Fraunhofer Institute for Digital Medicine

Artificial Intelligence (AI) is increasingly integrated into clinical practice, but its influence on patient decision-making, particularly when AI and physicians disagree, remains unclear. To examine collective advice, we investigated a breast cancer screening scenario using (1) a qualitative interview study (N=9) and (2) a quantitative experiment (N=339) where participants received either consistent or conflicting biopsy recommendations. Qualitative findings include the need for empathetic care, the importance of patient autonomy, and a desire for a four-eyes principle. Quantitative findings accordingly show that patients generally trust physicians more than AI but still tend to follow AI recommendations due to risk aversion. When both advised a biopsy, 99% adhered; if both advised against it, 25% still proceeded. In conflicting scenarios, 97% followed the physician's advice, whereas 66% followed the AI if it recommended the biopsy. These results underscore the need for careful interaction design of collective healthcare advice to prevent unnecessary healthcare procedures.

Magnetic resonance imaging (MRI) is one of the most prevalent imaging modalities used for diagnosis, treatment planning, and outcome control in various medical conditions. MRI sequences provide physicians with the ability to view and monitor tissues at multiple contrasts within a single scan and serve as input for automated systems to perform downstream tasks. However, in clinical practice, there is usually no concise set of identically acquired sequences for a whole group of patients. As a consequence, medical professionals and automated systems both face difficulties due to the lack of complementary information from such missing sequences. This problem is well known in computer vision, particularly in medical image processing tasks such as tumor segmentation, tissue classification, and image generation. With the aim of helping researchers, this literature review examines a significant number of recent approaches that attempt to mitigate these problems. Basic techniques such as early synthesis methods, as well as later approaches that deploy deep learning, such as common latent space models, knowledge distillation networks, mutual information maximization, and generative adversarial networks (GANs) are examined in detail. We investigate the novelty, strengths, and weaknesses of the aforementioned strategies. Moreover, using a case study on the segmentation task, our survey offers quantitative benchmarks to further analyze the effectiveness of these methods for addressing the missing modalities challenge. Furthermore, a discussion offers possible future research directions.

The size of hepatic microwave ablations (MWA) is often difficult to predict due to cooling effects from liver vessels. This study introduces a simplified predictive model, the Ablation Success Ratio (ASR), which estimates the likelihood of a successful ablation based on tumor size and specific ablation parameters. The ASR model is based on the three-dimensional minimum ablation radius (r3Dmin), defining the spherical region within which complete ablation is achieved. To validate the ASR, standardized MWAs were performed in an ex vivo porcine liver model using a glass tube to simulate the vascular cooling effect. Ablations (n = 148) were conducted at 100 W for 5 min, with antenna-to-vessel (A-V) distances set at 2.5, 5.0, and 10.0 mm. Subsequently, the r3Dmin was calculated. Without vascular cooling (0 ml/min, corresponding to an intraoperative Pringle maneuver), an ASR of 100% was achieved for ablation diameters up to 20 mm. However, in the presence of vascular cooling (1–500 ml/min), the ASR reached 100% only for ablation diameters up to 12 mm, demonstrating that the ASR effectively includes the impact of vascular cooling effects. The ASR is a promising and simple approach for predicting ablation success while also accounting for vascular cooling effects in hepatic MWA.

Background: Rupture of the ACL is a common injury among men and women athletes. While planning the surgical ACL reconstruction procedure, the eventual graft’s diameter is extremely important. Many parameters are therefore evaluated pre-surgery to ensure access to reliable data for estimating the graft diameter. Considering this, magnetic resonance imaging (MRI), particularly qualitative analyses of the hamstring tendons, offers a promising approach. Methods: In a retrospective analysis, we carried out 3D segmentation of the gracilis (GT) and semitendinosus tendon (ST) utilizing MRI with varying slice thicknesses and field strengths. The cross-sectional area (CSA) was calculated on different levels (by relying on the models we had thus created) to generate a mean of CSA with six specific segments. We then correlated the mean CSA with the diameter of the graft measured during surgery. Results: A total of 32 patients were included (12 female, 20 male) in this retrospective analysis. We observed the largest CSA in segment 10 mm–0 (16.8 ± 6.1) with differences between men and women. The graft size and tendon diameter correlated significantly in all segments throughout our study cohort. The strongest correlation was apparent in the segment 10 mm–0 (r = 0.552). Conclusions: MRI-based 3D segmentation and the STGT CSA represent a reliable method for estimating preoperatively a quadrupled hamstring graft diameter. The 10 mm–0 mm segment above the joint line showed a strong correlation, making it an ideal reference for graft planning.

