applications in this Part II, we complete the methodological review provided in Part I.}, language = {en} } @article{HoeserKuenzer2020, author = {Hoeser, Thorsten and Kuenzer, Claudia}, title = {Object detection and image segmentation with deep learning on Earth observation data: a review-part I: evolution and recent trends}, series = {Remote Sensing}, volume = {12}, journal = {Remote Sensing}, number = {10}, issn = {2072-4292}, doi = {10.3390/rs12101667}, url = {http://nbn-resolving.de/urn:nbn:de:bvb:20-opus-205918}, year = {2020}, abstract = {Deep learning (DL) has great influence on large parts of science and increasingly established itself as an adaptive method for new challenges in the field of Earth observation (EO). Nevertheless, the entry barriers for EO researchers are high due to the dense and rapidly developing field mainly driven by advances in computer vision (CV). To lower the barriers for researchers in EO, this review gives an overview of the evolution of DL with a focus on image segmentation and object detection in convolutional neural networks (CNN). The survey starts in 2012, when a CNN set new standards in image recognition, and lasts until late 2019. Thereby, we highlight the connections between the most important CNN architectures and cornerstones coming from CV in order to alleviate the evaluation of modern DL models. Furthermore, we briefly outline the evolution of the most popular DL frameworks and provide a summary of datasets in EO. By discussing well performing DL architectures on these datasets as well as reflecting on advances made in CV and their impact on future research in EO, we narrow the gap between the reviewed, theoretical concepts from CV and practical application in EO.}, language = {en} } @article{AnkenbrandLohrSchloetelburgetal.2021, author = {Ankenbrand, Markus Johannes and Lohr, David and Schl{\"o}telburg, Wiebke and Reiter, Theresa and Wech, Tobias and Schreiber, Laura Maria}, title = {Deep learning-based cardiac cine segmentation: Transfer learning application to 7T ultrahigh-field MRI}, series = {Magnetic Resonance in Medicine}, volume = {86}, journal = {Magnetic Resonance in Medicine}, number = {4}, doi = {10.1002/mrm.28822}, url = {http://nbn-resolving.de/urn:nbn:de:bvb:20-opus-257604}, pages = {2179-2191}, year = {2021}, abstract = {Purpose Artificial neural networks show promising performance in automatic segmentation of cardiac MRI. However, training requires large amounts of annotated data and generalization to different vendors, field strengths, sequence parameters, and pathologies is limited. Transfer learning addresses this challenge, but specific recommendations regarding type and amount of data required is lacking. In this study, we assess data requirements for transfer learning to experimental cardiac MRI at 7T where the segmentation task can be challenging. In addition, we provide guidelines, tools, and annotated data to enable transfer learning approaches by other researchers and clinicians. Methods A publicly available segmentation model was used to annotate a publicly available data set. This labeled data set was subsequently used to train a neural network for segmentation of left ventricle and myocardium in cardiac cine MRI. The network is used as starting point for transfer learning to 7T cine data of healthy volunteers (n = 22; 7873 images) by updating the pre-trained weights. Structured and random data subsets of different sizes were used to systematically assess data requirements for successful transfer learning. Results Inconsistencies in the publically available data set were corrected, labels created, and a neural network trained. On 7T cardiac cine images the model pre-trained on public imaging data, acquired at 1.5T and 3T, achieved DICE\(_{LV}\) = 0.835 and DICE\(_{MY}\) = 0.670. Transfer learning using 7T cine data and ImageNet weight initialization improved model performance to DICE\(_{LV}\) = 0.900 and DICE\(_{MY}\) = 0.791. Using only end-systolic and end-diastolic images reduced training data by 90\%, with no negative impact on segmentation performance (DICE\(_{LV}\) = 0.908, DICE\(_{MY}\) = 0.805). Conclusions This work demonstrates and quantifies the benefits of transfer learning for cardiac cine image segmentation. We provide practical guidelines for researchers planning transfer learning projects in cardiac MRI and make data, models, and code publicly available.}, language = {en} } @article{DirscherlDietzKneiseletal.2021, author = {Dirscherl, Mariel and Dietz, Andreas J. and Kneisel, Christof and Kuenzer, Claudia}, title = {A novel method for automated supraglacial lake mapping in Antarctica using Sentinel-1 SAR imagery and deep learning}, series = {Remote Sensing}, volume = {13}, journal = {Remote Sensing}, number = {2}, issn = {2072-4292}, doi = {10.3390/rs13020197}, url = {http://nbn-resolving.de/urn:nbn:de:bvb:20-opus-222998}, year = {2021}, abstract = {Supraglacial meltwater accumulation on ice sheets can be a main driver for accelerated ice discharge, mass loss, and global sea-level-rise. With further increasing surface air temperatures, meltwater-induced hydrofracturing, basal sliding, or surface thinning will cumulate and most likely trigger unprecedented ice mass loss on the Greenland and Antarctic ice sheets. While the Greenland surface hydrological network as well as its impacts on ice dynamics and mass balance has been studied in much detail, Antarctic supraglacial lakes remain understudied with a circum-Antarctic record of their spatio-temporal development entirely