The Current and Future State of AI Interpretation of Medical Images

Advancing Neurocritical Care with Artificial Intelligence and Machine Learning: The Promise, Practicalities, and Pitfalls ahead

Expansion of artificial intelligence (AI) in the field of medicine is changing the paradigm of clinical practice at a rapid pace. Incorporation of AI in medicine offers new tools as well as challenges, and physicians and learners need to adapt to assimilate AI into practice and education. AI can expedite early diagnosis and intervention with real-time multimodal monitoring. AI assistants can decrease the clerical burden of heath care improving the productivity of work force while mitigating burnout. There are still no regulatory parameters for use of AI and regulatory framework is needed for the implementation of AI systems in medicine to ensure transparency, accountability, and equitable access.

Value-Based Radiology in Canada: Reducing Low-Value Care and Improving System Efficiency

Radiology departments are increasingly tasked with managing growing demands on services including long waitlists for scanning and interventional procedures, human health resource shortages, equipment needs, and challenges incorporating advanced imaging solutions. The burden of system inefficiencies and the overuse of "low-value" imaging causes downstream impact on patients at the individual level, the economy and healthcare system at the societal level, and planetary health at an overarching level. Low value imaging includes those performed for an inappropriate clinical indication, with little to no value to the management of the patient, and resulting in healthcare resource waste; it is estimated that up to a quarter of advanced imaging studies in Canada meet this criterion. Strategies to reduce low-value imaging include the development and use of referral guidelines, use of appropriateness criteria, optimization of existing protocols, and integration of clinical decision support tools into the ordering provider's workflow. Additional means of optimizing system efficiency such as centralized intake models, improved access to electronic medical records and outside imaging, enhanced communication with patients and referrers, and the utilization of artificial intelligence will further increase the value of radiology provided to patients and care providers.

Cross-modality PET image synthesis for Parkinson's Disease diagnosis: a leap from [18F]FDG to [11C]CFT

PURPOSE

Dopamine transporter [11C]CFT PET is highly effective for diagnosing Parkinson's Disease (PD), whereas it is not widely available in most hospitals. To develop a deep learning framework to synthesize [11C]CFT PET images from real [18F]FDG PET images and leverage their cross-modal correlation to distinguish PD from normal control (NC).

METHODS

We developed a deep learning framework to synthesize [11C]CFT PET images from real [18F]FDG PET images, and leveraged their cross-modal correlation to distinguish PD from NC. A total of 604 participants (274 with PD and 330 with NC) who underwent [11C]CFT and [18F]FDG PET scans were included. The quality of the synthetic [11C]CFT PET images was evaluated through quantitative comparison with the ground-truth images and radiologist visual assessment. The evaluations of PD diagnosis performance were conducted using biomarker-based quantitative analyses (using striatal binding ratios from synthetic [11C]CFT PET images) and the proposed PD classifier (incorporating both real [18F]FDG and synthetic [11C]CFT PET images).

RESULTS

Visualization result shows that the synthetic [11C]CFT PET images resemble the real ones with no significant differences visible in the error maps. Quantitative evaluation demonstrated that synthetic [11C]CFT PET images exhibited a high peak signal-to-noise ratio (PSNR: 25.0-28.0) and structural similarity (SSIM: 0.87-0.96) across different unilateral striatal subregions. The radiologists achieved a diagnostic accuracy of 91.9% (± 2.02%) based on synthetic [11C]CFT PET images, while biomarker-based quantitative analysis of the posterior putamen yielded an AUC of 0.912 (95% CI, 0.889-0.936), and the proposed PD Classifier achieved an AUC of 0.937 (95% CI, 0.916-0.957).

CONCLUSION

By bridging the gap between [18F]FDG and [11C]CFT, our deep learning framework can significantly enhance PD diagnosis without the need for [11C]CFT tracers, thereby expanding the reach of advanced diagnostic tools to clinical settings where [11C]CFT PET imaging is inaccessible.

Machine Learning for the Early Prediction of Delayed Cerebral Ischemia in Patients With Subarachnoid Hemorrhage: Systematic Review and Meta-Analysis

BACKGROUND

Delayed cerebral ischemia (DCI) is a primary contributor to death after subarachnoid hemorrhage (SAH), with significant incidence. Therefore, early determination of the risk of DCI is an urgent need. Machine learning (ML) has received much attention in clinical practice. Recently, some studies have attempted to apply ML models for early noninvasive prediction of DCI. However, systematic evidence for its predictive accuracy is still lacking.

OBJECTIVE

The aim of this study was to synthesize the prediction accuracy of ML models for DCI to provide evidence for the development or updating of intelligent detection tools.

METHODS

PubMed, Cochrane, Embase, and Web of Science databases were systematically searched up to May 18, 2023. The risk of bias in the included studies was assessed using PROBAST (Prediction Model Risk of Bias Assessment Tool). During the analysis, we discussed the performance of different models in the training and validation sets.

RESULTS

We finally included 48 studies containing 16,294 patients with SAH and 71 ML models with logistic regression as the main model type. In the training set, the pooled concordance index (C index), sensitivity, and specificity of all the models were 0.786 (95% CI 0.737-0.835), 0.77 (95% CI 0.69-0.84), and 0.83 (95% CI 0.75-0.89), respectively, while those of the logistic regression models were 0.770 (95% CI 0.724-0.817), 0.75 (95% CI 0.67-0.82), and 0.71 (95% CI 0.63-0.78), respectively. In the validation set, the pooled C index, sensitivity, and specificity of all the models were 0.767 (95% CI 0.741-0.793), 0.66 (95% CI 0.53-0.77), and 0.78 (95% CI 0.71-0.84), respectively, while those of the logistic regression models were 0.757 (95% CI 0.715-0.800), 0.59 (95% CI 0.57-0.80), and 0.80 (95% CI 0.71-0.87), respectively.

CONCLUSIONS

ML models appear to have relatively desirable power for early noninvasive prediction of DCI after SAH. However, enhancing the prediction sensitivity of these models is challenging. Therefore, efficient, noninvasive, or minimally invasive low-cost predictors should be further explored in future studies to improve the prediction accuracy of ML models.

TRIAL REGISTRATION

PROSPERO (CRD42023438399); https://tinyurl.com/yfuuudde.

Multiparametric MRI for Assessment of the Biological Invasiveness and Prognosis of Pancreatic Ductal Adenocarcinoma in the Era of Artificial Intelligence

Pancreatic ductal adenocarcinoma (PDAC) is the deadliest malignant tumor, with a grim 5-year overall survival rate of about 12%. As its incidence and mortality rates rise, it is likely to become the second-leading cause of cancer-related death. The radiological assessment determined the stage and management of PDAC. However, it is a highly heterogeneous disease with the complexity of the tumor microenvironment, and it is challenging to adequately reflect the biological aggressiveness and prognosis accurately through morphological evaluation alone. With the dramatic development of artificial intelligence (AI), multiparametric magnetic resonance imaging (mpMRI) using specific contrast media and special techniques can provide morphological and functional information with high image quality and become a powerful tool in quantifying intratumor characteristics. Besides, AI has been widespread in the field of medical imaging analysis. Radiomics is the high-throughput mining of quantitative image features from medical imaging that enables data to be extracted and applied for better decision support. Deep learning is a subset of artificial neural network algorithms that can automatically learn feature representations from data. AI-enabled imaging biomarkers of mpMRI have enormous promise to bridge the gap between medical imaging and personalized medicine and demonstrate huge advantages in predicting biological characteristics and the prognosis of PDAC. However, current AI-based models of PDAC operate mainly in the realm of a single modality with a relatively small sample size, and the technical reproducibility and biological interpretation present a barrage of new potential challenges. In the future, the integration of multi-omics data, such as radiomics and genomics, alongside the establishment of standardized analytical frameworks will provide opportunities to increase the robustness and interpretability of AI-enabled image biomarkers and bring these biomarkers closer to clinical practice. EVIDENCE LEVEL: 3 TECHNICAL EFFICACY: Stage 4.

