Austrian Research Institute for Artificial Intelligence

Abstract Early detection of newly emerging diseases, lesion severity assessment, differentiation of medical conditions and automated screening are examples for the wide applicability and importance of anomaly detection (AD) and unsupervised segmentation in medicine. Normal fine-grained tissue variability such as present in pulmonary anatomy is a major challenge for existing generative AD methods. Here, we propose a novel generative AD approach addressing this issue. It consists of an image-to-image translation for anomaly-free reconstruction and a subsequent patch similarity scoring between observed and generated image-pairs for precise anomaly localization. We validate the new method on chest computed tomography (CT) scans for the detection and segmentation of infectious disease lesions. To assess generalizability, we evaluate the method on an ischemic stroke lesion segmentation task in T1-weighted brain MRI. Results show improved pixel-level anomaly segmentation in both chest CTs and brain MRIs, with relative DICE score improvements of +1.9% and +4.4%, respectively, compared to other state-of-the-art reconstruction-based methods.

Transthyretin cardiac amyloidosis (ATTR-CM) is a progressive, underdiagnosed disease with high morbidity and mortality. While disease-modifying therapies (DMTs) slow progression, early treatment response markers remain scarce. This study assessed AI-quantified thoracic [99mTc]Tc-DPD SPECT/CT markers as potential non-invasive biomarkers for monitoring therapeutic efficacy. This longitudinal study included ATTR-CM patients receiving DMTs (transthyretin stabilizers, RNA interference, or antisense oligonucleotides). [99mTc]Tc-DPD SPECT/CT scans were acquired at baseline and after treatment (median interval 9 months, IQR 7-10). AI-driven segmentation and quantification extracted 26 markers (SUV metrics, retention index, amyloid-affected volume, and amyloid activity). Functional, clinical, and blood parameters, as well as clinical outcomes, were evaluated for their association with changes in imaging markers. In 45 patients (37 ATTRwt-CM, 8 ATTRv-CM), 65% (17/26) of AI-extracted SPECT/CT markers significantly decreased after treatment (all p < 0.001), including SUVmax reductions in the left ventricle (18.6 to 14.1) and myocardium (19.5 to 15.5). None of the markers significantly increased (p > 0.05). Six of the imaging markers, most notably SUVpeak (p = 0.007) of the myocardium and amyloid activity of the left ventricle (p = 0.009), were associated with reductions in NT-proBNP. Lower values for three markers, including amyloid activity of the myocardium, retention index, and SUVmean of the left atrium (all p = 0.016), were associated with improved NYHA class. An increase in amyloid-affected volume of the right ventricle (HR 3.19, 95% CI [1.29; 7.86], p = 0.005) and a decrease in right ventricular SUVmean (adjHR 0.15 95% CI [0.02;1.10], logrank p = 0.030) were associated with death or heart failure-associated hospitalization before and after multivariate adjustment. AI-driven analysis extracted imaging markers substantially faster and eliminated inter-rater variability. AI-driven [99mTc]Tc-DPD SPECT/CT analysis effectively detects treatment-induced reductions in cardiac amyloid burden, offering a non-invasive biomarker for early response assessment in ATTR-CM. AI-enabled imaging markers enhance reproducibility and efficiency, providing valuable support for personalized treatment strategies as new therapeutic options for ATTR-CM become available.

SummaryTrained immunity is a long-lasting innate immune cell phenotype with benefits in infection control and recognized anti-cancer effects. Conversely, inappropriately induced trained immunity contributes to pathological inflammation, warranting the exploration of regulatory pathways. We explore interleukin-38 (IL-38) as a regulator of trained immunity in vivo in a cohort of 325 healthy adults vaccinated with Bacillus Calmette-Guérin (BCG). Using multi-omics profiling, we find that IL-38 is negatively associated with trained immunity on metabolic and epigenetic level. Genetic variants in IL1F10, encoding for IL-38, further link IL-38 to diminished training responses. These associations were validated in human and murine models. We confirmed that IL-38 functionally impairs anti-microbial traits of trained immunity in trained immunity-infection models in vivo (IL-38KO mice) and in vitro (human monocytes). Our study therefore suggests that IL-38 endogenously regulates the induction of trained immunity in humans in vivo.