lacking. This study provides the first automated supraglacial lake extent mapping method using Sentinel-1 synthetic aperture radar (SAR) imagery over Antarctica and complements the developed optical Sentinel-2 supraglacial lake detection algorithm presented in our companion paper. In detail, we propose the use of a modified U-Net for semantic segmentation of supraglacial lakes in single-polarized Sentinel-1 imagery. The convolutional neural network (CNN) is implemented with residual connections for optimized performance as well as an Atrous Spatial Pyramid Pooling (ASPP) module for multiscale feature extraction. The algorithm is trained on 21,200 Sentinel-1 image patches and evaluated in ten spatially or temporally independent test acquisitions. In addition, George VI Ice Shelf is analyzed for intra-annual lake dynamics throughout austral summer 2019/2020 and a decision-level fused Sentinel-1 and Sentinel-2 maximum lake extent mapping product is presented for January 2020 revealing a more complete supraglacial lake coverage (~770 km\(^2\)) than the individual single-sensor products. Classification results confirm the reliability of the proposed workflow with an average Kappa coefficient of 0.925 and a F\(_1\)-score of 93.0\% for the supraglacial water class across all test regions. Furthermore, the algorithm is applied in an additional test region covering supraglacial lakes on the Greenland ice sheet which further highlights the potential for spatio-temporal transferability. Future work involves the integration of more training data as well as intra-annual analyses of supraglacial lake occurrence across the whole continent and with focus on supraglacial lake development throughout a summer melt season and into Antarctic winter.}, language = {en} } @article{JanieschZschechHeinrich2021, author = {Janiesch, Christian and Zschech, Patrick and Heinrich, Kai}, title = {Machine learning and deep learning}, series = {Electronic Markets}, volume = {31}, journal = {Electronic Markets}, number = {3}, issn = {1422-8890}, doi = {10.1007/s12525-021-00475-2}, url = {http://nbn-resolving.de/urn:nbn:de:bvb:20-opus-270155}, pages = {685-695}, year = {2021}, abstract = {Today, intelligent systems that offer artificial intelligence capabilities often rely on machine learning. Machine learning describes the capacity of systems to learn from problem-specific training data to automate the process of analytical model building and solve associated tasks. Deep learning is a machine learning concept based on artificial neural networks. For many applications, deep learning models outperform shallow machine learning models and traditional data analysis approaches. In this article, we summarize the fundamentals of machine learning and deep learning to generate a broader understanding of the methodical underpinning of current intelligent systems. In particular, we provide a conceptual distinction between relevant terms and concepts, explain the process of automated analytical model building through machine learning and deep learning, and discuss the challenges that arise when implementing such intelligent systems in the field of electronic markets and networked business. These naturally go beyond technological aspects and highlight issues in human-machine interaction and artificial intelligence servitization.}, language = {en} } @article{AnkenbrandShainbergHocketal.2021, author = {Ankenbrand, Markus J. and Shainberg, Liliia and Hock, Michael and Lohr, David and Schreiber, Laura M.}, title = {Sensitivity analysis for interpretation of machine learning based segmentation models in cardiac MRI}, series = {BMC Medical Imaging}, volume = {21}, journal = {BMC Medical Imaging}, number = {1}, doi = {10.1186/s12880-021-00551-1}, url = {http://nbn-resolving.de/urn:nbn:de:bvb:20-opus-259169}, pages = {27}, year = {2021}, abstract = {Background Image segmentation is a common task in medical imaging e.g., for volumetry analysis in cardiac MRI. Artificial neural networks are used to automate this task with performance similar to manual operators. However, this performance is only achieved in the narrow tasks networks are trained on. Performance drops dramatically when data characteristics differ from the training set properties. Moreover, neural networks are commonly considered black boxes, because it is hard to understand how they make decisions and why they fail. Therefore, it is also hard to predict whether they will generalize and work well with new data. Here we present a generic method for segmentation model interpretation. Sensitivity analysis is an approach where model input is modified in a controlled manner and the effect of these modifications on the model output is evaluated. This method yields insights into the sensitivity of the model to these alterations and therefore to the importance of certain features on segmentation performance. Results We present an open-source Python library (misas), that facilitates the use of sensitivity analysis with arbitrary data and models. We show that this method is a suitable approach to answer practical questions regarding use and functionality of segmentation models. We demonstrate this in two case studies on cardiac magnetic resonance imaging. The first case study explores the suitability of a published network for use on a public dataset the network has not been trained on. The second case study demonstrates how sensitivity analysis can be used to evaluate the robustness of a newly trained model. Conclusions Sensitivity analysis is a useful tool for deep learning developers as well as users such as clinicians. It extends their toolbox, enabling and improving interpretability of segmentation models. Enhancing our understanding of neural networks through sensitivity analysis also assists in decision making. Although demonstrated only on cardiac magnetic resonance images this approach and software are much more broadly applicable.