Computer-assisted diagnosis for axillary lymph node metastasis of early breast cancer based on transformer with dual-modal adaptive mid-term fusion using ultrasound elastography

Accurate preoperative qualitative assessment of axillary lymph node metastasis (ALNM) in early breast cancer patients is crucial for precise clinical staging and selection of axillary treatment strategies. Although previous studies have introduced artificial intelligence (AI) to enhance the assessment performance of ALNM, they all focus on the prediction performances of their AI models and neglect the clinical assistance to the radiologists, which brings some issues to the clinical practice. To this end, we propose a human-AI collaboration strategy for ALNM diagnosis of early breast cancer, in which a novel deep learning framework, termed DAMF-former, is designed to assist radiologists in evaluating ALNM. Specifically, the DAMF-former focuses on the axillary region rather than the primary tumor area in previous studies. To mimic the radiologists' alternative integration of the UE images of the target axillary lymph nodes for comprehensive analysis, adaptive mid-term fusion is proposed to alternatively extract and adaptively fuse the high-level features from the dual-modal UE images (i.e., B-mode ultrasound and Shear Wave Elastography). To further improve the diagnostic outcome of the DAMF-former, an adaptive Youden index scheme is proposed to deal with the fully fused dual-modal UE image features at the end of the framework, which can balance the diagnostic performance in terms of sensitivity and specificity. The clinical experiment indicates that the designed DAMF-former can assist and improve the diagnostic abilities of less-experienced radiologists for ALNM. Especially, the junior radiologists can significantly improve the diagnostic outcome from 0.807 AUC [95% CI: 0.781, 0.830] to 0.883 AUC [95% CI: 0.861, 0.902] (P-value <0.0001). Moreover, there are great agreements among radiologists of different levels when assisted by the DAMF-former (Kappa value ranging from 0.805 to 0.895; P-value <0.0001), suggesting that less-experienced radiologists can potentially achieve a diagnostic level similar to that of experienced radiologists through human-AI collaboration. This study explores a potential solution to human-AI collaboration for ALNM diagnosis based on UE images.

An evaluation framework for clinical use of large language models in patient interaction tasks

The integration of large language models (LLMs) into clinical diagnostics has the potential to transform doctor-patient interactions. However, the readiness of these models for real-world clinical application remains inadequately tested. This paper introduces the Conversational Reasoning Assessment Framework for Testing in Medicine (CRAFT-MD) approach for evaluating clinical LLMs. Unlike traditional methods that rely on structured medical examinations, CRAFT-MD focuses on natural dialogues, using simulated artificial intelligence agents to interact with LLMs in a controlled environment. We applied CRAFT-MD to assess the diagnostic capabilities of GPT-4, GPT-3.5, Mistral and LLaMA-2-7b across 12 medical specialties. Our experiments revealed critical insights into the limitations of current LLMs in terms of clinical conversational reasoning, history-taking and diagnostic accuracy. These limitations also persisted when analyzing multimodal conversational and visual assessment capabilities of GPT-4V. We propose a comprehensive set of recommendations for future evaluations of clinical LLMs based on our empirical findings. These recommendations emphasize realistic doctor-patient conversations, comprehensive history-taking, open-ended questioning and using a combination of automated and expert evaluations. The introduction of CRAFT-MD marks an advancement in testing of clinical LLMs, aiming to ensure that these models augment medical practice effectively and ethically.

Advances and challenges in precision imaging

Technological innovations in genomics and related fields have facilitated large sequencing efforts, supported new biological discoveries in cancer, and spawned an era of liquid biopsy biomarkers. Despite these advances, precision oncology has practical constraints, partly related to cancer's biological diversity and spatial and temporal complexity. Advanced imaging technologies are being developed to address some of the current limitations in early detection, treatment selection and planning, drug delivery, and therapeutic response, as well as difficulties posed by drug resistance, drug toxicity, disease monitoring, and metastatic evolution. We discuss key areas of advanced imaging for improving cancer outcomes and survival. Finally, we discuss practical challenges to the broader adoption of precision imaging in the clinic and the need for a robust translational infrastructure.

Foundation model of ECG diagnosis: Diagnostics and explanations of any form and rhythm on ECG

We propose a knowledge-enhanced electrocardiogram (ECG) diagnosis foundation model (KED) that utilizes large language models to incorporate domain-specific knowledge of ECG signals. This model is trained on 800,000 ECGs from nearly 160,000 unique patients. Despite being trained on single-center data, KED demonstrates exceptional zero-shot diagnosis performance across various regions, including different locales in China, the United States, and other regions. This performance spans across all age groups for various conditions such as morphological abnormalities, rhythm abnormalities, conduction blocks, hypertrophy, myocardial ischemia, and infarction. Moreover, KED exhibits robust performance on diseases it has not encountered during its training. When compared to three experienced cardiologists on real clinical datasets, the model achieves comparable performance in zero-shot diagnosis of seven common clinical ECG types. We concentrate on the zero-shot diagnostic capability and the generalization performance of the proposed ECG foundation model, particularly in the context of external multi-center data and previously unseen disease.

Interpretable multi-modal artificial intelligence model for predicting gastric cancer response to neoadjuvant chemotherapy

Neoadjuvant chemotherapy assessment is imperative for prognostication and clinical management of locally advanced gastric cancer. We propose an incremental supervised contrastive learning model (iSCLM), an interpretable artificial intelligence framework integrating pretreatment CT scans and H&E-stained biopsy images, for improved decision-making regarding neoadjuvant chemotherapy. We have constructed and tested iSCLM using retrospective data from 2,387 patients across 10 medical centers and evaluated its discriminative ability in a prospective cohort (132 patients; ChiCTR2300068917). iSCLM achieves areas under receiver operating characteristic curves of 0.846-0.876 across different test cohorts. Computed tomography (CT) and pathological attention heatmaps from Shapley additive explanations and global sort pooling illustrate additional benefits for capturing morphological features through supervised contrastive learning. Specifically, pathological top-ranked tiles exhibit decreased distances to tumor-invasive borders and increased inflammatory cell infiltration in responders compared with non-responders. Moreover, CD11c expression is elevated in responders. The developed interpretable model at the molecular pathology level accurately predicts chemotherapy efficacy.

