Stroke remains a leading cause of morbidity and mortality worldwide, despite advances in treatment modalities. Endovascular thrombectomy (EVT), a revolutionary intervention for ischemic stroke, is limited by its reliance on 2D fluoroscopic imaging, which lacks depth and comprehensive vascular detail. We propose a novel AI-driven pipeline for 3D CTA to 2D DSA cross-modality registration, termed DeepIterReg. The proposed pipeline integrates neural network-based initialization with iterative optimization to align pre-intervention and peri-intervention data. Our approach addresses the challenges of cross-modality alignment, particularly in scenarios involving limited shared vascular structures, by leveraging synthetic data, vein-centric anchoring, and differentiable rendering techniques. We assess the efficacy of DeepIterReg through quantitative analysis of capture ranges and registration accuracy. Results indicate that our method is able to accurately register 70% of a testset of 20 patients, and is able to improve capture ranges when performing an initial pose estimation using a convolutional neural network.

Language models (LLMs) have demonstrated impressive performance on knowledge-intensive tasks like question answering when supported by external knowledge. However, their success relies not only on their reasoning capabilities and the accuracy of the external knowledge but also on the truthfulness of the prompts provided. False premises in prompts can lead to "hallucinations," where the generated content appears plausible but is factually incorrect. This issue is common in online questions, particularly when users search for information on unfamiliar topics, leading to confirmation bias in information retrieval.

Existing methods for detecting hallucinations may not effectively handle false premises, as they can be misled by coherent responses that align with the false premises. Fact-checking methods may also be unsuitable, as LLMs can exhibit sycophantic behavior in attempting to satisfy user requirements.

To address this challenge, we propose a False Premise Detection with Abductive Reasoning FPDAR method for question answering with LLMs. Abductive reasoning enables backward thinking, minimizing less plausible assumptions and working towards the correct answer through a bottom-up approach. FPDAR is designed as a plug-and-play module that can be integrated after the question-answering process.

FPDAR employs a two-stage abductive reasoning process. First, it infers the most plausible question intent based on factual context and generated response without considering the potentially problematic question as input. This allows for the identification of false premises by comparing the inferred intent with the original question. Second, abductive reasoning helps generate a more plausible explanation aligned with factual context, increasing the likelihood of correctness and ruling out less plausible, potentially hallucinated responses.

To the best of our knowledge, this is the first study introducing abductive reasoning for identifying and diagnosing false premises. We conduct extensive experiments on two question-answering benchmarks containing false premises to validate the effectiveness of FPDAR. The results show that FPDAR can achieve high accuracy in terms of response correctness, although it may struggle to effectively detect false premises. Nevertheless, it achieves substantial accuracy improvements over state-of-the-art methods.

Alzheimer's disease (AD) is becoming more prevalent as the world population gets older. The formation of Amyloid-beta (\AB) plaques is one of the pathologies related to AD. Recent work has shown that the \ab load in brain tissue has a negative correlation with cognitive performance in cognitively healthy centenarians.
This work aims to expand this research by investigating whether the types of \ab plaque present are linked to cognition and by comparing the types of plaques in the centenarian cohort with an AD cohort.
For this task, a system is developed that can identify \ab plaques in images of brain tissue. It first automatically segments the grey matter using a fine-tuned U-net. Then the plaques are located using traditional image processing techniques. Lastly, shape and size features are extracted from the plaque in addition to a feature vector made by a pre-trained AlexNet. K-means clustering is used on AlexNet features to find categories for the plaques.
The clustering approach failed to yield good results. However, the area and roundness are differently distributed between the AD and centenarian cohorts, but the differences are small. Correlations have been found between the area and roundness of plaques in the occipital pole and cognitive performance in centenarians. They indicate that cognitively stronger individuals have smaller and less round \ab plaques in their brains. More research is necessary to reveal the true extent of the impact of plaque types on cognition.

Celiac disease is a genetic autoimmune disorder caused by a negative reaction to gluten associated with alterations in the gut microbiome. This study explored the potential of machine learning models and feature selection methods in identifying biomarkers for celiac disease using gut microbiome data. The performance of several machine learning models was evaluated, and the impact of different feature selection methods, including MRMR, ANOVA, and information gain, was examined. The findings revealed comparable performance among the models without feature selection. However, the choice of feature selection method had varying effects on model performance, with logistic regression and support vector machines being more sensitive than random forest and XGBoost models. Notably, several identified bacteria species, such as Bacteroides eggerthii, Parabacteroides johnsonii, Faecalibacterium prausnitzii, and Ruminococcus_D bicirculans, have been previously associated with celiac disease, reinforcing their potential as biomarkers.