}, language = {en} } @article{WengHeidenreichMetzetal.2021, author = {Weng, Andreas M. and Heidenreich, Julius F. and Metz, Corona and Veldhoen, Simon and Bley, Thorsten A. and Wech, Tobias}, title = {Deep learning-based segmentation of the lung in MR-images acquired by a stack-of-spirals trajectory at ultra-short echo-times}, series = {BMC Medical Imaging}, volume = {21}, journal = {BMC Medical Imaging}, doi = {10.1186/s12880-021-00608-1}, url = {http://nbn-resolving.de/urn:nbn:de:bvb:20-opus-260520}, year = {2021}, abstract = {Background Functional lung MRI techniques are usually associated with time-consuming post-processing, where manual lung segmentation represents the most cumbersome part. The aim of this study was to investigate whether deep learning-based segmentation of lung images which were scanned by a fast UTE sequence exploiting the stack-of-spirals trajectory can provide sufficiently good accuracy for the calculation of functional parameters. Methods In this study, lung images were acquired in 20 patients suffering from cystic fibrosis (CF) and 33 healthy volunteers, by a fast UTE sequence with a stack-of-spirals trajectory and a minimum echo-time of 0.05 ms. A convolutional neural network was then trained for semantic lung segmentation using 17,713 2D coronal slices, each paired with a label obtained from manual segmentation. Subsequently, the network was applied to 4920 independent 2D test images and results were compared to a manual segmentation using the S{\o}rensen-Dice similarity coefficient (DSC) and the Hausdorff distance (HD). Obtained lung volumes and fractional ventilation values calculated from both segmentations were compared using Pearson's correlation coefficient and Bland Altman analysis. To investigate generalizability to patients outside the CF collective, in particular to those exhibiting larger consolidations inside the lung, the network was additionally applied to UTE images from four patients with pneumonia and one with lung cancer. Results The overall DSC for lung tissue was 0.967 ± 0.076 (mean ± standard deviation) and HD was 4.1 ± 4.4 mm. Lung volumes derived from manual and deep learning based segmentations as well as values for fractional ventilation exhibited a high overall correlation (Pearson's correlation coefficent = 0.99 and 1.00). For the additional cohort with unseen pathologies / consolidations, mean DSC was 0.930 ± 0.083, HD = 12.9 ± 16.2 mm and the mean difference in lung volume was 0.032 ± 0.048 L. Conclusions Deep learning-based image segmentation in stack-of-spirals based lung MRI allows for accurate estimation of lung volumes and fractional ventilation values and promises to replace the time-consuming step of manual image segmentation in the future.}, language = {en} } @article{PennigHoyerKrauskopfetal.2021, author = {Pennig, Lenhard and Hoyer, Ulrike Cornelia Isabel and Krauskopf, Alexandra and Shahzad, Rahil and J{\"u}nger, Stephanie T. and Thiele, Frank and Laukamp, Kai Roman and Grunz, Jan-Peter and Perkuhn, Michael and Schlamann, Marc and Kabbasch, Christoph and Borggrefe, Jan and Goertz, Lukas}, title = {Deep learning assistance increases the detection sensitivity of radiologists for secondary intracranial aneurysms in subarachnoid hemorrhage}, series = {Neuroradiology}, volume = {63}, journal = {Neuroradiology}, number = {12}, issn = {0028-3940}, doi = {10.1007/s00234-021-02697-9}, url = {http://nbn-resolving.de/urn:nbn:de:bvb:20-opus-308117}, pages = {1985-1994}, year = {2021}, abstract = {Purpose To evaluate whether a deep learning model (DLM) could increase the detection sensitivity of radiologists for intracranial aneurysms on CT angiography (CTA) in aneurysmal subarachnoid hemorrhage (aSAH). Methods Three different DLMs were trained on CTA datasets of 68 aSAH patients with 79 aneurysms with their outputs being combined applying ensemble learning (DLM-Ens). The DLM-Ens was evaluated on an independent test set of 104 aSAH patients with 126 aneuryms (mean volume 129.2 ± 185.4 mm3, 13.0\% at the posterior circulation), which were determined by two radiologists and one neurosurgeon in consensus using CTA and digital subtraction angiography scans. CTA scans of the test set were then presented to three blinded radiologists (reader 1: 13, reader 2: 4, and reader 3: 3 years of experience in diagnostic neuroradiology), who assessed them individually for aneurysms. Detection sensitivities for aneurysms of the readers with and without the assistance of the DLM were compared. Results In the test set, the detection sensitivity of the DLM-Ens (85.7\%) was comparable to the radiologists (reader 1: 91.2\%, reader 2: 86.5\%, and reader 3: 86.5\%; Fleiss κ of 0.502). DLM-assistance significantly increased the detection sensitivity (reader 1: 97.6\%, reader 2: 97.6\%,and reader 3: 96.0\%; overall P=.024; Fleiss κ of 0.878), especially for secondary aneurysms (88.2\% of the additional aneurysms provided by the DLM). Conclusion Deep learning significantly improved the detection sensitivity of radiologists for aneurysms in aSAH, especially for secondary aneurysms. It therefore represents a valuable adjunct for physicians to establish an accurate diagnosis in order to optimize patient treatment.