Purpose To evaluate the feasibility of interleaved ²³Na/¹H cardiac MRI at 7 T using ¹H parallel transmission (pTx) pulses. Methods A combined setup consisting of a ²³Na volume coil and two ¹H transceiver arrays was employed and the transmit and receive characteristics were compared in vitro with those of the uncombined radiofrequency coils. Furthermore, the implemented interleaved ²³Na/¹H pTx sequence was validated in phantom measurements and applied to four healthy subjects. For the latter, three customized ¹H excitation pulses (universal and individual phase shims (UPS/IPS) and individual 4kT pulses (4kT)) were employed in the interleaved ²³Na/¹H pTx sequence and compared with the vendor‐provided default cardiac phase shim (DPS). Results Combining both coils resulted in a reduction of the mean ²³Na transmit field (B1⁺) efficiency and ²³Na signal‐to‐noise ratio by 18.9% and 15.4% for the combined setup, whereas the ¹H B1⁺ efficiency was less influenced (−4.7%). Compared with single‐nuclear acquisitions, interleaved dual‐nuclear ²³Na/¹H MRI showed negligible influence on ²³Na and ¹H image quality. For all three customized ¹H pTx pulses the B1⁺ homogeneity was improved (coefficients of variation: CVUPS = 0.30, CVIPS = 0.23, CV4kT = 0.15) and no ¹H signal dropouts occurred compared with the vendor‐provided default phase shim (CVDPS = 0.37). Conclusion The incorporation of customized ¹H pTx pulses in an interleaved ²³Na/¹H sequence scheme was successfully demonstrated at 7 T and improvements of the ¹H B1⁺ homogeneity within the heart were shown. Combining interleaved ²³Na/¹H MRI with ¹H pTx is an important tool to enable robust quantification of myocardial tissue sodium concentrations at 7 T within clinically acceptable acquisition times.

Objective. In preclinical research, in vivo imaging of mice and rats is more common than any other animal species, since their physiopathology is very well-known and many genetically altered disease models exist. Animal studies based on small rodents are usually performed using dedicated preclinical imaging systems with high spatial resolution. For studies that require animal models such as mini-pigs or New-Zealand White (NZW) rabbits, imaging systems with larger bore sizes are required. In case of hybrid imaging using positron emission tomography (PET) and magnetic resonance imaging (MRI), clinical systems have to be used, as these animal models do not typically fit in preclinical simultaneous PET-MRI scanners. Approach. In this paper, we present initial imaging results obtained with the Hyperion IID PET insert which can accommodate NZW rabbits when combined with a large volume MRI RF coil. First, we developed a rabbit-sized image quality phantom of comparable size to a NZW rabbit in order to evaluate the PET imaging performance of the insert under high count rates. For this phantom, radioactive spheres with inner diameters between 3.95 and 7.86mm were visible in a warm background with a tracer activity ratio of 4.1 to 1 and with a total ¹⁸F activity in the phantom of 58MBq at measurement start. Second, we performed simultaneous PET-MR imaging of atherosclerotic plaques in a rabbit in vivo using a single injection containing ¹⁸F-FDG for detection of inflammatory activity, and Gd-ESMA for visualization of the aortic vessel wall and plaques with MRI. Main results. The fused PET-MR images reveal ¹⁸F-FDG uptake within an active plaques with plaque thicknesses in the sub-millimeter range. Histology showed colocalization of ¹⁸F-FDG uptake with macrophages in the aortic vessel wall lesions. Significance. Our initial results demonstrate that this PET insert is a promising system for simultaneous high-resolution PET-MR atherosclerotic plaque imaging studies in NZW rabbits.

Background Over the next 5 years, new breast cancer screening guidelines recommending magnetic resonance imaging (MRI) for certain patients will significantly increase the volume of imaging data to be analyzed. While this increase poses challenges for radiologists, artificial intelligence (AI) offers potential solutions to manage this workload. However, the development of AI models is often hindered by manual annotation requirements and strict data-sharing regulations between institutions. Methods In this study, we present an integrated pipeline combining weakly supervised learning—reducing the need for detailed annotations—with local AI model training via swarm learning (SL), which circumvents centralized data sharing. We utilized three datasets comprising 1372 female bilateral breast MRI exams from institutions in three countries: the United States (US), Switzerland, and the United Kingdom (UK) to train models. These models were then validated on two external datasets consisting of 649 bilateral breast MRI exams from Germany and Greece. Results Upon systematically benchmarking various weakly supervised two-dimensional (2D) and three-dimensional (3D) deep learning (DL) methods, we find that the 3D-ResNet-101 demonstrates superior performance. By implementing a real-world SL setup across three international centers, we observe that these collaboratively trained models outperform those trained locally. Even with a smaller dataset, we demonstrate the practical feasibility of deploying SL internationally with on-site data processing, addressing challenges such as data privacy and annotation variability. Conclusions Combining weakly supervised learning with SL enhances inter-institutional collaboration, improving the utility of distributed datasets for medical AI training without requiring detailed annotations or centralized data sharing.

