FedStenoNet: tackling domain shift in x-ray coronary angiography through a personalized federated detection framework.

Ultrasound is considered a key modality for the clinical assessment of hepatic steatosis (i.e., fatty liver) due to its non-invasiveness and availability. Deep learning methods have attracted considerable interest in this field, as they are capable of learning patterns in a collection of images and achieve clinically comparable levels of accuracy in steatosis grading. However, variations in patient populations, acquisition protocols, equipment, and operator expertise across clinical sites can introduce domain shifts that reduce model performance when applied outside the original training setting. In response, unsupervised domain adaptation techniques are being investigated to address these shifts, allowing models to generalize more effectively across diverse clinical environments. In this work, we propose a test-time batch normalization technique designed to handle domain shift, especially for changes in label distribution, by adapting selected features of batch normalization layers in a trained convolutional neural network model. This approach operates in an unsupervised manner, allowing robust adaptation to new distributions without access to label data. The method was evaluated on two abdominal ultrasound datasets collected at different institutions, assessing its capability in mitigating domain shift for hepatic steatosis classification. The proposed method reduced the mean absolute error in steatosis grading by 37% and improved the area under the receiver operating characteristic curve for steatosis detection from 0.78 to 0.97, compared to non-adapted models. These findings demonstrate the potential of the proposed method to address domain shift in ultrasound-based hepatic steatosis diagnosis, minimizing risks associated with deploying trained models in various clinical settings.

Neurocritical care is devoted to the care of critically ill patients with acute neurological or neurosurgical emergencies. There is limited information regarding epidemiological data, disease characteristics, variability of clinical care, and in-hospital mortality of neurocritically ill patients worldwide. We addressed these issues in the Point PRevalence In Neurocritical CarE (PRINCE) study, a prospective, cross-sectional, observational study. We recruited patients from various intensive care units (ICUs) admitted on a pre-specified date, and the investigators recorded specific clinical care activities they performed on the subjects during their first 7days of admission or discharge (whichever came first) from their ICUs and at hospital discharge. In this manuscript, we analyzed the final data set of the study that included patient admission characteristics, disease type and severity, ICU resources, ICU and hospital length of stay, and in-hospital mortality. We present descriptive statistics to summarize data from the case report form. We tested differences between geographically grouped data using parametric and nonparametric testing as appropriate. We used a multivariable logistic regression model to evaluate factors associated with in-hospital mortality. We analyzed data from 1545 patients admitted to 147 participating sites from 31 countries of which most were from North America (69%, N = 1063). Globally, there was variability in patient characteristics, admission diagnosis, ICU treatment team and resource allocation, and in-hospital mortality. Seventy-three percent of the participating centers were academic, and the most common admitting diagnosis was subarachnoid hemorrhage (13%). The majority of patients were male (59%), a half of whom had at least two comorbidities, and median Glasgow Coma Scale (GCS) of 13. Factors associated with in-hospital mortality included age (OR 1.03; 95% CI, 1.02 to 1.04); lower GCS (OR 1.20; 95% CI, 1.14 to 1.16 for every point reduction in GCS); pupillary reactivity (OR 1.8; 95% CI, 1.09 to 3.23 for bilateral unreactive pupils); admission source (emergency room versus direct admission [OR 2.2; 95% CI, 1.3 to 3.75]; admission from a general ward versus direct admission [OR 5.85; 95% CI, 2.75 to 12.45; and admission from another ICU versus direct admission [OR 3.34; 95% CI, 1.27 to 8.8]); and the absence of a dedicated neurocritical care unit (NCCU) (OR 1.7; 95% CI, 1.04 to 2.47). PRINCE is the first study to evaluate care patterns of neurocritical patients worldwide. The data suggest that there is a wide variability in clinical care resources and patient characteristics. Neurological severity of illness and the absence of a dedicated NCCU are independent predictors of in-patient mortality.