}, language = {en} } @phdthesis{Hoeser2022, author = {H{\"o}ser, Thorsten}, title = {Global Dynamics of the Offshore Wind Energy Sector Derived from Earth Observation Data - Deep Learning Based Object Detection Optimised with Synthetic Training Data for Offshore Wind Energy Infrastructure Extraction from Sentinel-1 Imagery}, doi = {10.25972/OPUS-29285}, url = {http://nbn-resolving.de/urn:nbn:de:bvb:20-opus-292857}, school = {Universit{\"a}t W{\"u}rzburg}, year = {2022}, abstract = {The expansion of renewable energies is being driven by the gradual phaseout of fossil fuels in order to reduce greenhouse gas emissions, the steadily increasing demand for energy and, more recently, by geopolitical events. The offshore wind energy sector is on the verge of a massive expansion in Europe, the United Kingdom, China, but also in the USA, South Korea and Vietnam. Accordingly, the largest marine infrastructure projects to date will be carried out in the upcoming decades, with thousands of offshore wind turbines being installed. In order to accompany this process globally and to provide a database for research, development and monitoring, this dissertation presents a deep learning-based approach for object detection that enables the derivation of spatiotemporal developments of offshore wind energy infrastructures from satellite-based radar data of the Sentinel-1 mission. For training the deep learning models for offshore wind energy infrastructure detection, an approach is presented that makes it possible to synthetically generate remote sensing data and the necessary annotation for the supervised deep learning process. In this synthetic data generation process, expert knowledge about image content and sensor acquisition techniques is made machine-readable. Finally, extensive and highly variable training data sets are generated from this knowledge representation, with which deep learning models can learn to detect objects in real-world satellite data. The method for the synthetic generation of training data based on expert knowledge offers great potential for deep learning in Earth observation. Applications of deep learning based methods can be developed and tested faster with this procedure. Furthermore, the synthetically generated and thus controllable training data offer the possibility to interpret the learning process of the optimised deep learning models. The method developed in this dissertation to create synthetic remote sensing training data was finally used to optimise deep learning models for the global detection of offshore wind energy infrastructure. For this purpose, images of the entire global coastline from ESA's Sentinel-1 radar mission were evaluated. The derived data set includes over 9,941 objects, which distinguish offshore wind turbines, transformer stations and offshore wind energy infrastructures under construction from each other. In addition to this spatial detection, a quarterly time series from July 2016 to June 2021 was derived for all objects. This time series reveals the start of construction, the construction phase and the time of completion with subsequent operation for each object. The derived offshore wind energy infrastructure data set provides the basis for an analysis of the development of the offshore wind energy sector from July 2016 to June 2021. For this analysis, further attributes of the detected offshore wind turbines were derived. The most important of these are the height and installed capacity of a turbine. The turbine height was calculated by a radargrammetric analysis of the previously detected Sentinel-1 signal and then used to statistically model the installed capacity. The results show that in June 2021, 8,885 offshore wind turbines with a total capacity of 40.6 GW were installed worldwide. The largest installed capacities are in the EU (15.2 GW), China (14.1 GW) and the United Kingdom (10.7 GW). From July 2016 to June 2021, China has expanded 13 GW of offshore wind energy infrastructure. The EU has installed 8 GW and the UK 5.8 GW of offshore wind energy infrastructure in the same period. This temporal analysis shows that China was the main driver of the expansion of the offshore wind energy sector in the period under investigation. The derived data set for the description of the offshore wind energy sector was made publicly available. It is thus freely accessible to all decision-makers and stakeholders involved in the development of offshore wind energy projects. Especially in the scientific context, it serves as a database that enables a wide range of investigations. Research questions regarding offshore wind turbines themselves as well as the influence of the expansion in the coming decades can be investigated. This supports the imminent and urgently needed expansion of offshore wind energy in order to promote sustainable expansion in addition to the expansion targets that have been set.