Refine
Has Fulltext
- yes (52)
Is part of the Bibliography
- yes (52)
Document Type
- Journal article (50)
- Doctoral Thesis (2)
Language
- English (52) (remove)
Keywords
- machine learning (52) (remove)
Institute
- Institut für Geographie und Geologie (9)
- Institut für Informatik (9)
- Center for Computational and Theoretical Biology (5)
- Institut für Klinische Epidemiologie und Biometrie (5)
- Medizinische Klinik und Poliklinik II (4)
- Pathologisches Institut (4)
- Theodor-Boveri-Institut für Biowissenschaften (4)
- Betriebswirtschaftliches Institut (3)
- Institut für diagnostische und interventionelle Radiologie (Institut für Röntgendiagnostik) (3)
- Klinik und Poliklinik für Dermatologie, Venerologie und Allergologie (3)
Sonstige beteiligte Institutionen
Background
A basic requirement for artificial intelligence (AI)–based image analysis systems, which are to be integrated into clinical practice, is a high robustness. Minor changes in how those images are acquired, for example, during routine skin cancer screening, should not change the diagnosis of such assistance systems.
Objective
To quantify to what extent minor image perturbations affect the convolutional neural network (CNN)–mediated skin lesion classification and to evaluate three possible solutions for this problem (additional data augmentation, test-time augmentation, anti-aliasing).
Methods
We trained three commonly used CNN architectures to differentiate between dermoscopic melanoma and nevus images. Subsequently, their performance and susceptibility to minor changes (‘brittleness’) was tested on two distinct test sets with multiple images per lesion. For the first set, image changes, such as rotations or zooms, were generated artificially. The second set contained natural changes that stemmed from multiple photographs taken of the same lesions.
Results
All architectures exhibited brittleness on the artificial and natural test set. The three reviewed methods were able to decrease brittleness to varying degrees while still maintaining performance. The observed improvement was greater for the artificial than for the natural test set, where enhancements were minor.
Conclusions
Minor image changes, relatively inconspicuous for humans, can have an effect on the robustness of CNNs differentiating skin lesions. By the methods tested here, this effect can be reduced, but not fully eliminated. Thus, further research to sustain the performance of AI classifiers is needed to facilitate the translation of such systems into the clinic.
Among the defense strategies developed in microbes over millions of years, the innate adaptive CRISPR-Cas immune systems have spread across most of bacteria and archaea. The flexibility, simplicity, and specificity of CRISPR-Cas systems have laid the foundation for CRISPR-based genetic tools. Yet, the efficient administration of CRISPR-based tools demands rational designs to maximize the on-target efficiency and off-target specificity. Specifically, the selection of guide RNAs (gRNAs), which play a crucial role in the target recognition of CRISPR-Cas systems, is non-trivial. Despite the fact that the emerging machine learning techniques provide a solution to aid in gRNA design with prediction algorithms, design rules for many CRISPR-Cas systems are ill-defined, hindering their broader applications.
CRISPR interference (CRISPRi), an alternative gene silencing technique using a catalytically dead Cas protein to interfere with transcription, is a leading technique in bacteria for functional interrogation, pathway manipulation, and genome-wide screens. Although the application is promising, it also is hindered by under-investigated design rules. Therefore, in this work, I develop a state-of-art predictive machine learning model for guide silencing efficiency in bacteria leveraging the advantages of feature engineering, data integration, interpretable AI, and automated machine learning. I first systematically investigate the influential factors that attribute to the extent of depletion in multiple CRISPRi genome-wide essentiality screens in Escherichia coli and demonstrate the surprising dominant contribution of gene-specific effects, such as gene expression level. These observations allowed me to segregate the confounding gene-specific effects using a mixed-effect random forest (MERF) model to provide a better estimate of guide efficiency, together with the improvement led by integrating multiple screens. The MERF model outperformed existing tools in an independent high-throughput saturating screen. I next interpret the predictive model to extract the design rules for robust gene silencing, such as the preference for cytosine and disfavoring for guanine and thymine within and around the protospacer adjacent motif (PAM) sequence. I further incorporated the MERF model in a web-based tool that is freely accessible at www.ciao.helmholtz-hiri.de.