Deep learning-based multiclass segmentation in aneurysmal subarachnoid hemorrhage

Introduction

Radiological scores used to assess the extent of subarachnoid hemorrhage are limited by intrarater and interrater variability and do not utilize all available information from the imaging. Image segmentation enables precise identification and delineation of objects or regions of interest and offers the potential for automatization of score assessments using precise volumetric information. Our study aims to develop a deep learning model that enables automated multiclass segmentation of structures and pathologies relevant for aneurysmal subarachnoid hemorrhage outcome prediction.

Methods

A set of 73 non-contrast CT scans of patients with aneurysmal subarachnoid hemorrhage were included. Six target classes were manually segmented to create a multiclass segmentation ground truth: subarachnoid, intraventricular, intracerebral and subdural hemorrhage, aneurysms and ventricles. We used the 2d and 3d configurations of the nnU-Net deep learning biomedical image segmentation framework. Additionally, we performed an interrater reliability analysis in our internal test set (n = 20) and an external validation on a set of primary intracerebral hemorrhage patients (n = 104). Segmentation performance was evaluated using the Dice coefficient, volumetric similarity and sensitivity.

Results

The nnU-Net-based segmentation model demonstrated performance closely matching the interrater reliability between two senior raters for the subarachnoid hemorrhage, ventricles, intracerebral hemorrhage classes and overall hemorrhage segmentation. For the hemorrhage segmentation a median Dice coefficient of 0.664 was achieved by the 3d model (0.673 = 2d model). In the external test set a median Dice coefficient of 0.831 for the hemorrhage segmentation was achieved.

Conclusion

Deep learning enables automated multiclass segmentation of aneurysmal subarachnoid hemorrhage-related pathologies and achieves performance approaching that of a human rater. This enables automatized volumetries of pathologies identified on admission CTs in patients with subarachnoid hemorrhage potentially leading to imaging biomarkers for improved outcome prediction.

Large language models as an academic resource for radiologists stepping into artificial intelligence research

BACKGROUND

Radiologists increasingly use artificial intelligence (AI) to enhance diagnostic accuracy and optimize workflows. However, many lack the technical skills to effectively apply machine learning (ML) and deep learning (DL) algorithms, limiting the accessibility of these methods to radiology researchers who could otherwise benefit from them. Large language models (LLMs), such as GPT-4o, may serve as virtual advisors, offering tailored algorithm recommendations for specific research needs. This study evaluates GPT-4o's effectiveness as a recommender system to enhance radiologists' understanding and implementation of AI in research.

INTERVENTION

GPT-4o was used to recommend ML and DL algorithms based on specific details provided by researchers, including dataset characteristics, modality types, data sizes, and research objectives. The model acted as a virtual advisor, guiding researchers in selecting the most appropriate models for their studies.

METHODS

The study systematically evaluated GPT-4o's recommendations for clarity, task alignment, model diversity, and baseline selection. Responses were graded to assess the model's ability to meet the needs of radiology researchers.

RESULTS

GPT-4o effectively recommended appropriate ML and DL algorithms for various radiology tasks, including segmentation, classification, and regression in medical imaging. The model suggested a diverse range of established and innovative algorithms, such as U-Net, Random Forest, Attention U-Net, and EfficientNet, aligning well with accepted practices.

CONCLUSION

GPT-4o shows promise as a valuable tool for radiologists and early career researchers by providing clear and relevant AI and ML algorithm recommendations. Its ability to bridge the knowledge gap in AI implementation could democratize access to advanced technologies, fostering innovation and improving radiology research quality. Further studies should explore integrating LLMs into routine workflows and their role in ongoing professional development.

Reimbursement in the age of generalist radiology artificial intelligence

We argue that generalist radiology artificial intelligence (GRAI) challenges current healthcare reimbursement frameworks. Unlike narrow AI tools, GRAI's multi-task capabilities render existing pathways inadequate. This perspective examines key questions surrounding GRAI reimbursement, including issues of coding, valuation, and coverage policies. We aim to catalyze dialogue among stakeholders about how reimbursement might evolve to accommodate GRAI, potentially influencing AI reimbursement strategies in radiology and beyond.

ChromTR: chromosome detection in raw metaphase cell images via deformable transformers

Chromosome karyotyping is a critical way to diagnose various hematological malignancies and genetic diseases, of which chromosome detection in raw metaphase cell images is the most critical and challenging step. In this work, focusing on the joint optimization of chromosome localization and classification, we propose ChromTR to accurately detect and classify 24 classes of chromosomes in raw metaphase cell images. ChromTR incorporates semantic feature learning and class distribution learning into a unified DETR-based detection framework. Specifically, we first propose a Semantic Feature Learning Network (SFLN) for semantic feature extraction and chromosome foreground region segmentation with object-wise supervision. Next, we construct a Semantic-Aware Transformer (SAT) with two parallel encoders and a Semantic-Aware decoder to integrate global visual and semantic features. To provide a prediction with a precise chromosome number and category distribution, a Category Distribution Reasoning Module (CDRM) is built for foreground-background objects and chromosome class distribution reasoning. We evaluate ChromTR on 1404 newly collected R-band metaphase images and the public G-band dataset AutoKary2022. Our proposed ChromTR outperforms all previous chromosome detection methods with an average precision of 92.56% in R-band chromosome detection, surpassing the baseline method by 3.02%. In a clinical test, ChromTR is also confident in tackling normal and numerically abnormal karyotypes. When extended to the chromosome enumeration task, ChromTR also demonstrates state-of-the-art performances on R-band and G-band two metaphase image datasets. Given these superior performances to other methods, our proposed method has been applied to assist clinical karyotype diagnosis.

From theoretical models to practical deployment: A perspective and case study of opportunities and challenges in AI-driven cardiac auscultation research for low-income settings

This article includes a literature review and a case study of artificial intelligence (AI) heart murmur detection models to analyse the opportunities and challenges in deploying AI in cardiovascular healthcare in low- or medium-income countries (LMICs). This study has two parallel components: (1) The literature review assesses the capacity of AI to aid in addressing the observed disparity in healthcare between high- and low-income countries. Reasons for the limited deployment of machine learning models are discussed, as well as model generalisation. Moreover, the literature review discusses how emerging human-centred deployment research is a promising avenue for overcoming deployment barriers. (2) A predictive AI screening model is developed and tested in a case study on heart murmur detection in rural Brazil. Our binary Bayesian ResNet model leverages overlapping log mel spectrograms of patient heart sound recordings and integrates demographic data and signal features via XGBoost to optimise performance. This is followed by a discussion of the model's limitations, its robustness, and the obstacles preventing its practical application. The difficulty with which this model, and other state-of-the-art models, generalise to out-of-distribution data is also discussed. By integrating the results of the case study with those of the literature review, the NASSS framework was applied to evaluate the key challenges in deploying AI-supported heart murmur detection in low-income settings. The research accentuates the transformative potential of AI-enabled healthcare, particularly for affordable point-of-care screening systems in low-income settings. It also emphasises the necessity of effective implementation and integration strategies to guarantee the successful deployment of these technologies.