}, language = {en} } @article{WechAnkenbrandBleyetal.2022, author = {Wech, Tobias and Ankenbrand, Markus Johannes and Bley, Thorsten Alexander and Heidenreich, Julius Frederik}, title = {A data-driven semantic segmentation model for direct cardiac functional analysis based on undersampled radial MR cine series}, series = {Magnetic Resonance in Medicine}, volume = {87}, journal = {Magnetic Resonance in Medicine}, number = {2}, doi = {10.1002/mrm.29017}, url = {http://nbn-resolving.de/urn:nbn:de:bvb:20-opus-257616}, pages = {972-983}, year = {2022}, abstract = {Purpose Image acquisition and subsequent manual analysis of cardiac cine MRI is time-consuming. The purpose of this study was to train and evaluate a 3D artificial neural network for semantic segmentation of radially undersampled cardiac MRI to accelerate both scan time and postprocessing. Methods A database of Cartesian short-axis MR images of the heart (148,500 images, 484 examinations) was assembled from an openly accessible database and radial undersampling was simulated. A 3D U-Net architecture was pretrained for segmentation of undersampled spatiotemporal cine MRI. Transfer learning was then performed using samples from a second database, comprising 108 non-Cartesian radial cine series of the midventricular myocardium to optimize the performance for authentic data. The performance was evaluated for different levels of undersampling by the Dice similarity coefficient (DSC) with respect to reference labels, as well as by deriving ventricular volumes and myocardial masses. Results Without transfer learning, the pretrained model performed moderately on true radial data [maximum number of projections tested, P = 196; DSC = 0.87 (left ventricle), DSC = 0.76 (myocardium), and DSC =0.64 (right ventricle)]. After transfer learning with authentic data, the predictions achieved human level even for high undersampling rates (P = 33, DSC = 0.95, 0.87, and 0.93) without significant difference compared with segmentations derived from fully sampled data. Conclusion A 3D U-Net architecture can be used for semantic segmentation of radially undersampled cine acquisitions, achieving a performance comparable with human experts in fully sampled data. This approach can jointly accelerate time-consuming cine image acquisition and cumbersome manual image analysis.}, language = {en} } @phdthesis{Griebel2022, author = {Griebel, Matthias}, title = {Applied Deep Learning: from Data to Deployment}, doi = {10.25972/OPUS-27765}, url = {http://nbn-resolving.de/urn:nbn:de:bvb:20-opus-277650}, school = {Universit{\"a}t W{\"u}rzburg}, year = {2022}, abstract = {Novel deep learning (DL) architectures, better data availability, and a significant increase in computing power have enabled scientists to solve problems that were considered unassailable for many years. A case in point is the "protein folding problem", a 50-year-old grand challenge in biology that was recently solved by the DL-system AlphaFold. Other examples comprise the development of large DL-based language models that, for instance, generate newspaper articles that hardly differ from those written by humans. However, developing unbiased, reliable, and accurate DL models for various practical applications remains a major challenge - and many promising DL projects get stuck in the piloting stage, never to be completed. In light of these observations, this thesis investigates the practical challenges encountered throughout the life cycle of DL projects and proposes solutions to develop and deploy rigorous DL models. The first part of the thesis is concerned with prototyping DL solutions in different domains. First, we conceptualize guidelines for applied image recognition and showcase their application in a biomedical research project. Next, we illustrate the bottom-up development of a DL backend for an augmented intelligence system in the manufacturing sector. We then turn to the fashion domain and present an artificial curation system for individual fashion outfit recommendations that leverages DL techniques and unstructured data from social media and fashion blogs. After that, we showcase how DL solutions can assist fashion designers in the creative process. Finally, we present our award-winning DL solution for the segmentation of glomeruli in human kidney tissue images that was developed for the Kaggle data science competition HuBMAP - Hacking the Kidney. The second part continues the development path of the biomedical research project beyond the prototyping stage. Using data from five laboratories, we show that ground truth estimation from multiple human annotators and training of DL model ensembles help to establish objectivity, reliability, and validity in DL-based bioimage analyses. In the third part, we present deepflash2, a DL solution that addresses the typical challenges encountered during training, evaluation, and application of DL models in bioimaging. The tool facilitates the objective and reliable segmentation of ambiguous bioimages through multi-expert annotations and integrated quality assurance. It is embedded in an easy-to-use graphical user interface and offers best-in-class predictive performance for semantic and instance segmentation under economical usage of computational resources.}, language = {en} } @article{KrenzerMakowskiHekaloetal.2022, author = {Krenzer, Adrian and Makowski, Kevin and Hekalo, Amar and Fitting, Daniel and Troya, Joel and Zoller, Wolfram G. and Hann, Alexander and Puppe, Frank}, title = {Fast machine learning annotation in the medical domain: a semi-automated video annotation tool for gastroenterologists}, series = {BioMedical Engineering OnLine}, volume = {21}, journal = {BioMedical Engineering OnLine}, number = {1}, doi = {10.1186/s12938-022-01001-x}, url = {http://nbn-resolving.de/urn:nbn:de:bvb:20-opus-300231}, year = {2022}, abstract = {Background Machine learning, especially deep learning, is becoming more and more relevant in research and development in the medical domain. For all the supervised deep learning applications, data is the most critical factor in securing successful implementation and sustaining the