Type 2 Diabetes is a very prevalent disease in current times and leads to significant adverse effects. Recently, there has been a growing interest in the association of the human gut microbiome with respect to chronic diseases like Type 2 Diabetes with the aim to identify biomarkers. In this study, we researched the effect of different machine learning and feature selection techniques to identify biomarkers for Type 2 Diabetes that can later be used for diagnosis and prediction. The main methods that we explored were Random Forests,Linear Regression, Support Vector Machines andXGBoost along with mRMR and CMIM as feature selection techniques. These methods were applied to data taken from Europe and China. We found that mRMR improved the performance of the Random Forest classifier compared to CMIM.Apart from finding biomarkers specific to one location, we found that Clostridiales, Clostridium, Roseburia and Lactobacillus could be of interestin the prediction of Type 2 Diabetes irrespective of location. This study verified biomarkers found in previous literature and evaluated several techniquesfor the prediction of the disease across different regions.

Colorectal cancer (CRC), one of the leading causes of mortality, is challenging to diagnose. By using metagenomic analysis with machine learning methods, this can be done in a non-invasive manner. In this research, a neural network has been trained on relative pathway abundance data, a way to measure the functional potential of a microbiome, in order to find biomarkers for colorectal cancer. The accuracy achieved by the neural network is 57%. The most important features used by the model are compared to established biomarkers in literature. Besides overlapping pathways, this research also found new potential biomarkers for CRC.

Parkinson's disease (PD) is a neurodegenerative disorder characterized by motor function loss and potential mental and behavioral changes. The identification of biomarkers in the gut microbiota of PD patients can significantly aid in fast and accurate diagnosis. This study investigates the application of machine learning (ML) models, including Logistic Regression (LR), Random Forest (RF), and Support Vector Machines (SVM), to discover biomarkers in the gut metagenomic data of PD patients. The ML models were optimized using various feature selection techniques, and a comparative analysis of the most influential species in sample discrimination was conducted to verify potential PD-associated biomarkers.
The results demonstrate that all three ML models exhibit moderate performance, indicating their limited discriminatory power. However, the comparison of significant species across different classifiers demonstrates substantial overlap and indicates PD-associated species that align with existing literature findings. These outcomes provide promising evidence that LR, RF, and SVM classifiers can effectively identify biomarkers for PD. However, confounding analysis on a small subset of the dataset failed to identify meaningful PD-associated species. Therefore, caution is advised when interpreting the findings of ML model, considering factors such as classifier performance, dataset limitations, potential biases, influence of feature selection methods, and inherent model differences.
We validate the potential usefulness of ML approaches for biomarker discovery and highlight areas for further investigation into building a sufficiently accurate ML model for PD diagnosis.

There is mounting evidence indicating a relation- ship between the gut microbiome composition and the development of mental diseases but the mech- anisms remain unclear. Shotgun sequenced data from 90 schizophrenic patients and 81 sex, age, weight, and location matched controls was used for three machine learning models: Logistic Re- gression, Random Forests, and XGBoost. The 20 most relevant species in the decision mak- ing of each classifier was retained and the over- lap between models recorded. There is a total 19 overlapping species between the models’ top 20 most relevant species, with 10 species over- lapping on all three models. Bifidobacterium bi- fidum, Akkermansia muciniphila, Eubacterium sir- aeum, Alistipes finegoldii, Intestinibacter bartlet- tii, Bifidobacterium pseudocatenulatum, and Strep- tococcus thermophilus are of particular interest as they are reported as enriched in schizophrenia sam- ples in existing literatures. Phoceicola vulgatus has been found to play a significant role in the classi- fiers decisions and is enriched in healthy samples in the literature. One species, Ruthenibacterium lactatiformans, and one co-abundant gene group, Eubacterium sp. CAG:180, consistently ranked as the most important features across all three classi- fiers, despite the absence of reporting in existing literature. This study could be expanded by using genus-level data. Further research should be done to validate the species mentioned above as potential biomarkers for schizophrenia.

In modern neurosurgical practice, a surgeon can see a patient’s fiber tracts (nerve tracts) on a monitor in the operating room. This design study investigates the benefit of adding the uncertainty of the tracts and aims to improve the surgeon’s orientation while reducing visual clutter.

Based on an interview with a neurosurgeon and our experience during a surgery visit, we designed, implemented, and evaluated a new visualization method for intraoperative use. We present two new visualization techniques: a partial hull that appears when a surgeon approaches it and a distance meter that serves as a warning tool.

We conclude that there is a need for the intraoperative visualization of fiber uncertainty and recommend doing clinical trials using these methods.