Impact of human and artificial intelligence collaboration on workload reduction in medical image interpretation

Clinicians face increasing workloads in medical imaging interpretation, and artificial intelligence (AI) offers potential relief. This meta-analysis evaluates the impact of human-AI collaboration on image interpretation workload. Four databases were searched for studies comparing reading time or quantity for image-based disease detection before and after AI integration. The Quality Assessment of Studies of Diagnostic Accuracy was modified to assess risk of bias. Workload reduction and relative diagnostic performance were pooled using random-effects model. Thirty-six studies were included. AI concurrent assistance reduced reading time by 27.20% (95% confidence interval, 18.22%-36.18%). The reading quantity decreased by 44.47% (40.68%-48.26%) and 61.72% (47.92%-75.52%) when AI served as the second reader and pre-screening, respectively. Overall relative sensitivity and specificity are 1.12 (1.09, 1.14) and 1.00 (1.00, 1.01), respectively. Despite these promising results, caution is warranted due to significant heterogeneity and uneven study quality.

Diagnostic electron microscopy in human infectious diseases - Methods and applications

Diagnostic electron microscopy (EM) is indispensable in all cases of infectious diseases which deserve or profit from the detection of the entire pathogen (i.e. the infectious unit). The focus of its application has shifted during the last decades from routine diagnostics to diagnostics of special cases, emergencies and the investigation of disease pathogenesis. While the focus of application has changed, the methods remain more or less the same. However, since the number of cases for diagnostic EM has declined as the number of laboratories that are able to perform such investigations, the preservation of the present knowledge is important. The aim of this article is to provide a review of the methods and strategies which are useful for diagnostic EM related to infectious diseases in our days. It also addresses weaknesses as well as useful variants or extensions of established methods. The main techniques, negative staining and thin section EM, are described in detail with links to suitable protocols and more recent improvements, such as thin section EM of small volume suspensions. Sample collection, transport and conservation/inactivation are discussed. Strategies of sample examination and requirements for a proper recognition of structures are outlined. Finally, some examples for the actual application of diagnostic EM related to infectious diseases are presented. The outlook section will discuss recent trends in microscopy, such as automated object recognition by machine learning, regarding their potential in supporting diagnostic EM.

The Impact of Artificial Intelligence on Radiologists' Reading Time in Bone Age Radiograph Assessment: A Preliminary Retrospective Observational Study

To evaluate the real-world impact of artificial intelligence (AI) on radiologists' reading time during bone age (BA) radiograph assessments. Patients (<19 year-old) who underwent left-hand BA radiographs between December 2021 and October 2023 were retrospectively included. A commercial AI software was installed from October 2022. Radiologists' reading times, automatically recorded in the PACS log, were compared between the AI-unaided and AI-aided periods using linear regression tests and factors affecting reading time were identified. A total of 3643 radiographs (M:F=1295:2348, mean age 9.12 ± 2.31 years) were included and read by three radiologists, with 2937 radiographs (80.6%) in the AI-aided period. Overall reading times were significantly shorter in the AI-aided period compared to the AI-unaided period (mean 17.2 ± 12.9 seconds vs. mean 22.3 ± 14.7 seconds, p < 0.001). Staff reading times significantly decreased in the AI-aided period (mean 15.9 ± 11.4 seconds vs. mean 19.9 ± 13.4 seconds, p < 0.001), while resident reading times increased (mean 38.3 ± 16.4 seconds vs. 33.6 ± 15.3 seconds, p = 0.013). The use of AI and years of experience in radiology were significant factors affecting reading time (all, p≤0.001). The degree of decrease in reading time as experience increased was larger when utilizing AI (-1.151 for AI-unaided, -1.866 for AI-aided, difference =-0.715, p<0.001). In terms of AI exposure time, the staff's reading time decreased by 0.62 seconds per month (standard error 0.07, p<0.001) during the AI-aided period. The reading time of radiologists for BA assessment was influenced by AI. The time-saving effect of utilizing AI became more pronounced as the radiologists' experience and AI exposure time increased.

Collaboration between clinicians and vision-language models in radiology report generation

Automated radiology report generation has the potential to improve patient care and reduce the workload of radiologists. However, the path toward real-world adoption has been stymied by the challenge of evaluating the clinical quality of artificial intelligence (AI)-generated reports. We build a state-of-the-art report generation system for chest radiographs, called Flamingo-CXR, and perform an expert evaluation of AI-generated reports by engaging a panel of board-certified radiologists. We observe a wide distribution of preferences across the panel and across clinical settings, with 56.1% of Flamingo-CXR intensive care reports evaluated to be preferable or equivalent to clinician reports, by half or more of the panel, rising to 77.7% for in/outpatient X-rays overall and to 94% for the subset of cases with no pertinent abnormal findings. Errors were observed in human-written reports and Flamingo-CXR reports, with 24.8% of in/outpatient cases containing clinically significant errors in both report types, 22.8% in Flamingo-CXR reports only and 14.0% in human reports only. For reports that contain errors we develop an assistive setting, a demonstration of clinician-AI collaboration for radiology report composition, indicating new possibilities for potential clinical utility.

Addressing the Generalizability of AI in Radiology Using a Novel Data Augmentation Framework with Synthetic Patient Image Data: Proof-of-Concept and External Validation for Classification Tasks in Multiple Sclerosis

Artificial intelligence (AI) models often face performance drops after deployment to external datasets. This study evaluated the potential of a novel data augmentation framework based on generative adversarial networks (GANs) that creates synthetic patient image data for model training to improve model generalizability. Model development and external testing were performed for a given classification task, namely the detection of new fluid-attenuated inversion recovery lesions at MRI during longitudinal follow-up of patients with multiple sclerosis (MS). An internal dataset of 669 patients with MS (n = 3083 examinations) was used to develop an attention-based network, trained both with and without the inclusion of the GAN-based synthetic data augmentation framework. External testing was performed on 134 patients with MS from a different institution, with MR images acquired using different scanners and protocols than images used during training. Models trained using synthetic data augmentation showed a significant performance improvement when applied on external data (area under the receiver operating characteristic curve [AUC], 83.6% without synthetic data vs 93.3% with synthetic data augmentation; P = .03), achieving comparable results to the internal test set (AUC, 95.0%; P = .53), whereas models without synthetic data augmentation demonstrated a performance drop upon external testing (AUC, 93.8% on internal dataset vs 83.6% on external data; P = .03). Data augmentation with synthetic patient data substantially improved performance of AI models on unseen MRI data and may be extended to other clinical conditions or tasks to mitigate domain shift, limit class imbalance, and enhance the robustness of AI applications in medical imaging. Keywords: Brain, Brain Stem, Multiple Sclerosis, Synthetic Data Augmentation, Generative Adversarial Network Supplemental material is available for this article. © RSNA, 2024.

Medical large language models are susceptible to targeted misinformation attacks

Large language models (LLMs) have broad medical knowledge and can reason about medical information across many domains, holding promising potential for diverse medical applications in the near future. In this study, we demonstrate a concerning vulnerability of LLMs in medicine. Through targeted manipulation of just 1.1% of the weights of the LLM, we can deliberately inject incorrect biomedical facts. The erroneous information is then propagated in the model's output while maintaining performance on other biomedical tasks. We validate our findings in a set of 1025 incorrect biomedical facts. This peculiar susceptibility raises serious security and trustworthiness concerns for the application of LLMs in healthcare settings. It accentuates the need for robust protective measures, thorough verification mechanisms, and stringent management of access to these models, ensuring their reliable and safe use in medical practice.