progress of the machine learning model. Especially gastroenterological data, which often involves endoscopic videos, are cumbersome to annotate. Domain experts are needed to interpret and annotate the videos. To support those domain experts, we generated a framework. With this framework, instead of annotating every frame in the video sequence, experts are just performing key annotations at the beginning and the end of sequences with pathologies, e.g., visible polyps. Subsequently, non-expert annotators supported by machine learning add the missing annotations for the frames in-between. Methods In our framework, an expert reviews the video and annotates a few video frames to verify the object's annotations for the non-expert. In a second step, a non-expert has visual confirmation of the given object and can annotate all following and preceding frames with AI assistance. After the expert has finished, relevant frames will be selected and passed on to an AI model. This information allows the AI model to detect and mark the desired object on all following and preceding frames with an annotation. Therefore, the non-expert can adjust and modify the AI predictions and export the results, which can then be used to train the AI model. Results Using this framework, we were able to reduce workload of domain experts on average by a factor of 20 on our data. This is primarily due to the structure of the framework, which is designed to minimize the workload of the domain expert. Pairing this framework with a state-of-the-art semi-automated AI model enhances the annotation speed further. Through a prospective study with 10 participants, we show that semi-automated annotation using our tool doubles the annotation speed of non-expert annotators compared to a well-known state-of-the-art annotation tool. Conclusion In summary, we introduce a framework for fast expert annotation for gastroenterologists, which reduces the workload of the domain expert considerably while maintaining a very high annotation quality. The framework incorporates a semi-automated annotation system utilizing trained object detection models. The software and framework are open-source.}, language = {en} } @article{PhilippDietzUllmannetal.2022, author = {Philipp, Marius and Dietz, Andreas and Ullmann, Tobias and Kuenzer, Claudia}, title = {Automated extraction of annual erosion rates for Arctic permafrost coasts using Sentinel-1, Deep Learning, and Change Vector Analysis}, series = {Remote Sensing}, volume = {14}, journal = {Remote Sensing}, number = {15}, issn = {2072-4292}, doi = {10.3390/rs14153656}, url = {http://nbn-resolving.de/urn:nbn:de:bvb:20-opus-281956}, year = {2022}, abstract = {Arctic permafrost coasts become increasingly vulnerable due to environmental drivers such as the reduced sea-ice extent and duration as well as the thawing of permafrost itself. A continuous quantification of the erosion process on large to circum-Arctic scales is required to fully assess the extent and understand the consequences of eroding permafrost coastlines. This study presents a novel approach to quantify annual Arctic coastal erosion and build-up rates based on Sentinel-1 (S1) Synthetic Aperture RADAR (SAR) backscatter data, in combination with Deep Learning (DL) and Change Vector Analysis (CVA). The methodology includes the generation of a high-quality Arctic coastline product via DL, which acted as a reference for quantifying coastal erosion and build-up rates from annual median and standard deviation (sd) backscatter images via CVA. The analysis was applied on ten test sites distributed across the Arctic and covering about 1038 km of coastline. Results revealed maximum erosion rates of up to 160 m for some areas and an average erosion rate of 4.37 m across all test sites within a three-year temporal window from 2017 to 2020. The observed erosion rates within the framework of this study agree with findings published in the previous literature. The proposed methods and data can be applied on large scales and, prospectively, even for the entire Arctic. The generated products may be used for quantifying the loss of frozen ground, estimating the release of stored organic material, and can act as a basis for further related studies in Arctic coastal environments.}, language = {en} } @article{VollmerSaraviVollmeretal.2022, author = {Vollmer, Andreas and Saravi, Babak and Vollmer, Michael and Lang, Gernot Michael and Straub, Anton and Brands, Roman C. and K{\"u}bler, Alexander and Gubik, Sebastian and Hartmann, Stefan}, title = {Artificial intelligence-based prediction of oroantral communication after tooth extraction utilizing preoperative panoramic radiography}, series = {Diagnostics}, volume = {12}, journal = {Diagnostics}, number = {6}, issn = {2075-4418}, doi = {10.3390/diagnostics12061406}, url = {http://nbn-resolving.de/urn:nbn:de:bvb:20-opus-278814}, year = {2022}, abstract = {Oroantral communication (OAC) is a common complication after tooth extraction of upper molars. Profound preoperative panoramic radiography analysis might potentially help predict OAC following tooth extraction. In this exploratory study, we evaluated n = 300 consecutive cases (100 OAC and 200 controls) and trained five machine learning algorithms (VGG16, InceptionV3, MobileNetV2, EfficientNet, and ResNet50) to predict OAC versus non-OAC (binary classification task) from the input images. Further, four oral and maxillofacial experts evaluated the respective panoramic radiography and determined performance metrics (accuracy, area under the curve (AUC), precision, recall, F1-score, and receiver operating characteristics curve) of all diagnostic approaches. Cohen's kappa was used to evaluate the agreement between expert evaluations. The deep learning algorithms reached high specificity (highest specificity 100\% for InceptionV3) but low sensitivity (highest sensitivity 42.86\% for MobileNetV2). The AUCs from VGG16, InceptionV3, MobileNetV2, EfficientNet, and ResNet50 were 0.53, 0.60, 0.67, 0.51, and 0.56, respectively. Expert 1-4 reached an AUC of 0.550, 0.629, 0.500, and 0.579, respectively. The specificity of the expert evaluations ranged from 51.74\% to 95.02\%, whereas sensitivity ranged from 14.14\% to 59.60\%. Cohen's kappa revealed a poor agreement for the oral and maxillofacial expert evaluations (Cohen's kappa: 0.1285). Overall, present data indicate that OAC cannot be sufficiently predicted from preoperative panoramic radiography. The false-negative rate, i.e., the rate of positive cases (OAC) missed by the deep learning algorithms, ranged from 57.14\% to 95.24\%. Surgeons should not solely rely on panoramic radiography when evaluating the probability of OAC occurrence. Clinical testing of OAC is warranted after each upper-molar tooth extraction.}, language = {en} } @article{BrandTroyaKrenzeretal.2022, author = {Brand, Markus and Troya, Joel and Krenzer, Adrian and Saßmannshausen, Zita and Zoller, Wolfram G. and Meining, Alexander and Lux, Thomas J. and Hann, Alexander}, title = {Development and evaluation of a deep learning model to improve the usability of polyp detection systems during interventions}, series = {United European Gastroenterology Journal}, volume = {10}, journal = {United European Gastroenterology Journal}, number = {5}, doi = {10.1002/ueg2.12235}, url = {http://nbn-resolving.de/urn:nbn:de:bvb:20-opus-312708}, pages = {477-484}, year = {2022}, abstract = {Background The efficiency of artificial intelligence as computer-aided detection (CADe) systems for colorectal polyps has been demonstrated in several randomized trials. However, CADe systems generate many distracting detections, especially during interventions such as polypectomies. Those distracting CADe detections are often induced by the introduction of snares or biopsy forceps as the systems have not been trained for such situations. In addition, there are a significant number of non-false but not relevant detections, since the polyp has already been previously detected. All these detections have the potential to disturb the examiner's work. Objectives Development and evaluation of a convolutional neuronal network that recognizes instruments in the endoscopic image, suppresses distracting CADe detections, and reliably detects endoscopic interventions. Methods A total of 580 different examination videos from 9 different centers using 4 different processor types were screened for instruments and represented the training dataset (519,856 images in total, 144,217 contained a visible instrument). The test dataset included 10 full-colonoscopy videos that were analyzed for the recognition of visible instruments and detections by a commercially available CADe system (GI Genius, Medtronic). Results The test dataset contained 153,623 images, 8.84\% of those presented visible instruments (12 interventions, 19 instruments used). The convolutional neuronal network reached an overall accuracy in the detection of visible instruments of 98.59\%. Sensitivity and specificity were 98.55\% and 98.92\%, respectively. A mean of 462.8 frames containing distracting CADe detections per colonoscopy were avoided using the convolutional neuronal network. This accounted for 95.6\% of all distracting CADe detections. Conclusions Detection of endoscopic instruments in colonoscopy using artificial intelligence technology is reliable and achieves high sensitivity and specificity. Accordingly, the new convolutional neuronal network could be used to reduce distracting CADe detections during endoscopic procedures. Thus, our study demonstrates the great potential of artificial intelligence technology beyond mucosal assessment.}, language = {en} } @article{LuxBanckSassmannshausenetal.2022, author = {Lux, Thomas J. and Banck, Michael and Saßmannshausen, Zita and Troya, Joel and Krenzer, Adrian and Fitting, Daniel and Sudarevic, Boban and Zoller, Wolfram G. and Puppe, Frank and Meining, Alexander and Hann, Alexander}, title = {Pilot study of a new freely available computer-aided polyp detection system in clinical practice}, series = {International Journal of Colorectal Disease}, volume = {37}, journal = {International Journal of Colorectal Disease}, number = {6}, doi = {10.1007/s00384-022-04178-8}, url = {http://nbn-resolving.de/urn:nbn:de:bvb:20-opus-324459}, pages = {1349-1354}, year = {2022}, abstract = {Purpose Computer-aided polyp detection (CADe) systems for colonoscopy are already presented to increase adenoma detection rate (ADR) in randomized clinical trials. Those commercially available closed systems often do not allow for data collection and algorithm optimization, for example regarding the usage of different endoscopy processors. Here, we present the first clinical experiences of a, for research purposes publicly available, CADe system. Methods We developed an end-to-end data acquisition and polyp detection system named EndoMind. Examiners of four centers utilizing four different endoscopy processors used EndoMind during their clinical routine. Detected polyps, ADR, time to first detection of a polyp (TFD), and system usability were evaluated (NCT05006092). Results During 41 colonoscopies, EndoMind detected 29 of 29 adenomas in 66 of 66 polyps resulting in an ADR of 41.5\%. Median TFD was 130 ms (95\%-CI, 80-200 ms) while maintaining a median false positive rate of 2.2\% (95\%-CI, 1.7-2.8\%). The four participating centers rated the system using the System Usability Scale with a median of 96.3 (95\%-CI, 70-100). Conclusion EndoMind's ability to acquire data, detect polyps in real-time, and high usability score indicate substantial practical value for research and clinical practice. Still, clinical benefit, measured by ADR, has to be determined in a prospective randomized controlled trial.