The BraTS Generalizability Across Tumors (BraTS-GoAT) initiative addresses the critical need for robust and generalizable models in brain tumor segmentation. Despite advancements in automated segmentation techniques, the variability in tumor characteristics and imaging modalities across clinical settings presents a significant challenge. This study aims to develop an advanced CNN-based model for brain tumor segmentation that enhances consistency and utility across diverse clinical environments. The objective is to improve the generalizability of CNN models by applying them to large-scale datasets and integrating robust preprocessing techniques. The proposed approach involves the application of advanced CNN models to the BraTS 2024 challenge dataset, incorporating preprocessing techniques such as standardization, feature extraction, and segmentation. The model's performance was evaluated based on accuracy, mean Intersection over Union (IOU), average Dice coefficient, Hausdorff 95 score, precision, sensitivity, and specificity. The model achieved an accuracy of 98.47%, a mean IOU of 0.8185, an average Dice coefficient of 0.7, an average Hausdorff 95 score of 1.66, a precision of 98.55%, a sensitivity of 98.40%, and a specificity of 99.52%. These results demonstrate a significant improvement over the current gold standard in brain tumor segmentation. The findings of this study contribute to establishing benchmarks for generalizability in medical imaging, promoting the adoption of CNN-based brain tumor segmentation models in diverse clinical environments. This work has the potential to improve outcomes for patients with brain tumors by enhancing the reliability and effectiveness of automated segmentation techniques.

BackgroundMRI sequence classification becomes challenging in multicenter studies due to variability in imaging protocols, leading to unreliable metadata and requiring labor-intensive manual annotation. While numerous automated MRI sequence identification models are available, they frequently encounter the issue of domain shift, which detrimentally impacts their accuracy. This study addresses domain shift, particularly from adult to pediatric MRI data, by evaluating the effectiveness of pre-trained models under these conditions.MethodsThis retrospective and multicentric study explored the efficiency of a pre-trained convolutional (ResNet) and CNN-Transformer hybrid model (MedViT) to handle domain shift. The study involved training ResNet-18 and MedVit models on an adult MRI dataset and testing them on a pediatric dataset, with expert domain knowledge adjustments applied to account for differences in sequence types.ResultsThe MedViT model demonstrated superior performance compared to ResNet-18 and benchmark models, achieving an accuracy of 0.893 (95% CI 0.880–0.904). Expert domain knowledge adjustments further improved the MedViT model’s accuracy to 0.905 (95% CI 0.893–0.916), showcasing its robustness in handling domain shift.ConclusionAdvanced neural network architectures like MedViT and expert domain knowledge on the target dataset significantly enhance the performance of MRI sequence classification models under domain shift conditions. By combining the strengths of CNNs and transformers, hybrid architectures offer enhanced robustness for reliable automated MRI sequence classification in diverse research and clinical settings.Key PointsQuestionDomain shift between adult and pediatric MRI data limits deep learning model accuracy, requiring solutions for reliable sequence classification across diverse patient populations.FindingsThe MedViT model outperformed ResNet-18 in pediatric imaging; expert domain knowledge adjustment further improved accuracy, demonstrating robustness across diverse datasets.Clinical relevanceThis study enhances MRI sequence classification by leveraging advanced neural networks and expert domain knowledge to mitigate domain shift, boosting diagnostic precision and efficiency across diverse patient populations in multicenter environments.Graphical

Background Diagnosis of coronary ischaemia by perfusion cardiovascular magnetic resonance (CMR) has high sensitivity and specificity when using X-ray coronary angiography as the reference standard. Potential reasons for false negative perfusion CMR studies include suboptimal image quality, technical reasons, or the potential discrepancy between angiographic stenosis and detectable myocardial hypoperfusion. The rates at which these factors occur have not been specifically studied to date. The CE-MARC study prospectively enrolled 752 patients with suspected coronary artery disease, scheduled to undergo CMR, SPECT and X-ray coronary angiography. We assessed potential reasons for the false negative CMR perfusion studies within CE-MARC. Table 1 Patient characteristics for false negative perfusion CMR in CE-MARC. False negative patients CE-MARC whole population N 35 752 Age (years) 61±7 60±10 Male 29 (83%) 471 (63%) Body-mass index (kg/m2) 28.3±4.0 29.2±4.4 Resting BP (mm Hg) 125/71±20/10 138/79±21/11 LAD disease 19 (54%) 183 (25%) Circumflex disease 16 (46%) 133 (18%) RCA disease 8 (23%) 110 (15%) Left main disease 2 (6%) 23 (3%) Data as n (%) or mean±SD. Methods All patients with significant coronary stenosis (≥70% stenosis of a first order coronary artery ≥2 mm diameter, or left main stem stenosis ≥50% as measured by quantitative coronary angiography (QCA)), who had a normal or probably normal CMR perfusion analysis from the original, blinded read were selected from the CE-MARC population. Patient and imaging characteristics were analysed. Myocardial perfusion reserve (MPR) was calculated offline (PMI v0.4) from CMR stress and rest perfusion images using the Fermi model, with arterial input defined in LV blood pool, and the whole mid-LV short axis myocardial slice used as tissue response. Results 36 patients with a false-negative CMR result were identified (table 1). 1 patient had ‘unusable’ image quality grading, and was excluded from further analysis. 4 (11%) patients had images graded as ‘poor quality.’ 10 patients (29%) had inadequate hemodynamic response to adenosine (SBP decrease <10 mm Hg or heart rate increase <10 beats/min). 1 patient (3%) had angiographic 3-vessel disease, supporting balanced ischaemia. A further 6 patients (17%) had an adequate hemodynamic response but MPR <1.5, suggesting possible inadequate vasodilatation (in the absence of triple vessel disease). Of the remaining 14 patients, mean QCA diameter of culprit stenoses was 74%±12%, close to the angiographic cut-off of ≥70% for significant disease (figure 1). Figure 1 Factors associated with a false negative CMR perfusion scan. One factor is shown per patient (n=35). Conclusions Of the many potential factors contributing to false negative CMR perfusion studies, over one third of false negative studies may have been related to lack of efficacy of pharmacological stress at the standard adenosine dose of 140 µg/kg/min. A substantial proportion of patients had coronary stenosis severity close to the angiographic cut-off of 70%, which may represent discordance between anatomical and functional assessment. Non-diagnostic image quality and three-vessel disease made a relatively small contribution to false-negative CMR studies.

