Refine
Has Fulltext
- yes (18)
Is part of the Bibliography
- yes (18)
Year of publication
Document Type
- Journal article (16)
- Doctoral Thesis (1)
- Preprint (1)
Language
- English (18)
Keywords
- deep learning (4)
- cardiac magnetic resonance (2)
- foraging (2)
- honeybee (2)
- juvenile hormone (2)
- neural networks (2)
- segmentation (2)
- triglycerides (2)
- 16S metabarcoding (1)
- 7 T (1)
Institute
- Theodor-Boveri-Institut für Biowissenschaften (11)
- Center for Computational and Theoretical Biology (7)
- Deutsches Zentrum für Herzinsuffizienz (DZHI) (5)
- Institut für diagnostische und interventionelle Radiologie (Institut für Röntgendiagnostik) (4)
- Julius-von-Sachs-Institut für Biowissenschaften (4)
- Medizinische Klinik und Poliklinik I (2)
- Graduate School of Life Sciences (1)
ResearcherID
- D-1221-2009 (1)
EU-Project number / Contract (GA) number
Sensitivity analysis for interpretation of machine learning based segmentation models in cardiac MRI
(2021)
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.
RNA sequencing (RNA-seq) has become a powerful tool to understand molecular mechanisms and/or developmental programs. It provides a fast, reliable and cost-effective method to access sets of expressed elements in a qualitative and quantitative manner. Especially for non-model organisms and in absence of a reference genome, RNA-seq data is used to reconstruct and quantify transcriptomes at the same time. Even SNPs, InDels, and alternative splicing events are predicted directly from the data without having a reference genome at hand. A key challenge, especially for non-computational personnal, is the management of the resulting datasets, consisting of different data types and formats. Here, we present TBro, a flexible de novo transcriptome browser, tackling this challenge. TBro aggregates sequences, their annotation, expression levels as well as differential testing results. It provides an easy-to-use interface to mine the aggregated data and generate publication-ready visualizations. Additionally, it supports users with an intuitive cart system, that helps collecting and analysing biological meaningful sets of transcripts. TBro’s modular architecture allows easy extension of its functionalities in the future. Especially, the integration of new data types such as proteomic quantifications or array-based gene expression data is straightforward. Thus, TBro is a fully featured yet flexible transcriptome browser that supports approaching complex biological questions and enhances collaboration of numerous researchers.
New experimental methods have drastically accelerated the pace and quantity at which biological data is generated. High-throughput DNA sequencing is one of the pivotal new technologies. It offers a number of novel applications in various fields of biology, including ecology, evolution, and genomics. However, together with those opportunities many new challenges arise. Specialized algorithms and software are required to cope with the amount of data, often requiring substantial training in bioinformatic methods. Another way to make those data accessible to non-bioinformaticians is the development of programs with intuitive user interfaces.
In my thesis I developed analyses and programs to tackle current problems with high-throughput data in biology. In the field of ecology this covers the establishment of the bioinformatic workflow for pollen DNA meta-barcoding. Furthermore, I developed an application that facilitates the analysis of ecological communities in the context of their traits. Information from multiple public databases have been aggregated and can now be mapped automatically to existing community tables for interactive inspection. In evolution the new data are used to reconstruct phylogenetic trees from multiple genes. I developed the tool bcgTree to automate this process for bacteria. Many plant genomes have been sequenced in current years. Sequencing reads of those projects also contain data from the chloroplasts. The tool chloroExtractor supports the targeted extraction and analysis of the chloroplast genome. To compare the structure of multiple genomes specialized software is required for calculation and visualization of the relationships. I developed AliTV to address this. In contrast to existing programs for this task it allows interactive adjustments of produced graphics. Thus, facilitating the discovery of biologically relevant information. Another application I developed helps to analyze transcriptomes even if no reference genome is present. This is achieved by aggregating the different pieces of information, like functional annotation and expression level, for each transcript in a web platform. Scientists can then search, filter, subset, and visualize the transcriptome.
Together the methods and tools expedite insights into biological systems that were not possible before.
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.
Although the concept of botanical carnivory has been known since Darwin's time, the molecular mechanisms that allow animal feeding remain unknown, primarily due to a complete lack of genomic information. Here, we show that the transcriptomic landscape of the Dionaea trap is dramatically shifted toward signal transduction and nutrient transport upon insect feeding, with touch hormone signaling and protein secretion prevailing. At the same time, a massive induction of general defense responses is accompanied by the repression of cell death-related genes/processes. We hypothesize that the carnivory syndrome of Dionaea evolved by exaptation of ancient defense pathways, replacing cell death with nutrient acquisition.