}, language = {en} } @article{DawoodBreuerStebanietal.2023, author = {Dawood, Peter and Breuer, Felix and Stebani, Jannik and Burd, Paul and Homolya, Istv{\´a}n and Oberberger, Johannes and Jakob, Peter M. and Blaimer, Martin}, title = {Iterative training of robust k-space interpolation networks for improved image reconstruction with limited scan specific training samples}, series = {Magnetic Resonance in Medicine}, volume = {89}, journal = {Magnetic Resonance in Medicine}, number = {2}, doi = {10.1002/mrm.29482}, url = {http://nbn-resolving.de/urn:nbn:de:bvb:20-opus-312306}, pages = {812 -- 827}, year = {2023}, abstract = {To evaluate an iterative learning approach for enhanced performance of robust artificial-neural-networks for k-space interpolation (RAKI), when only a limited amount of training data (auto-calibration signals [ACS]) are available for accelerated standard 2D imaging. Methods In a first step, the RAKI model was tailored for the case of limited training data amount. In the iterative learning approach (termed iterative RAKI [iRAKI]), the tailored RAKI model is initially trained using original and augmented ACS obtained from a linear parallel imaging reconstruction. Subsequently, the RAKI convolution filters are refined iteratively using original and augmented ACS extracted from the previous RAKI reconstruction. Evaluation was carried out on 200 retrospectively undersampled in vivo datasets from the fastMRI neuro database with different contrast settings. Results For limited training data (18 and 22 ACS lines for R = 4 and R = 5, respectively), iRAKI outperforms standard RAKI by reducing residual artifacts and yields better noise suppression when compared to standard parallel imaging, underlined by quantitative reconstruction quality metrics. Additionally, iRAKI shows better performance than both GRAPPA and standard RAKI in case of pre-scan calibration with varying contrast between training- and undersampled data. Conclusion RAKI benefits from the iterative learning approach, which preserves the noise suppression feature, but requires less original training data for the accurate reconstruction of standard 2D images thereby improving net acceleration.}, language = {en} } @phdthesis{Somody2023, author = {Somody, Joseph Christian Campbell}, title = {Leveraging deep learning for identification and structural determination of novel protein complexes from \(in\) \(situ\) electron cryotomography of \(Mycoplasma\) \(pneumoniae\)}, doi = {10.25972/OPUS-31344}, url = {http://nbn-resolving.de/urn:nbn:de:bvb:20-opus-313447}, school = {Universit{\"a}t W{\"u}rzburg}, year = {2023}, abstract = {The holy grail of structural biology is to study a protein in situ, and this goal has been fast approaching since the resolution revolution and the achievement of atomic resolution. A cell's interior is not a dilute environment, and proteins have evolved to fold and function as needed in that environment; as such, an investigation of a cellular component should ideally include the full complexity of the cellular environment. Imaging whole cells in three dimensions using electron cryotomography is the best method to accomplish this goal, but it comes with a limitation on sample thickness and produces noisy data unamenable to direct analysis. This thesis establishes a novel workflow to systematically analyse whole-cell electron cryotomography data in three dimensions and to find and identify instances of protein complexes in the data to set up a determination of their structure and identity for success. Mycoplasma pneumoniae is a very small parasitic bacterium with fewer than 700 protein-coding genes, is thin enough and small enough to be imaged in large quantities by electron cryotomography, and can grow directly on the grids used for imaging, making it ideal for exploratory studies in structural proteomics. As part of the workflow, a methodology for training deep-learning-based particle-picking models is established. As a proof of principle, a dataset of whole-cell Mycoplasma pneumoniae tomograms is used with this workflow to characterize a novel membrane-associated complex observed in the data. Ultimately, 25431 such particles are picked from 353 tomograms and refined to a density map with a resolution of 11 {\AA}. Making good use of orthogonal datasets to filter search space and verify results, structures were predicted for candidate proteins and checked for suitable fit in the density map. In the end, with this approach, nine proteins were found to be part of the complex, which appears to be associated with chaperone activity and interact with translocon machinery. Visual proteomics refers to the ultimate potential of in situ electron cryotomography: the comprehensive interpretation of tomograms. The workflow presented here is demonstrated to help in reaching that potential.}, subject = {Kryoelektronenmikroskopie}, language = {en} }