Deep learning has demonstrated exceptional performance in medical image analysis, but its effectiveness degrades significantly when applied to different medical centers due to domain shifts. Lesion detection, a critical task in medical imaging, is particularly impacted by this challenge due to the diversity and complexity of lesions, which can arise from different organs, diseases, imaging devices, and other factors. While collecting data and labels from target domains is a feasible solution, annotating medical images is often tedious, expensive, and requires professionals. To address this problem, we combine active learning with domain-invariant feature learning. We propose a Domain-shift Active Learning (DistAL) framework, which includes a transferable feature learning algorithm and a hybrid sample selection strategy. Feature learning incorporates contrastive-consistency training to learn discriminative and domain-invariant features. The sample selection strategy is called RUDY, which jointly considers Representativeness, Uncertainty, and DiversitY. Its goal is to select samples from the unlabeled target domain for cost-effective annotation. It first selects representative samples to deal with domain shift, as well as uncertain ones to improve class separability, and then leverages K-means++ initialization to remove redundant candidates to achieve diversity. We evaluate our method for the task of lesion detection. By selecting only 1.7% samples from the target domain to annotate, DistAL achieves comparable performance to the method trained with all target labels. It outperforms other AL methods in five experiments on eight datasets collected from different hospitals, using different imaging protocols, annotation conventions, and etiologies.

Intraoperative stenosis detection in X-ray coronary angiography via temporal fusion and attention-based CNN.

Background Non-invasive imaging has a key role in the detection of coronary artery disease (CAD). Its importance has been affirmed by recent National Institute of Clinical Excellence (NICE) guidelines. Localisation of ischaemia to a coronary territory is also important in patient management. Cardiac Magnetic Resonance (CMR) perfusion imaging is a well-established and radiation-free test for these purposes. However, there are few data comparing perfusion CMR with Positron Emission Tomography (PET), which is widely regarded as the non-invasive gold standard. Furthermore novel CMR methods, including those based on k-t acceleration techniques, allow myocardial perfusion imaging with unprecedented spatial resolution. Methods 31 patients with known or suspected CAD referred for diagnostic x-ray coronary angiography (XCA) underwent both CMR and PET examinations. Both PET and CMR protocols included adenosine stress and rest perfusion imaging. CMR perfusion imaging was performed at 1.5T with a k-t -accelerated steady-state free-precession sequence. PET imaging was performed with 13 N-Ammonia. The Abstract 118 figure 1 shows an example. Experts blinded to the clinical data analysed the imaging data and experts blinded to the imaging results visually analysed the XCA data. A significant coronary artery stenosis was defined as ≥70% reduction in diameter or a fractional flow reserve Results Patient characteristics are shown in the Abstract 118 table 1. Mean age ± SD was 64±9 years. One CMR examination was non-diagnostic. The interval between PET and CMR was 2±6 days (77% same day), between PET and XCA 22±28 days and between CMR and XCA 22±29 days. The prevalence of CAD was 81%. For the detection of CAD PET sensitivity was 80% (95% CI 59% to 92%) and specificity was 67% (24% to 94%). CMR sensitivity was 83% (95% CI 62 to 95%) and specificity was also 83% (36% to 99%). In patients with CAD ischaemia was localised to 63% of the territories supplied by stenotic arteries by PET and 76% by CMR. Remote ischaemia was detected in 24% of territories by PET and 16% by CMR. Conclusions CMR is at least as accurate as PET for the diagnosis of CAD and also for the localisation of ischaemia to coronary territories. Relatively low numbers mean that CIs are wide and further work is required. Using an anatomic test as the reference-standard for functional tests has well-described limitations. Remote ischaemia is likely to occur for several reasons including underestimation of disease severity at XCA, microvascular disease and also false positive results.