Single-cell sequencing allows measuring individual cells' molecular features and their responses to perturbations. Understanding which cells respond to a particular perturbation and how these responses vary across populations can be used to, for example, improve vaccine immunogenicity. However, an exhaustive exploration of single-cell perturbation responses in every population is usually experimentally unfeasible. Several machine learning models have been developed to predict perturbation responses, but they are limited to single-modality data. Single-modality data alone, such as only transcriptomics, is not suited to capture all cell responses accurately. For example, the identification of immune responses requires transcriptomic and proteomic measurements. Here, we introduce cellPMVI, a method built to predict perturbation responses from multi-modality data. cellPMVI combines the single-cell data modeling from scVI with a mixture-of-experts posterior integration to allow for multi-modality input data. In this work, we validate cellPMVI for immune response prediction of adjuvants across populations. The model is trained on two-modality CITE-seq data containing gene and protein measurements from three different populations. We show that cellPMVI can model both modalities of the CITE-seq data without information loss in either modality and predict immune responses with a high correlation to the observed responses across different populations. Hence, cellPMVI is the first model to capture and predict immune response for multi-modality data with the potential to be applied for other perturbations, such as drugs.

Color information has been shown to provide useful information during image classification. Yet current popular deep convolutional neural networks use 2-dimensional convolutional layers. The first 2-dimensional convolutional layer in the network combines the color channels of the input images, which produces feature maps per channel with only spatial dimensions, height and width, getting rid of the color dimension. In this work we introduce Full Color Deep networks which use 3-dimensional convolutions to retain the color dimension beyond the first layer. The 3D kernels convolve over the color and spatial dimensions. The network can extract features from all three dimensions in all layers which are subsequently used by the classifier. We show that the Full Color Deep networks perform at least as well as the current CNNs but outperform them in learning color information and using that information in other downstream tasks.

Out-of-Domain (OOD) generalization is a challenging problem in machine learning about learning a model from one or more domains and making the model perform well on an unseen domain. Empirical Risk Minimization (ERM), the standard machine learning method, suffers from learning spurious correlation in the training domain, therefore may perform badly when the unseen domain has different distribution from the training domain. Group Distributionally Robust Optimization (group DRO) is a method proposed to handle the OOD generalization problem. In this paper, the goals are to 1) measure if group DRO has a better OOD generalization performance than ERM. 2) evaluate if group DRO finds causally invariant relationships between the input and output. Semi-synthetic bird images with different backgrounds are used to form our data sets to construct a binary image classification problem for experiments. Results show that group DRO improves OOD generalization performance over ERM, and group DRO can find invariant relationships. However, the ability of group DRO to find invariant relationships is limited when the spurious correlation in the training domain is strong.

Out-of-domain (OOD) generalization refers to learning a model from one or more different but related domain(s) that can be used in an unknown test domain. It is challenging for existing machine learning models. Several methods have been proposed to solve this problem, and multi-domain calibration is one of these methods. Model selection with the average expected calibration error (ECE) across training domains and naive calibration are two approaches to implementing multi-domain calibration. However, it might happen that neither approach can learn a genuinely well-calibrated model in the multi-domain setting. Hence, this paper intends to evaluate how naive calibration and model selection with average ECE perform in the OOD generalization problem for binary classifiers. We generated many synthetic datasets and set up three experiments to answer this question. Finally, the conclusions based on empirical results are obtained: 1) Although naive calibration can improve the average accuracy across unseen domains (OOD accuracy) and the average area under the ROC Curve across unseen domains (OOD AUROC) for some binary classifiers, it does not work for all binary classifiers. However, at least it does not make the model worse for OOD generalization. 2) On the synthetic datasets we generated, if the number of training domains increases, most binary classifiers' OOD accuracy will also increase. 3) Average ECE is a reasonable metric for selecting a model in the OOD generalization problem and is better than validation accuracy. This is because a strong linear relationship exists between OOD accuracy and the average ECE across the training domains. This linear relationship is stronger than the linear relationship between OOD accuracy and validation accuracy.

Generalizing models for new unknown datasets is a common problem in machine learning. Algorithms that perform well for test instances with the same distribution as their training dataset often perform severely on new datasets with a different distribution. This problem is caused by distributional shifts between the training of the model and applying that model to a test domain. This paper addresses whether and in what situations Risk Extrapolation (REx) can tackle this problem of Out-Of-Distribution generalization by exploiting invariant relationships. These relationships are based on features that are invariant across all domains. By learning these relationships, REx aims to learn the concept of the problem we are trying to solve. We show in what situations REx can learn these invariant relationships and when it does not. We translate the definition of an invariant relationship into a homoscedastic synthetic dataset with either covariate, confounded, anti-causal, or hybrid shift. We expose REx to experiments in sample complexity, the number of training domains, and the training domain distance. We show that REx performs better for invariant prediction in situations with larger sample sizes and training domain distance and that if these criteria are met, REx performs equivalently in all four distributional shifts. We also compare REx to Invariant- and Empirical Risk Minimization and show that; REx is less sensitive and thus robust to the shifting of the average distributional variance in the training domains; REx asymptotically out-performs the methods in the more complex distributional shifts.