When comparing the MERF model with existing tools, the performance of the alternative gRNA design tool optimized for CRISPRi in eukaryotes when applied to bacteria was far from satisfying, questioning the robustness of prediction algorithms across organisms. In addition, the CRISPR-Cas systems exhibit diverse mechanisms albeit with some similarities. The captured predictive patterns from one dataset thereby are at risk of poor generalization when applied across organisms and CRISPR-Cas techniques. To fill the gap, the machine learning approach I present here for CRISPRi could serve as a blueprint for the effective development of prediction algorithms for specific organisms or CRISPR-Cas systems of interest. The explicit workflow includes three principle steps: 1) accommodating the feature set for the CRISPR-Cas system or technique; 2) optimizing a machine learning model using automated machine learning; 3) explaining the model using interpretable AI. To illustrate the applicability of the workflow and diversity of results when applied across different bacteria and CRISPR-Cas systems, I have applied this workflow to analyze three distinct CRISPR-Cas genome-wide screens. From the CRISPR base editor essentiality screen in E. coli, I have determined the PAM preference and sequence context in the editing window for efficient editing, such as A at the 2nd position of PAM, A/TT/TG downstream of PAM, and TC at the 4th to 5th position of gRNAs. From the CRISPR-Cas13a screen in E. coli, in addition to the strong correlation with the guide depletion, the target expression level is the strongest predictor in the model, supporting it as a main determinant of the activation of Cas13-induced immunity and better characterizing the CRISPR-Cas13 system. From the CRISPR-Cas12a screen in Klebsiella pneumoniae, I have extracted the design rules for robust antimicrobial activity across K. pneumoniae strains and provided a predictive algorithm for gRNA design, facilitating CRISPR-Cas12a as an alternative technique to tackle antibiotic resistance.
Overall, this thesis presents an accurate prediction algorithm for CRISPRi guide efficiency in bacteria, providing insights into the determinants of efficient silencing and guide designs. The systematic exploration has led to a robust machine learning approach for effective model development in other bacteria and CRISPR-Cas systems. Applying the approach in the analysis of independent CRISPR-Cas screens not only sheds light on the design rules but also the mechanisms of the CRISPR-Cas systems. Together, I demonstrate that applied machine learning paves the way to a deeper understanding and a broader application of CRISPR-Cas systems.
Objectives
Open-access cancer imaging datasets have become integral for evaluating novel AI approaches in radiology. However, their use in quantitative analysis with radiomics features presents unique challenges, such as incomplete documentation, low visibility, non-uniform data formats, data inhomogeneity, and complex preprocessing. These issues may cause problems with reproducibility and standardization in radiomics studies.
Methods
We systematically reviewed imaging datasets with public copyright licenses, published up to March 2023 across four large online cancer imaging archives. We included only datasets with tomographic images (CT, MRI, or PET), segmentations, and clinical annotations, specifically identifying those suitable for radiomics research. Reproducible preprocessing and feature extraction were performed for each dataset to enable their easy reuse.
Results
We discovered 29 datasets with corresponding segmentations and labels in the form of health outcomes, tumor pathology, staging, imaging-based scores, genetic markers, or repeated imaging. We compiled a repository encompassing 10,354 patients and 49,515 scans. Of the 29 datasets, 15 were licensed under Creative Commons licenses, allowing both non-commercial and commercial usage and redistribution, while others featured custom or restricted licenses. Studies spanned from the early 1990s to 2021, with the majority concluding after 2013. Seven different formats were used for the imaging data. Preprocessing and feature extraction were successfully performed for each dataset.
Conclusion
RadiomicsHub is a comprehensive public repository with radiomics features derived from a systematic review of public cancer imaging datasets. By converting all datasets to a standardized format and ensuring reproducible and traceable processing, RadiomicsHub addresses key reproducibility and standardization challenges in radiomics.
Critical relevance statement
This study critically addresses the challenges associated with locating, preprocessing, and extracting quantitative features from open-access datasets, to facilitate more robust and reliable evaluations of radiomics models.
Key points
- Through a systematic review, we identified 29 cancer imaging datasets suitable for radiomics research.
- A public repository with collection overview and radiomics features, encompassing 10,354 patients and 49,515 scans, was compiled.
- Most datasets can be shared, used, and built upon freely under a Creative Commons license.
- All 29 identified datasets have been converted into a common format to enable reproducible radiomics feature extraction.