Chapter 4 - Extracting heterogeneous vessels in X-ray coronary angiography via machine learning

X-ray coronary angiography is a primary imaging technique for diagnosing coronary diseases. Although quantitative coronary angiography (QCA) provides morphological information of coronary arteries with objective quantitative measures, considerable training is required to identify the target vessels and understand the tree structure of coronary arteries. Despite the use of computer-aided tools, such as the edge-detection method, manual correction is necessary for accurate segmentation of coronary vessels. In the present study, we proposed a robust method for major vessel segmentation using deep learning models with fully convolutional networks. When angiographic images of 3302 diseased major vessels from 2042 patients were tested, deep learning networks accurately identified and segmented the major vessels in X-ray coronary angiography. The average F1 score reached 0.917, and 93.7% of the images exhibited a high F1 score > 0.8. The most narrowed region at the stenosis was distinctly captured with high connectivity. Robust predictability was validated for the external dataset with different image characteristics. For major vessel segmentation, our approach demonstrated that prediction could be completed in real time with minimal image preprocessing. By applying deep learning segmentation, QCA analysis could be further automated, thereby facilitating the use of QCA-based diagnostic methods.

Improving convolutional neural network learning based on a hierarchical bezier generative model for stenosis detection in X-ray images

1 In a model study the influence of some variables on the accuracy of predictions of phenytoin doses which should produce desired plasma concentrations were studied. All predictions were extrapolations from low test doses with concurrent observations of plasma concentrations. 2 The variables in introduced were:--Variability in observed plasma concentration;--Variability in applied test doses;--Variability in patient characteristics as expressed in Km and Vmax values. 3 The model study generated the same sort of result as a previous one, executed in practice (Driessen. Van der Velde + Höppener, 1980), i.e. predictions are often inaccurate and of the method tested (Richens & Dunlop, 1975; Ludden, Hawkins, Allen & Hoffman, 1976; Martin, Tozer, Sheiner & Riegelman, 1977; Rambeck, Boenigk, Dunlop, Mullen, Wadsworth & Richens, 1980), Richens' first nomogram (Richens & Dunlop, 1975) is to be preferred.

Hybrid classical–quantum Convolutional Neural Network for stenosis detection in X-ray coronary angiography

To establish recruitment approaches that leverage electronic health records in multicenter prediabetes/diabetes clinical trials and compare recruitment outcomes between electronic health record-supported and conventional recruitment methods. Observational analysis of recruitment approaches in the vitamin D and type 2 diabetes (D2d) study, a multicenter trial in participants with prediabetes. Outcomes were adoption of electronic health record-supported recruitment approaches by sites, number of participants screened, recruitment performance (proportion screened who were randomized), and characteristics of participants from electronic health record-supported versus non-electronic health record methods. In total, 2423 participants were randomized: 1920 from electronic health record (mean age of 60 years, 41% women, 68% White) and 503 from non-electronic health record sources (mean age of 56.9 years, 58% women, 61% White). Electronic health record-supported recruitment was adopted by 21 of 22 sites. Electronic health record-supported recruitment was associated with more participants screened versus non-electronic health record methods (4969 vs 2166 participants screened), higher performance (38.6% vs 22.7%), and more randomizations (1918 vs 505). Participants recruited via electronic health record were older, included fewer women and minorities, and reported higher use of dietary supplements. Electronic health record-supported recruitment was incorporated in diverse clinical environments, engaging clinicians either at the individual or the healthcare system level. Establishing electronic health record-supported recruitment approaches across a multicenter prediabetes/diabetes trial is feasible and can be adopted by diverse clinical environments.

Heart disease is the leading cause of death worldwide, responsible for 16% of the world’s total deaths. Coronary artery disease is the most common type of heart disease and is caused by plaque buildup in the wall of the arteries. X-ray coronary angiography is the gold standard for assessing coronary artery disease via estimation of the percentage of narrowing, also known as stenosis. However, errors in physicians’ interpretation of stenosis severity could lead to overuse and underuse of revascularization. Automated techniques can help clinicians in their decision-making. In this paper, we review state-of-the-art techniques developed for coronary artery segmentation, vessel modeling, and stenosis detection. The methods are categorized using criteria such as segmentation methods, evaluation metrics, and validation approaches.