Exploring the changing landscape of medical imaging: insights from highly cited studies before and during the COVID-19 pandemic

OBJECTIVES

To investigate whether and how the COVID-19 pandemic has changed medical imaging trends by synthesizing the highly cited studies before and during the pandemic.

METHODS

In this cross-sectional study, we identified highly cited studies on medical imaging from the essential science indicators (ESI) database, categorizing them into two periods: before the pandemic (January 2016-December 2019) and during the pandemic (January 2020-December 2023). We conducted a global research landscape comparative analysis and utilized CiteSpace and VOSviewer software to create knowledge maps for analyzing the co-occurrences of keywords and references in this field.

RESULTS

A total of 2914 highly cited studies were included in this study, which revealed a notable 30.1% increase in medical imaging publications during the pandemic. Enhanced international cooperation has been observed, with European countries and the US leading the research efforts. Keyword analysis revealed that artificial intelligence (AI) has remained a dominant hotspot in medical imaging research before and during the pandemic. References analysis showed a shift in focus towards COVID-19-related studies, overshadowing some important areas including cancer imaging, cardiac imaging, and neuroimaging.

CONCLUSIONS

Over the past four years, the COVID-19 pandemic has led to changes in the research output, international collaborations, and hotspots within highly cited medical imaging studies. Navigating the post-COVID era, it is imperative to continue fostering international collaboration, prioritize resource allocation to refocus on overlooked research areas, and develop long-term strategic plans to prepare for and mitigate the impact of future public health crises.

KEY POINTS

Question Understanding how the COVID-19 pandemic has changed medical imaging trends and priorities, which is crucial for preparing against future public health crises, remains unclear. Findings The COVID-19 pandemic has led to increases in highly cited medical imaging studies, enhancements in international collaborations, and shifts in research hotspots. Clinical relevance This study provides a comprehensive overview of highly cited studies on medical imaging before and during the pandemic, highlighting the pivotal role of AI in the post-COVID era and emphasizing the need to refocus on potentially neglected research areas.

Use of artificial intelligence algorithms to analyse systemic sclerosis-interstitial lung disease imaging features

The use of artificial intelligence (AI) in high-resolution computed tomography (HRCT) for diagnosing systemic sclerosis-associated interstitial lung disease (SSc-ILD) is relatively limited. This study aimed to analyse lung HRCT images of patients with systemic sclerosis with interstitial lung disease (SSc-ILD) using artificial intelligence (AI), conduct correlation analysis with clinical manifestations and prognosis, and explore the features and prognosis of SSc-ILD. Overall, 72 lung HRCT images and clinical data of 58 patients with SSC-ILD were collected. ILD lesion type, location, and volume on HRCT images were identified and evaluated using AI. The imaging characteristics of diffuse SSC (dSSc)-ILD and limited SSc-ILD (lSSc-ILD) were statistically analysed. Furthermore, the correlations between lesion type, clinical indicators, and prognosis were investigated. dSSc and lSSc were more prevalent in patients with a disease duration of < 1 and ≥ 5 years, respectively. SSc-ILD mainly comprises non-specific interstitial pneumonia (NSIP), usual interstitial pneumonia (UIP), and unclassifiable idiopathic interstitial pneumonia. HRCT reveals various lesion types in the early stages of the disease, with an increase in the number of lesion types as the disease progresses. Lesions appearing as grid, ground-glass, and nodular shadows were dispersed throughout both lungs, while those appearing as consolidation shadows and honeycomb were distributed across the lungs. Ground-glass opacity lesion type was absent on HRCT images of patients with SSc-ILD and pulmonary hypertension. This study showed that AI can efficiently analyse imaging characteristics of SSc-ILD, demonstrating its potential to learn from complex images with high generalisation ability.

Gradient-Based Saliency Maps Are Not Trustworthy Visual Explanations of Automated AI Musculoskeletal Diagnoses

Saliency maps are popularly used to "explain" decisions made by modern machine learning models, including deep convolutional neural networks (DCNNs). While the resulting heatmaps purportedly indicate important image features, their "trustworthiness," i.e., utility and robustness, has not been evaluated for musculoskeletal imaging. The purpose of this study was to systematically evaluate the trustworthiness of saliency maps used in disease diagnosis on upper extremity X-ray images. The underlying DCNNs were trained using the Stanford MURA dataset. We studied four trustworthiness criteria-(1) localization accuracy of abnormalities, (2) repeatability, (3) reproducibility, and (4) sensitivity to underlying DCNN weights-across six different gradient-based saliency methods (Grad-CAM (GCAM), gradient explanation (GRAD), integrated gradients (IG), Smoothgrad (SG), smooth IG (SIG), and XRAI). Ground-truth was defined by the consensus of three fellowship-trained musculoskeletal radiologists who each placed bounding boxes around abnormalities on a holdout saliency test set. Compared to radiologists, all saliency methods showed inferior localization (AUPRCs: 0.438 (SG)-0.590 (XRAI); average radiologist AUPRC: 0.816), repeatability (IoUs: 0.427 (SG)-0.551 (IG); average radiologist IOU: 0.613), and reproducibility (IoUs: 0.250 (SG)-0.502 (XRAI); average radiologist IOU: 0.613) on abnormalities such as fractures, orthopedic hardware insertions, and arthritis. Five methods (GCAM, GRAD, IG, SG, XRAI) passed the sensitivity test. Ultimately, no saliency method met all four trustworthiness criteria; therefore, we recommend caution and rigorous evaluation of saliency maps prior to their clinical use.

A Responsible Framework for Applying Artificial Intelligence on Medical Images and Signals at the Point of Care: The PACS-AI Platform

The potential of artificial intelligence (AI) in medicine lies in its ability to enhance clinicians' capacity to analyse medical images, thereby improving diagnostic precision and accuracy and thus enhancing current tests. However, the integration of AI within health care is fraught with difficulties. Heterogeneity among health care system applications, reliance on proprietary closed-source software, and rising cybersecurity threats pose significant challenges. Moreover, before their deployment in clinical settings, AI models must demonstrate their effectiveness across a wide range of scenarios and must be validated by prospective studies, but doing so requires testing in an environment mirroring the clinical workflow, which is difficult to achieve without dedicated software. Finally, the use of AI techniques in health care raises significant legal and ethical issues, such as the protection of patient privacy, the prevention of bias, and the monitoring of the device's safety and effectiveness for regulatory compliance. This review describes challenges to AI integration in health care and provides guidelines on how to move forward. We describe an open-source solution that we developed that integrates AI models into the Picture Archives Communication System (PACS), called PACS-AI. This approach aims to increase the evaluation of AI models by facilitating their integration and validation with existing medical imaging databases. PACS-AI may overcome many current barriers to AI deployment and offer a pathway toward responsible, fair, and effective deployment of AI models in health care. In addition, we propose a list of criteria and guidelines that AI researchers should adopt when publishing a medical AI model to enhance standardisation and reproducibility.