Variability of gene expression due to stochasticity of transcription or variation of extrinsic signals, termed biological noise, is a potential driving force of cellular differentiation. Utilizing single-cell RNA-sequencing, we develop VarID2 for the quantification of biological noise at single-cell resolution. VarID2 reveals enhanced nuclear versus cytoplasmic noise, and distinct regulatory modes stratified by correlation between noise, expression, and chromatin accessibility. Noise levels are minimal in murine hematopoietic stem cells (HSCs) and increase during differentiation and ageing. Differential noise identifies myeloid-biased Dlk1+ long-term HSCs in aged mice with enhanced quiescence and self-renewal capacity. VarID2 reveals noise dynamics invisible to conventional single-cell transcriptome analysis.
Highlights
• Brain connectivity states identified by cofluctuation strength.
• CMEP as new method to robustly predict human traits from brain imaging data.
• Network-identifying connectivity ‘events’ are not predictive of cognitive ability.
• Sixteen temporally independent fMRI time frames allow for significant prediction.
• Neuroimaging-based assessment of cognitive ability requires sufficient scan lengths.
Abstract
Human functional brain connectivity can be temporally decomposed into states of high and low cofluctuation, defined as coactivation of brain regions over time. Rare states of particularly high cofluctuation have been shown to reflect fundamentals of intrinsic functional network architecture and to be highly subject-specific. However, it is unclear whether such network-defining states also contribute to individual variations in cognitive abilities – which strongly rely on the interactions among distributed brain regions. By introducing CMEP, a new eigenvector-based prediction framework, we show that as few as 16 temporally separated time frames (< 1.5% of 10 min resting-state fMRI) can significantly predict individual differences in intelligence (N = 263, p < .001). Against previous expectations, individual's network-defining time frames of particularly high cofluctuation do not predict intelligence. Multiple functional brain networks contribute to the prediction, and all results replicate in an independent sample (N = 831). Our results suggest that although fundamentals of person-specific functional connectomes can be derived from few time frames of highest connectivity, temporally distributed information is necessary to extract information about cognitive abilities. This information is not restricted to specific connectivity states, like network-defining high-cofluctuation states, but rather reflected across the entire length of the brain connectivity time series.
Background
Colorectal cancer is a leading cause of cancer-related deaths worldwide. The best method to prevent CRC is a colonoscopy. However, not all colon polyps have the risk of becoming cancerous. Therefore, polyps are classified using different classification systems. After the classification, further treatment and procedures are based on the classification of the polyp. Nevertheless, classification is not easy. Therefore, we suggest two novel automated classifications system assisting gastroenterologists in classifying polyps based on the NICE and Paris classification.
Methods
We build two classification systems. One is classifying polyps based on their shape (Paris). The other classifies polyps based on their texture and surface patterns (NICE). A two-step process for the Paris classification is introduced: First, detecting and cropping the polyp on the image, and secondly, classifying the polyp based on the cropped area with a transformer network. For the NICE classification, we design a few-shot learning algorithm based on the Deep Metric Learning approach. The algorithm creates an embedding space for polyps, which allows classification from a few examples to account for the data scarcity of NICE annotated images in our database.
Results
For the Paris classification, we achieve an accuracy of 89.35 %, surpassing all papers in the literature and establishing a new state-of-the-art and baseline accuracy for other publications on a public data set. For the NICE classification, we achieve a competitive accuracy of 81.13 % and demonstrate thereby the viability of the few-shot learning paradigm in polyp classification in data-scarce environments. Additionally, we show different ablations of the algorithms. Finally, we further elaborate on the explainability of the system by showing heat maps of the neural network explaining neural activations.
Conclusion
Overall we introduce two polyp classification systems to assist gastroenterologists. We achieve state-of-the-art performance in the Paris classification and demonstrate the viability of the few-shot learning paradigm in the NICE classification, addressing the prevalent data scarcity issues faced in medical machine learning.
During the COVID-19 pandemic, the novel coronavirus had an impact not only on public health but also on the mental health of the population. Public sentiment on mental health and depression is often captured only in small, survey-based studies, while work based on Twitter data often only looks at the period during the pandemic and does not make comparisons with the pre-pandemic situation. We collected tweets that included the hashtags #MentalHealth and #Depression from before and during the pandemic (8.5 months each). We used LDA (Latent Dirichlet Allocation) for topic modeling and LIWC, VADER, and NRC for sentiment analysis. We used three machine-learning classifiers to seek evidence regarding an automatically detectable change in tweets before vs. during the pandemic: (1) based on TF-IDF values, (2) based on the values from the sentiment libraries, (3) based on tweet content (deep-learning BERT classifier). Topic modeling revealed that Twitter users who explicitly used the hashtags #Depression and especially #MentalHealth did so to raise awareness. We observed an overall positive sentiment, and in tough times such as during the COVID-19 pandemic, tweets with #MentalHealth were often associated with gratitude. Among the three classification approaches, the BERT classifier showed the best performance, with an accuracy of 81% for #MentalHealth and 79% for #Depression. Although the data may have come from users familiar with mental health, these findings can help gauge public sentiment on the topic. The combination of (1) sentiment analysis, (2) topic modeling, and (3) tweet classification with machine learning proved useful in gaining comprehensive insight into public sentiment and could be applied to other data sources and topics.