Abstract Endothelial cells (ECs) are critical regulators of vascular function and exhibit specialized, organ-specific roles across tissues. During aging, these cells become dysfunctional, resulting in increased susceptibility to cardiovascular disease and its associated mortality. While single-cell transcriptomics studies have revealed extensive endothelial heterogeneity across tissues and conditions, a comprehensive atlas of human EC transcriptomes over the course of the adult human lifespan is still lacking. Here, we present the Human Aging Endothelial Cell Atlas (HAECA), a harmonized single-cell transcriptomic compendium of over 375,000 ECs from 12 human tissues throughout adulthood. Using HAECA, we identified age-associated transcriptional shifts, including a decline in angiogenic gene expression in venous ECs and widespread alterations in extracellular matrix (ECM)- and mechanotransduction-associated pathways. We validated these findings in aging human skin and further uncovered a p21-linked transcriptional program in ECs, confirmed in both in vitro and in vivo models and linked to cellular senescence. Together, our study provides a high-resolution transcriptome reference across spatial as well as temporal axes of the human endothelium.

The Ubiquitous Pharmacogenomics consortium (www.upgx.eu) has recently completed and published the Preemptive Pharmacogenomic Testing for Preventing Adverse Drug Reactions (PREPARE) study on the implementation of panel-based pharmacogenetic testing. PREPARE has provided interesting lessons for the design, execution, and interpretation of future clinical implementation studies. In this paper, we share our experience and lessons learned from the PREPARE study for future pharmacogenetic implementation studies. Issues addressed are the study population, intervention, endpoint, randomization, blinding, crossover, ethics, real-world changes during the study, and data analysis and reporting.

Automated, anatomically coherent retinal layer segmentation in optical coherence tomography (OCT) is one of the most important components of retinal disease management. However, current methods rely on large amounts of labeled data, which can be difficult and expensive to obtain. In addition, these systems tend often propose anatomically impossible results, which undermines their clinical reliability. This study introduces a semi-supervised approach to retinal layer segmentation that leverages large amounts of unlabeled data and anatomical prior knowledge related to the structure of the retina. During training, we use a novel topological engine that converts inferred retinal layer boundaries into pixel-wise structured segmentations. These compose a set of anatomically valid disentangled representations which, together with predicted style factors, are used to reconstruct the input image. At training time, the retinal layer boundaries and pixel-wise predictions are both guided by reference annotations, where available, but more importantly by innovatively exploiting anatomical priors that improve the performance, robustness and coherence of the method even if only a small amount of labeled data is available. Exhaustive experiments with respect to label efficiency, contribution of unsupervised data and robustness to different acquisition settings were conducted. The proposed method showed state of-the-art performance on all the studied public and internal datasets, specially in low annotated data regimes. Additionally, the model was able to make use of unlabeled data from a different domain with only a small performance drop in comparison to a fully-supervised setting. A novel, robust, label-efficient retinal layer segmentation method was proposed. The approach has shown state-of-the-art layer segmentation performance with a fraction of the training data available, while at the same time, its robustness against domain shift was also shown.

Chimeric antigen receptor (CAR) T cell therapy has shown remarkable success in treating blood cancers, but CAR T cell dysfunction remains a common cause of treatment failure1. Here we present CELLFIE, a CRISPR screening platform for enhancing CAR T cells across multiple clinical objectives. We performed genome-wide screens in human primary CAR T cells, with readouts capturing key aspects of T cell biology, including proliferation, target cell recognition, activation, apoptosis and fratricide, and exhaustion. Screening hits were prioritized using a new in vivo CROP-seq2 method in a xenograft model of human leukaemia, establishing several gene knockouts that boost CAR T cell efficacy. Most notably, we discovered that RHOG knockout is a potent and unexpected CAR T cell enhancer, both individually and together with FAS knockout, which was validated across multiple in vivo models, CAR designs and sample donors, and in patient-derived cells. Demonstrating the versatility of the CELLFIE platform, we also conducted combinatorial CRISPR screens to identify synergistic gene pairs and saturation base-editing screens to characterize RHOG variants. In summary, we discovered, validated and biologically characterized CRISPR-boosted CAR T cells that outperform standard CAR T cells in widely used benchmarks, establishing a foundational resource for optimizing cell-based immunotherapies.

We present a new methodological approach based on integrating Arden-Syntax-based clinical decision support (CDS) with an upstream ontology service. Incoming linguistic patient data, such as single reports about detected germs or viruses, shall be identified by the applied ontology at a low level. Then, higher-level concepts are activated by ontology-based bottom-up reasoning. Access to these high-level concepts is then provided by Arden-Syntax-based CDS. The results suggest promising directions for future enhancements in knowledge-based artificial intelligence.