Enhancing representation in radiography-reports foundation model: a granular alignment algorithm using masked contrastive learning

Recently, multi-modal vision-language foundation models have gained significant attention in the medical field. While these models offer great opportunities, they still face crucial challenges, such as the requirement for fine-grained knowledge understanding in computer-aided diagnosis and the capability of utilizing very limited or even no task-specific labeled data in real-world clinical applications. In this study, we present MaCo, a masked contrastive chest X-ray foundation model that tackles these challenges. MaCo explores masked contrastive learning to simultaneously achieve fine-grained image understanding and zero-shot learning for a variety of medical imaging tasks. It designs a correlation weighting mechanism to adjust the correlation between masked chest X-ray image patches and their corresponding reports, thereby enhancing the model's representation learning capabilities. To evaluate the performance of MaCo, we conducted extensive experiments using 6 well-known open-source X-ray datasets. The experimental results demonstrate the superiority of MaCo over 10 state-of-the-art approaches across tasks such as classification, segmentation, detection, and phrase grounding. These findings highlight the significant potential of MaCo in advancing a wide range of medical image analysis tasks.

A knowledge-based decision support system to support family doctors in personalizing type-2 diabetes mellitus medical nutrition therapy

BACKGROUND

Type-2 Diabetes Mellitus (T2D) is a growing concern worldwide, and family doctors are called to help diabetic patients manage this chronic disease, also with Medical Nutrition Therapy (MNT). However, MNT for Diabetes is usually standardized, while it would be much more effective if tailored to the patient. There is a gap in patient-tailored MNT which, if addressed, could support family doctors in delivering effective recommendations. In this context, decision support systems (DSSs) are valuable tools for physicians to support MNT for T2D patients - as long as DSSs are transparent to humans in their decision-making process. Indeed, the lack of transparency in data-driven DSS might hinder their adoption in clinical practice, thus leaving family physicians to adopt general nutrition guidelines provided by the national healthcare systems.

METHOD

This work presents a prototypical ontology-based clinical Decision Support System (OnT2D- DSS) aimed at assisting general practice doctors in managing T2D patients, specifically in creating a tailored dietary plan, leveraging clinical expert knowledge. OnT2D-DSS exploits clinical expert knowledge formalized as a domain ontology to identify a patient's phenotype and potential comorbidities, providing personalized MNT recommendations for macro- and micro-nutrient intake. The system can be accessed via a prototypical interface.

RESULTS

Two preliminary experiments are conducted to assess both the quality and correctness of the inferences provided by the system and the usability and acceptance of the OnT2D-DSS (conducted with nutrition experts and family doctors, respectively).

CONCLUSIONS

Overall, the system is deemed accurate by the nutrition experts and valuable by the family doctors, with minor suggestions for future improvements collected during the experiments.

Artificial Intelligence-Enhanced Neurocritical Care for Traumatic Brain Injury : Past, Present and Future

In neurointensive care units (NICUs), particularly in cases involving traumatic brain injury (TBI), swift and accurate decision-making is critical because of rapidly changing patient conditions and the risk of secondary brain injury. The use of artificial intelligence (AI) in NICU can enhance clinical decision support and provide valuable assistance in these complex scenarios. This article aims to provide a comprehensive review of the current status and future prospects of AI utilization in the NICU, along with the challenges that must be overcome to realize this. Presently, the primary application of AI in NICU is outcome prediction through the analysis of preadmission and high-resolution data during admission. Recent applications include augmented neuromonitoring via signal quality control and real-time event prediction. In addition, AI can integrate data gathered from various measures and support minimally invasive neuromonitoring to increase patient safety. However, despite the recent surge in AI adoption within the NICU, the majority of AI applications have been limited to simple classification tasks, thus leaving the true potential of AI largely untapped. Emerging AI technologies, such as generalist medical AI and digital twins, harbor immense potential for enhancing advanced neurocritical care through broader AI applications. If challenges such as acquiring high-quality data and ethical issues are overcome, these new AI technologies can be clinically utilized in the actual NICU environment. Emphasizing the need for continuous research and development to maximize the potential of AI in the NICU, we anticipate that this will further enhance the efficiency and accuracy of TBI treatment within the NICU.

Capability of multimodal large language models to interpret pediatric radiological images

BACKGROUND

There is a dearth of artificial intelligence (AI) development and research dedicated to pediatric radiology. The newest iterations of large language models (LLMs) like ChatGPT can process image and video input in addition to text. They are thus theoretically capable of providing impressions of input radiological images.

OBJECTIVE

To assess the ability of multimodal LLMs to interpret pediatric radiological images.

MATERIALS AND METHODS

Thirty medically significant cases were collected and submitted to GPT-4 (OpenAI, San Francisco, CA), Gemini 1.5 Pro (Google, Mountain View, CA), and Claude 3 Opus (Anthropic, San Francisco, CA) with a short history for a total of 90 images. AI responses were recorded and independently assessed for accuracy by a resident and attending physician. 95% confidence intervals were determined using the adjusted Wald method.

RESULTS

Overall, the models correctly diagnosed 27.8% (25/90) of images (95% CI=19.5-37.8%), were partially correct for 13.3% (12/90) of images (95% CI=2.7-26.4%), and were incorrect for 58.9% (53/90) of images (95% CI=48.6-68.5%).

CONCLUSION

Multimodal LLMs are not yet capable of interpreting pediatric radiological images.

AI supported detection of cerebral multiple sclerosis lesions decreases radiologic reporting times

PURPOSE

Multiple Sclerosis (MS) is a common autoimmune disease of the central nervous system. MRI plays a crucial role in diagnosing as well as in disease and treatment monitoring. Therefore, evaluation of cerebral MRI of MS patients is part of daily clinical routine. A growing number of companies offer commercial software to support the reporting with automated lesion detection. Aim of this study was to evaluate the effect of such a software with AI supported lesion detection to the radiologic reporting.

METHOD

Four radiologist each counted MS-lesions in MRI examinations of 50 patients separated by the locations periventricular, cortical/juxtacortical, infrantentorial and unspecific white matter. After at least six weeks they repeated the evaluation, this time using the AI based software mdbrain for lesion detection. In both settings the required time was documented. Further the radiologists evaluated follow-up MRI of 50 MS-patients concerning new and enlarging lesions in the same manner.

RESULTS

To determine the lesion-load the average reporting time decreased from 286.85 sec to 196.34 sec (p > 0.001). For the evaluation of the follow-up images the reporting time dropped from 196.17 sec to 120.87 sec (p < 0.001). The interrater reliabilities showed no significant differences for the determination of lesion-load (0.83 without vs. 0.8 with software support) and for the detection of new/enlarged lesions (0.92 without vs. 0.82 with software support).

CONCLUSION

For the evaluation of MR images of MS patients, an AI-based support for image-interpretation can significantly decreases reporting times.