Bioimages frequently exhibit low signal-to-noise ratios due to experimental conditions, specimen characteristics, and imaging trade-offs. Reliable segmentation of such ambiguous images is difficult and laborious. Here we introduce deepflash2, a deep learning-enabled segmentation tool for bioimage analysis. The tool addresses typical challenges that may arise during the training, evaluation, and application of deep learning models on ambiguous data. The tool’s training and evaluation pipeline uses multiple expert annotations and deep model ensembles to achieve accurate results. The application pipeline supports various use-cases for expert annotations and includes a quality assurance mechanism in the form of uncertainty measures. Benchmarked against other tools, deepflash2 offers both high predictive accuracy and efficient computational resource usage. The tool is built upon established deep learning libraries and enables sharing of trained model ensembles with the research community. deepflash2 aims to simplify the integration of deep learning into bioimage analysis projects while improving accuracy and reliability.
Background
Medical resource management can be improved by assessing the likelihood of prolonged length of stay (LOS) for head and neck cancer surgery patients. The objective of this study was to develop predictive models that could be used to determine whether a patient's LOS after cancer surgery falls within the normal range of the cohort.
Methods
We conducted a retrospective analysis of a dataset consisting of 300 consecutive patients who underwent head and neck cancer surgery between 2017 and 2022 at a single university medical center. Prolonged LOS was defined as LOS exceeding the 75th percentile of the cohort. Feature importance analysis was performed to evaluate the most important predictors for prolonged LOS. We then constructed 7 machine learning and deep learning algorithms for the prediction modeling of prolonged LOS.
Results
The algorithms reached accuracy values of 75.40 (radial basis function neural network) to 97.92 (Random Trees) for the training set and 64.90 (multilayer perceptron neural network) to 84.14 (Random Trees) for the testing set. The leading parameters predicting prolonged LOS were operation time, ischemia time, the graft used, the ASA score, the intensive care stay, and the pathological stages. The results revealed that patients who had a higher number of harvested lymph nodes (LN) had a lower probability of recurrence but also a greater LOS. However, patients with prolonged LOS were also at greater risk of recurrence, particularly when fewer (LN) were extracted. Further, LOS was more strongly correlated with the overall number of extracted lymph nodes than with the number of positive lymph nodes or the ratio of positive to overall extracted lymph nodes, indicating that particularly unnecessary lymph node extraction might be associated with prolonged LOS.
Conclusions
The results emphasize the need for a closer follow-up of patients who experience prolonged LOS. Prospective trials are warranted to validate the present results.
Machine learning techniques are excellent to analyze expression data from single cells. These techniques impact all fields ranging from cell annotation and clustering to signature identification. The presented framework evaluates gene selection sets how far they optimally separate defined phenotypes or cell groups. This innovation overcomes the present limitation to objectively and correctly identify a small gene set of high information content regarding separating phenotypes for which corresponding code scripts are provided. The small but meaningful subset of the original genes (or feature space) facilitates human interpretability of the differences of the phenotypes including those found by machine learning results and may even turn correlations between genes and phenotypes into a causal explanation. For the feature selection task, the principal feature analysis is utilized which reduces redundant information while selecting genes that carry the information for separating the phenotypes. In this context, the presented framework shows explainability of unsupervised learning as it reveals cell-type specific signatures. Apart from a Seurat preprocessing tool and the PFA script, the pipeline uses mutual information to balance accuracy and size of the gene set if desired. A validation part to evaluate the gene selection for their information content regarding the separation of the phenotypes is provided as well, binary and multiclass classification of 3 or 4 groups are studied. Results from different single-cell data are presented. In each, only about ten out of more than 30000 genes are identified as carrying the relevant information. The code is provided in a GitHub repository at https://github.com/AC-PHD/Seurat_PFA_pipeline.