Software testing is essential for a successful development process, however, it can be troublesome as manually writing tests can be time demanding and error-prone. EvoSuite is a test case generating tool developed to address this [18]. It can generate test cases for different test criteria - Line Coverage, Branch Coverage, Input Diversity, etc. Branch Coverage puts the focus on covering branches in the code, whilst Input Diversity puts the focus on the use of diverse inputs as parameters in the test cases. The downside is that the user needs to select the best suited test criteria, out of the many that EvoSuite provides, for the classes under test. It is not feasible for the user to manually find the optimal one for the classes under test. This paper aims to shine a light on the effectiveness of the combination of Input Diversity and Branch Coverage as a test criteria. This study presents a machine learning technique to automatically select the best combination of test generation objectives according to static metrics. The model we chose for this task is a decision tree as it directly provides a pattern. Said pattern is a combination of conditions that the static metrics need to hold for the chosen test criteria to be effective. The evaluation of the effectiveness was done one a benchmark of 346 classes taken from SF-110 Corpus of Classes [9] and the Appache Commons. To evaluate the effectiveness of Input Diversity in combination with Branch Coverage, we compared the test criteria to two other test criteria - Branch Coverage and the Default coverage criteria used in EvoSuite. The decision tree models created achieve an accuracy upwards of 90% in the best case and deem metrics such as wmc, dit, fanin and others to be crucial for the effectiveness of Input Diversity in combination with Branch Coverage.

Recently, automating test suite generation is a problem that has drown attention in both industry and academia. One of the tools used to automatically generate test suites is EvoSuite, which is a state-of-the-art tool often used in research. It uses a genetic algorithm, which seeks to maximize certain coverage criteria, such as Branch Coverage or Exception Coverage. Previous research has investigated the possibility of combining multiple coverage criteria, but there is no single combination which performs best for all classes. This paper aims to investigate if it possible to predict when coverage criteria perform best, specifically Exception Coverage, according to the characteristics of the class under test. The paper shows that there is a significant difference between the performances of these coverage criteria. The paper also shows that this difference can be predicted using a Machine Learning model with an f1-score of 0,865, when performance was measured using Mutation Score. Lastly, an exploration is made into the characteristics of the Class-under-test which cause this.

The perpetual desire for more qualitative software has been an excellent incentive for software engineers to create automated tools to ease and improve the process of software testing. EvoSuite is an example of a state-of-the-art tool that synthesises test cases automatically. It uses a genetic algorithm to produce test cases based on given search targets. Previous studies have analysed the performance of single or combinations of targets but have not yet explored the differences between various combinations. In this research, we compare the Weak Mutation + Branch setting to Branch and the Default (combination of eight separate targets) of EvoSuite. We aim to provide insightful information about their differences in branch coverage and mutation scores. Moreover, we discuss machine-learning models that can predict which combination has the highest score (i.e., branch coverage, mutation score) based on characteristics of the tested classes, such as the number of lines of code. Our results highlight that the Weak Mutation + Branch combination outperforms Branch for the mutation score metric and Default for the branch coverage metric. They also show that Weak Mutation + Branch is outperformed by the branch criterion for Branch Coverage and by the Default combination for mutation score. Our findings also cover the performance of the models, having concluded that the Random Forest and Decision Tree classifiers produce the best results and are feasible options for predicting the best combinations from the ones analysed. Finally, static code metrics such as 'wmc', 'loc', and 'mathOperationsOty' often appear as relevant features for our models. We visualise how they influence the most suitable combination of criteria through our Decision Trees.

To ensure that a software system operates in the correct way, it is crucial to test it extensively. Manual software testing is severely time-consuming, and developers often underestimate its importance. Consequently, many tools for automatic test generation have been developed during the past decade. EvoSuite is a state-of-the-art tool for automatic generation of unit tests. It can produce test suites based on chosen coverage criteria, also known as a fitness function. Previous studies have widely assessed the performance of the different fitness functions available in EvoSuite. However, the combination of various coverage criteria has not been considered. In this paper, we assess the effectiveness of the combination of Branch coverage and Output diversity fitness functions. We compare it to two of the most popular fitness functions in EvoSuite - Branch coverage and the Default configuration (combines eight coverage criteria) to estimate its performance. We developed a machine learning tool that determines which fitness function will achieve better results based on class characteristics. The assessment criteria we consider are branch coverage and fault detection, represented by mutation score. We further examined how the time limit affects the performance of the considered fitness functions. The results have shown that the combination of Branch coverage and Output diversity outperforms the Default configuration significantly in branch coverage but has worse performance in fault detection capabilities. We have also found that the Branch and Output diversity coverage criteria achieve better results when compared with only using Branch coverage in terms of mutation score. Additionally, the static software metrics, especially CBO, LCOM* and LOC, are highly correlated with the performance of the fitness functions and can predict which coverage criteria will achieve better results.