Factors for increasing positive predictive value of pneumothorax detection on chest radiographs using artificial intelligence

This study evaluated the positive predictive value (PPV) of artificial intelligence (AI) in detecting pneumothorax on chest radiographs (CXRs) and its affecting factors. Patients determined to have pneumothorax on CXR by a commercial AI software from March to December 2021 were included retrospectively. The PPV was evaluated according to the true-positive (TP) and false-positive (FP) diagnosis determined by radiologists. To know the factors that might influence the results, logistic regression with generalized estimating equation was used. Among a total of 87,658 CXRs, 308 CXRs with 331 pneumothoraces from 283 patients were finally included. The overall PPV of AI about pneumothorax was 41.1% (TF:FP = 136:195). The PA view (odds ratio [OR], 29.837; 95% confidence interval [CI], 15.062-59.107), high abnormality score (OR, 1.081; 95% CI, 1.066-1.097), large amount of pneumothorax (OR, 1.005; 95% CI, 1.003-1.007), presence of ipsilateral atelectasis (OR, 3.508; 95% CI, 1.509-8.156) and a small amount of ipsilateral pleural effusion (OR, 5.277; 95% CI, 2.55-10.919) had significant effects on the increasing PPV. Therefore, PPV for pneumothorax diagnosis using AI can vary based on patients' factors, image-acquisition protocols, and the presence of concurrent lesions on CXR.

The state of artificial intelligence in medical research: A survey of corresponding authors from top medical journals

Natural Language Processing (NLP) is a subset of artificial intelligence that enables machines to understand and respond to human language through Large Language Models (LLMs)‥ These models have diverse applications in fields such as medical research, scientific writing, and publishing, but concerns such as hallucination, ethical issues, bias, and cybersecurity need to be addressed. To understand the scientific community's understanding and perspective on the role of Artificial Intelligence (AI) in research and authorship, a survey was designed for corresponding authors in top medical journals. An online survey was conducted from July 13th, 2023, to September 1st, 2023, using the SurveyMonkey web instrument, and the population of interest were corresponding authors who published in 2022 in the 15 highest-impact medical journals, as ranked by the Journal Citation Report. The survey link has been sent to all the identified corresponding authors by mail. A total of 266 authors answered, and 236 entered the final analysis. Most of the researchers (40.6%) reported having moderate familiarity with artificial intelligence, while a minority (4.4%) had no associated knowledge. Furthermore, the vast majority (79.0%) believe that artificial intelligence will play a major role in the future of research. Of note, no correlation between academic metrics and artificial intelligence knowledge or confidence was found. The results indicate that although researchers have varying degrees of familiarity with artificial intelligence, its use in scientific research is still in its early phases. Despite lacking formal AI training, many scholars publishing in high-impact journals have started integrating such technologies into their projects, including rephrasing, translation, and proofreading tasks. Efforts should focus on providing training for their effective use, establishing guidelines by journal editors, and creating software applications that bundle multiple integrated tools into a single platform.

Improved robustness for deep learning-based segmentation of multi-center myocardial perfusion cardiovascular MRI datasets using data-adaptive uncertainty-guided space-time analysis

BACKGROUND

Fully automatic analysis of myocardial perfusion cardiovascular magnetic resonance imaging datasets enables rapid and objective reporting of stress/rest studies in patients with suspected ischemic heart disease. Developing deep learning techniques that can analyze multi-center datasets despite limited training data and variations in software (pulse sequence) and hardware (scanner vendor) is an ongoing challenge.

METHODS

Datasets from three medical centers acquired at 3T (n = 150 subjects; 21,150 first-pass images) were included: an internal dataset (inD; n = 95) and two external datasets (exDs; n = 55) used for evaluating the robustness of the trained deep neural network (DNN) models against differences in pulse sequence (exD-1) and scanner vendor (exD-2). A subset of inD (n = 85) was used for training/validation of a pool of DNNs for segmentation, all using the same spatiotemporal U-Net architecture and hyperparameters but with different parameter initializations. We employed a space-time sliding-patch analysis approach that automatically yields a pixel-wise "uncertainty map" as a byproduct of the segmentation process. In our approach, dubbed data-adaptive uncertainty-guided space-time (DAUGS) analysis, a given test case is segmented by all members of the DNN pool and the resulting uncertainty maps are leveraged to automatically select the "best" one among the pool of solutions. For comparison, we also trained a DNN using the established approach with the same settings (hyperparameters, data augmentation, etc.).

RESULTS

The proposed DAUGS analysis approach performed similarly to the established approach on the inD (Dice score for the testing subset of inD: 0.896 ± 0.050 vs 0.890 ± 0.049; p = n.s.) whereas it significantly outperformed on the exDs (Dice for exD-1: 0.885 ± 0.040 vs 0.849 ± 0.065, p < 0.005; Dice for exD-2: 0.811 ± 0.070 vs 0.728 ± 0.149, p < 0.005). Moreover, the number of image series with "failed" segmentation (defined as having myocardial contours that include bloodpool or are noncontiguous in ≥1 segment) was significantly lower for the proposed vs the established approach (4.3% vs 17.1%, p < 0.0005).

CONCLUSION

The proposed DAUGS analysis approach has the potential to improve the robustness of deep learning methods for segmentation of multi-center stress perfusion datasets with variations in the choice of pulse sequence, site location, or scanner vendor.

Using AI to Identify Unremarkable Chest Radiographs for Automatic Reporting

Background Radiology practices have a high volume of unremarkable chest radiographs and artificial intelligence (AI) could possibly improve workflow by providing an automatic report. Purpose To estimate the proportion of unremarkable chest radiographs, where AI can correctly exclude pathology (ie, specificity) without increasing diagnostic errors. Materials and Methods In this retrospective study, consecutive chest radiographs in unique adult patients (≥18 years of age) were obtained January 1-12, 2020, at four Danish hospitals. Exclusion criteria included insufficient radiology reports or AI output error. Two thoracic radiologists, who were blinded to AI output, labeled chest radiographs as "remarkable" or "unremarkable" based on predefined unremarkable findings (reference standard). Radiology reports were classified similarly. A commercial AI tool was adapted to output a chest radiograph "remarkableness" probability, which was used to calculate specificity at different AI sensitivities. Chest radiographs with missed findings by AI and/or the radiology report were graded by one thoracic radiologist as critical, clinically significant, or clinically insignificant. Paired proportions were compared using the McNemar test. Results A total of 1961 patients were included (median age, 72 years [IQR, 58-81 years]; 993 female), with one chest radiograph per patient. The reference standard labeled 1231 of 1961 chest radiographs (62.8%) as remarkable and 730 of 1961 (37.2%) as unremarkable. At 99.9%, 99.0%, and 98.0% sensitivity, the AI had a specificity of 24.5% (179 of 730 radiographs [95% CI: 21, 28]), 47.1% (344 of 730 radiographs [95% CI: 43, 51]), and 52.7% (385 of 730 radiographs [95% CI: 49, 56]), respectively. With the AI fixed to have a similar sensitivity as radiology reports (87.2%), the missed findings of AI and reports had 2.2% (27 of 1231 radiographs) and 1.1% (14 of 1231 radiographs) classified as critical (P = .01), 4.1% (51 of 1231 radiographs) and 3.6% (44 of 1231 radiographs) classified as clinically significant (P = .46), and 6.5% (80 of 1231) and 8.1% (100 of 1231) classified as clinically insignificant (P = .11), respectively. At sensitivities greater than or equal to 95.4%, the AI tool exhibited less than or equal to 1.1% critical misses. Conclusion A commercial AI tool used off-label could correctly exclude pathology in 24.5%-52.7% of all unremarkable chest radiographs at greater than or equal to 98% sensitivity. The AI had equal or lower rates of critical misses than radiology reports at sensitivities greater than or equal to 95.4%. These results should be confirmed in a prospective study. © RSNA, 2024 Supplemental material is available for this article. See also the editorial by Yoon and Hwang in this issue.