Purpose Federated training is often challenging on heterogeneous datasets due to divergent data storage options, inconsistent naming schemes, varied annotation procedures, and disparities in label quality. This is particularly evident in the emerging multi-modal learning paradigms, where dataset harmonization including a uniform data representation and filtering options are of paramount importance. Methods DICOM-structured reports enable the standardized linkage of arbitrary information beyond the imaging domain and can be used within Python deep learning pipelines with . Building on this, we developed an open platform for data integration with interactive filtering capabilities, thereby simplifying the process of creation of patient cohorts over several sites with consistent multi-modal data. Results In this study, we extend our prior work by showing its applicability to more and divergent data types, as well as streamlining datasets for federated training within an established consortium of eight university hospitals in Germany. We prove its concurrent filtering ability by creating harmonized multi-modal datasets across all locations for predicting the outcome after minimally invasive heart valve replacement. The data include imaging and waveform data (i.e., computed tomography images, electrocardiography scans) as well as annotations (i.e., calcification segmentations, and pointsets), and metadata (i.e., prostheses and pacemaker dependency). Conclusion Structured reports bridge the traditional gap between imaging systems and information systems. Utilizing the inherent DICOM reference system arbitrary data types can be queried concurrently to create meaningful cohorts for multi-centric data analysis. The graphical interface as well as example structured report templates are available at https://github.com/Cardio-AI/fl-multi-modal-dataset-creation .

Objectives Caffeine, a known neurostimulant and adenosine antagonist, affects brain physiology by decreasing cerebral blood flow. It interacts with adenosine receptors to induce vasoconstriction, potentially disrupting brain homeostasis. However, the impact of caffeine on blood–brain barrier (BBB) permeability to water remains underexplored. This study investigated the water exchange via the BBB in a perturbed physiological condition caused by caffeine ingestion, using the multiple echo time (multi-TE) arterial spin labeling (ASL) technique. Material and methods Ten healthy, regular coffee drinkers (age = 31 ± 9 years, 3 females) were scanned to acquire five measurements before and six measurements after caffeine ingestion. Data were analyzed with a multi-TE two-compartment model to estimate exchange time (Tex), serving as a proxy for BBB permeability to water. Additionally, cerebral blood flow (CBF), arterial transit time (ATT), and intravoxel transit time (ITT) were investigated. Results Following caffeine intake, mean gray matter CBF showed a significant time-dependent decrease ( P < 0.01). In contrast, Tex, ATT, and ITT did not exhibit significant time-dependent change. However, a non-significant decreasing trend was observed for Tex and ITT, respectively, while ATT showed an increasing trend over time. Discussion The observed decreasing trend in Tex after caffeine ingestion suggests a potential increase in water flux across the BBB, which may represent a compensatory mechanism to maintain brain homeostasis in response to the caffeine-induced reduction in CBF. Further studies with larger sample sizes are needed to validate and expand upon these findings.

Background: Blood-brain barrier (BBB) dysfunction is suggested to be a potential mediator between vascular risk factors and cognitive impairment, leading to vascular cognitive impairment. Objective: To investigate the relationships between age, sex, and vascular risk factors and BBB water permeability as well as their relationship with cognition. Methods: To measure BBB permeability, a novel arterial spin labelling MRI technique (ME-ASL) was applied to derive the time of exchange (Tex), arterial time transit (ATT), and cerebral blood flow (CBF). The association of potential risk factors, such as age, sex, body mass index (BMI), blood pressure (BP), and medical history, with these BBB parameters were assessed in 144 community-dwelling adults (median age 59 years, 57% females). The relationship between BBB permeability and cognitive performance measured by the Montreal Cognitive Assessment (MoCA) was also assessed. Results: We found that increased BMI was significantly associated with decreased CBF (β = −0.06). Systolic BP and diastolic BP showed significant associations with all ASL parameters; systolic BP was negatively correlated with Tex (β = −0.02) and CBF (β = −0.01) but positively with ATT (β = 0.02). Diastolic BP was negatively associated with Tex (β = −0.03) and CBF (β = −0.03) but positively with ATT (β = 0.03). MoCA scores had a borderline significant association with Tex (OR = 1.51) and a significant association with CBF (OR = 1.84), which became non-significant after adjusting for confounders. Conclusions: These outcomes underscore the potential of using ME-ASL, warranting further research to strengthen these findings.