Abstract
Cell lineage decisions occur in three-dimensional spatial patterns that are difficult to identify by eye. There is an ongoing effort to replicate such patterns using mathematical modeling. One approach uses long ranging cell-cell communication to replicate common spatial arrangements like checkerboard and engulfing patterns. In this model, the cell-cell communication has been implemented as a signal that disperses throughout the tissue. On the other hand, machine learning models have been developed for pattern recognition and pattern reconstruction tasks. We combined synthetic data generated by the mathematical model with spatial summary statistics and deep learning algorithms to recognize and reconstruct cell fate patterns in organoids of mouse embryonic stem cells. Application of Moran’s index and pair correlation functions for in vitro and synthetic data from the model showed local clustering and radial segregation. To assess the patterns as a whole, a graph neural network was developed and trained on synthetic data from the model. Application to in vitro data predicted a low signal dispersion value. To test this result, we implemented a multilayer perceptron for the prediction of a given cell fate based on the fates of the neighboring cells. The results show a 70% accuracy of cell fate imputation based on the nine nearest neighbors of a cell. Overall, our approach combines deep learning with mathematical modeling to link cell fate patterns with potential underlying mechanisms.
Author summary
Mammalian embryo development relies on organized differentiation of stem cells into different lineages. Particularly at the early stages of embryogenesis, cells of different fates form three-dimensional spatial patterns that are difficult to identify by eye. Pattern quantification and mathematical modeling have produced first insights into potential mechanisms for the cell fate arrangements. However, these approaches have relied on classifications of the patterns such as inside-out or random, or used summary statistics such as pair correlation functions or cluster radii. Deep neural networks allow characterizing patterns directly. Since the tissue context can be readily reproduced by a graph, we implemented a graph neural network to characterize the patterns of embryonic stem cell organoids as a whole. In addition, we implemented a multilayer perceptron model to reconstruct the fate of a given cell based on its neighbors. To train and test the models, we used synthetic data generated by our mathematical model for cell-cell communication. This interplay of deep learning and mathematical modeling in combination with summary statistics allowed us to identify a potential mechanism for cell fate determination in mouse embryonic stem cells. Our results agree with a mechanism with a dispersion of the intercellular signal that links a cell’s fate to those of the local neighborhood.
Young grapevines (Vitis vinifera) suffer and eventually can die from the crown gall disease caused by the plant pathogen Allorhizobium vitis (Rhizobiaceae). Virulent members of A. vitis harbor a tumor-inducing plasmid and induce formation of crown galls due to the oncogenes encoded on the transfer DNA. The expression of oncogenes in transformed host cells induces unregulated cell proliferation and metabolic and physiological changes. The crown gall produces opines uncommon to plants, which provide an important nutrient source for A. vitis harboring opine catabolism enzymes. Crown galls host a distinct bacterial community, and the mechanisms establishing a crown gall–specific bacterial community are currently unknown. Thus, we were interested in whether genes homologous to those of the tumor-inducing plasmid coexist in the genomes of the microbial species coexisting in crown galls. We isolated 8 bacterial strains from grapevine crown galls, sequenced their genomes, and tested their virulence and opine utilization ability in bioassays. In addition, the 8 genome sequences were compared with 34 published bacterial genomes, including closely related plant-associated bacteria not from crown galls. Homologous genes for virulence and opine anabolism were only present in the virulent Rhizobiaceae. In contrast, homologs of the opine catabolism genes were present in all strains including the nonvirulent members of the Rhizobiaceae and non-Rhizobiaceae. Gene neighborhood and sequence identity of the opine degradation cluster of virulent and nonvirulent strains together with the results of the opine utilization assay support the important role of opine utilization for cocolonization in crown galls, thereby shaping the crown gall community.
Purpose
To fully automatically derive quantitative parameters from late gadolinium enhancement (LGE) cardiac MR (CMR) in patients with myocardial infarction and to investigate if phase sensitive or magnitude reconstructions or a combination of both results in best segmentation accuracy.