Purpose In brain tumors, disruption of the blood–brain barrier (BBB) indicates malignancy. Clinical assessment is qualitative; quantitative evaluation is feasible using the K2 leakage parameter from dynamic susceptibility contrast MRI. However, contrast agent–based techniques are limited in patients with renal dysfunction and insensitive to subtle impairments. Assessing water transport times across the BBB (Tex) by multi‐echo arterial spin labeling promises to detect BBB impairments noninvasively and potentially more sensitively. We hypothesized that reduced Tex indicates impaired BBB. Furthermore, we assumed higher sensitivity for Tex than dynamic susceptibility contrast–based K2, because arterial spin labeling uses water as a freely diffusible tracer. Methods We acquired 3T MRI data from 28 patients with intraparenchymal brain tumors (World Health Organization Grade 3 & 4 gliomas [n = 17] or metastases [n = 11]) and 17 age‐matched healthy controls. The protocol included multi‐echo and single‐echo Hadamard‐encoded arterial spin labeling, dynamic susceptibility contrast, and conventional clinical imaging. Tex was calculated using a T2‐dependent multi‐compartment model. Areas of contrast‐enhancing tissue, edema, and normal‐appearing tissue were automatically segmented, and parameter values were compared across volumes of interest and between patients and healthy controls. Results Tex was significantly reduced (−20.3%) in contrast‐enhancing tissue compared with normal‐appearing gray matter and correlated well with |K2| (r = −0.347). Compared with healthy controls, Tex was significantly lower in tumor patients' normal‐appearing gray matter (Tex,tumor = 0.141 ± 0.032 s vs. Tex,HC = 0.172 ± 0.036 s) and normal‐appearing white matter (Tex,tumor = 0.116 ± 0.015 vs. Tex,HC = 0.127 ± 0.017 s), whereas |K2| did not differ significantly. Receiver operating characteristic analysis showed a larger area under the curve for Tex (0.784) than K2 (0.604). Conclusion Tex is sensitive to pathophysiologically impaired BBB. It agrees with contrast agent–based K2 in contrast‐enhancing tissue and indicates sensitivity to subtle leakage.

Objective The German Health Data Lab is going to provide access to German statutory health insurance claims data ranging from 2009 to the present for research purposes. Due to evolving data formats within the German Health Data Lab, there is a need to standardize this data into a Common Data Model to facilitate collaborative health research and minimize the need for researchers to adapt to multiple data formats. For this purpose we selected transforming the data to the Observational Medical Outcomes Partnership Common Data Model. Methods We developed an Extract, Transform, and Load (ETL) pipeline for two distinct German Health Data Lab data formats: Format 1 (2009-2016) and Format 3 (2019 onwards). Due to the identical format structure of Format 1 and Format 2 (2017 -2018), the ETL pipeline of Format 1 can be applied on Format 2 as well. Our ETL process, supported by Observational Health Data Sciences and Informatics tools, includes specification development, SQL skeleton creation, and concept mapping. We detail the process characteristics and present a quality assessment that includes field coverage and concept mapping accuracy using example data. Results For Format 1, we achieved a field coverage of 92.7%. The Data Quality Dashboard showed 100.0% conformance and 80.6% completeness, although plausibility checks were disabled. The mapping coverage for the Condition domain was low at 18.3% due to invalid codes and missing mappings in the provided example data. For Format 3, the field coverage was 86.2%, with Data Quality Dashboard reporting 99.3% conformance and 75.9% completeness. The Procedure domain had very low mapping coverage (2.2%) due to the use of mocked data and unmapped local concepts The Condition domain results with 99.8% of unique codes mapped. The absence of real data limits the comprehensive assessment of quality. Conclusion The ETL process effectively transforms the data with high field coverage and conformance. It simplifies data utilization for German Health Data Lab users and enhances the use of OHDSI analysis tools. This initiative represents a significant step towards facilitating cross-border research in Europe by providing publicly available, standardized ETL processes (https://github.com/FraunhoferMEVIS/ETLfromHDLtoOMOP) and evaluations of their performance.