Methods
In this retrospective single center study, a convolutional neural network with a U-Net architecture with a self-configuring framework (“nnU-net”) was trained for segmentation of left ventricular myocardium and infarct zone in LGE-CMR. A database of 170 examinations from 78 patients with history of myocardial infarction was assembled. Separate fitting of the model was performed, using phase sensitive inversion recovery, the magnitude reconstruction or both contrasts as input channels.
Manual labelling served as ground truth. In a subset of 10 patients, the performance of the trained models was evaluated and quantitatively compared by determination of the Sørensen-Dice similarity coefficient (DSC) and volumes of the infarct zone compared with the manual ground truth using Pearson’s r correlation and Bland-Altman analysis.
Results
The model achieved high similarity coefficients for myocardium and scar tissue. No significant difference was observed between using PSIR, magnitude reconstruction or both contrasts as input (PSIR and MAG; mean DSC: 0.83 ± 0.03 for myocardium and 0.72 ± 0.08 for scars). A strong correlation for volumes of infarct zone was observed between manual and model-based approach (r = 0.96), with a significant underestimation of the volumes obtained from the neural network.
Conclusion
The self-configuring nnU-net achieves predictions with strong agreement compared to manual segmentation, proving the potential as a promising tool to provide fully automatic quantitative evaluation of LGE-CMR.
1.Honeybees Apis mellifera and other pollinating insects suffer from pesticides in agricultural landscapes. Flupyradifurone is the active ingredient of a novel pesticide by the name of ‘Sivanto’, introduced by Bayer AG (Crop Science Division, Monheim am Rhein, Germany). It is recommended against sucking insects and marketed as ‘harmless’ to honeybees. Flupyradifurone binds to nicotinergic acetylcholine receptors like neonicotinoids, but it has a different mode of action. So far, little is known on how sublethal flupyradifurone doses affect honeybees.
2. We chronically applied a sublethal and field‐realistic concentration of flupyradifurone to test for long‐term effects on flight behaviour using radio‐frequency identification. We examined haematoxylin/eosin‐stained brains of flupyradifurone‐treated bees to investigate possible changes in brain morphology and brain damage.
3. A field‐realistic flupyradifurone dose of approximately 1.0 μg/bee/day significantly increased mortality. Pesticide‐treated bees initiated foraging earlier than control bees. No morphological damage in the brain was observed.
4. Synthesis and applications. The early onset of foraging induced by a chronical application of flupyradifurone could be disadvantageous for honeybee colonies, reducing the period of in‐hive tasks and life expectancy of individuals. Radio‐frequency identification technology is a valuable tool for studying pesticide effects on lifetime foraging behaviour of insects.
Purpose
Inhomogeneities of the static magnetic B\(_{0}\) field are a major limiting factor in cardiac MRI at ultrahigh field (≥ 7T), as they result in signal loss and image distortions. Different magnetic susceptibilities of the myocardium and surrounding tissue in combination with cardiac motion lead to strong spatio‐temporal B\(_{0}\)‐field inhomogeneities, and their homogenization (B0 shimming) is a prerequisite. Limitations of state‐of‐the‐art shimming are described, regional B\(_{0}\) variations are measured, and a methodology for spherical harmonics shimming of the B\(_{0}\) field within the human myocardium is proposed.
Methods
The spatial B\(_{0}\)‐field distribution in the heart was analyzed as well as temporal B\(_{0}\)‐field variations in the myocardium over the cardiac cycle. Different shim region‐of‐interest selections were compared, and hardware limitations of spherical harmonics B\(_{0}\) shimming were evaluated by calibration‐based B0‐field modeling. The role of third‐order spherical harmonics terms was analyzed as well as potential benefits from cardiac phase–specific shimming.
Results
The strongest B\(_{0}\)‐field inhomogeneities were observed in localized spots within the left‐ventricular and right‐ventricular myocardium and varied between systolic and diastolic cardiac phases. An anatomy‐driven shim region‐of‐interest selection allowed for improved B\(_{0}\)‐field homogeneity compared with a standard shim region‐of‐interest cuboid. Third‐order spherical harmonics terms were demonstrated to be beneficial for shimming of these myocardial B\(_{0}\)‐field inhomogeneities. Initial results from the in vivo implementation of a potential shim strategy were obtained. Simulated cardiac phase–specific shimming was performed, and a shim term‐by‐term analysis revealed periodic variations of required currents.
Conclusion
Challenges in state‐of‐the‐art B\(_{0}\) shimming of the human heart at 7 T were described. Cardiac phase–specific shimming strategies were found to be superior to vendor‐supplied shimming.