Automated Real-Time Detection of Lung Sliding Using Artificial Intelligence: A Prospective Diagnostic Accuracy Study

BACKGROUND

Rapid evaluation for pneumothorax is a common clinical priority. Although lung ultrasound (LUS) often is used to assess for pneumothorax, its diagnostic accuracy varies based on patient and provider factors. To enhance the performance of LUS for pulmonary pathologic features, artificial intelligence (AI)-assisted imaging has been adopted; however, the diagnostic accuracy of AI-assisted LUS (AI-LUS) deployed in real time to diagnose pneumothorax remains unknown.

RESEARCH QUESTION

In patients with suspected pneumothorax, what is the real-time diagnostic accuracy of AI-LUS to recognize the absence of lung sliding?

STUDY DESIGN AND METHODS

We performed a prospective AI-assisted diagnostic accuracy study of AI-LUS to recognize the absence of lung sliding in a convenience sample of patients with suspected pneumothorax. After calibrating the model parameters and imaging settings for bedside deployment, we prospectively evaluated its diagnostic accuracy for lung sliding compared with a reference standard of expert consensus.

RESULTS

Two hundred forty-one lung sliding evaluations were derived from 62 patients. AI-LUS showed a sensitivity of 0.921 (95% CI, 0.792-0.973), specificity of 0.802 (95% CI, 0.735-0.856), area under the receiver operating characteristic curve of 0.885 (95% CI, 0.828-0.956), and accuracy of 0.824 (95% CI, 0.766-0.870) for the diagnosis of absent lung sliding.

INTERPRETATION

In this study, real-time AI-LUS showed high sensitivity and moderate specificity to identify the absence of lung sliding. Further research to improve model performance and optimize the integration of AI-LUS into existing diagnostic pathways is warranted.

Artificial Intelligence Assessment of Biological Age From Transthoracic Echocardiography: Discrepancies with Chronologic Age Predict Significant Excess Mortality

BACKGROUND

Age and sex can be estimated using artificial intelligence on the basis of various sources. The aims of this study were to test whether convolutional neural networks could be trained to estimate age and predict sex using standard transthoracic echocardiography and to evaluate the prognostic implications.

METHODS

The algorithm was trained on 76,342 patients, validated in 22,825 patients, and tested in 20,960 patients. It was then externally validated using data from a different hospital (n = 556). Finally, a prospective cohort of handheld point-of-care ultrasound devices (n = 319; ClinicalTrials.gov identifier NCT05455541) was used to confirm the findings. A multivariate Cox regression model was used to investigate the association between age estimation and chronologic age with overall survival.

RESULTS

The mean absolute error in age estimation was 4.9 years, with a Pearson correlation coefficient of 0.922. The probabilistic value of sex had an overall accuracy of 96.1% and an area under the curve of 0.993. External validation and prospective study cohorts yielded consistent results. Finally, survival analysis demonstrated that age prediction ≥5 years vs chronologic age was associated with an independent 34% increased risk for death during follow-up (P < .001).

CONCLUSIONS

Applying artificial intelligence to standard transthoracic echocardiography allows the prediction of sex and the estimation of age. Machine-based estimation is an independent predictor of overall survival and, with further evaluation, can be used for risk stratification and estimation of biological age.

The potential role for artificial intelligence in fracture risk prediction

Osteoporotic fractures are a major health challenge in older adults. Despite the availability of safe and effective therapies for osteoporosis, these therapies are underused in individuals at high risk for fracture, calling for better case-finding and fracture risk assessment strategies. Artificial intelligence (AI) and machine learning (ML) hold promise for enhancing identification of individuals at high risk for fracture by distilling useful features from high-dimensional data derived from medical records, imaging, and wearable devices. AI-ML could enable automated opportunistic screening for vertebral fractures and osteoporosis, home-based monitoring and intervention targeting lifestyle factors, and integration of multimodal features to leverage fracture prediction, ultimately aiding improved fracture risk assessment and individualised treatment. Optimism must be balanced with consideration for the explainability of AI-ML models, biases (including information inequity in numerically under-represented populations), model limitations, and net clinical benefit and workload impact. Clinical integration of AI-ML algorithms has the potential to transform osteoporosis management, offering a more personalised approach to reduce the burden of osteoporotic fractures.

Machine learning approach for prediction of outcomes in anticoagulated patients with atrial fibrillation

BACKGROUND

The accuracy of available prediction tools for clinical outcomes in patients with atrial fibrillation (AF) remains modest. Machine Learning (ML) has been used to predict outcomes in the AF population, but not in a population entirely on anticoagulant therapy.

METHODS AND AIMS

Different supervised ML models were applied to predict all-cause death, cardiovascular (CV) death, major bleeding and stroke in anticoagulated patients with AF, processing data from the multicenter START-2 Register.

RESULTS

11078 AF patients (male n = 6029, 54.3%) were enrolled with a median follow-up period of 1.5 years [IQR 1.0-2.6]. Patients on Vitamin K Antagonists (VKA) were 5135 (46.4%) and 5943 (53.6%) were on Direct Oral Anticoagulants (DOAC). Using Multi-Gate Mixture of Experts, a cross-validated AUC of 0.779 ± 0.016 and 0.745 ± 0.022 were obtained, respectively, for the prediction of all-cause death and CV-death in the overall population. The best ML model outperformed CHA2DSVA2SC and HAS-BLED for all-cause death prediction (p < 0.001 for both). When compared to HAS-BLED, Gradient Boosting improved major bleeding prediction in DOACs patients (0.711 vs. 0.586, p < 0.001). A very low number of events during follow-up (52) resulted in a suboptimal ischemic stroke prediction (best AUC of 0.606 ± 0.117 in overall population). Body mass index, age, renal function, platelet count and hemoglobin levels resulted the most important variables for ML prediction.

CONCLUSIONS

In AF patients, ML models showed good discriminative ability to predict all-cause death, regardless of the type of anticoagulation strategy, and major bleeding on DOAC therapy, outperforming CHA2DS2VASC and the HAS-BLED scores for risk prediction in these populations.

A Novel Video-Based Methodology for Automated Classification of Dystonia and Choreoathetosis in Dyskinetic Cerebral Palsy During a Lower Extremity Task

BACKGROUND

Movement disorders in children and adolescents with dyskinetic cerebral palsy (CP) are commonly assessed from video recordings, however scoring is time-consuming and expert knowledge is required for an appropriate assessment.

OBJECTIVE

To explore a machine learning approach for automated classification of amplitude and duration of distal leg dystonia and choreoathetosis within short video sequences.

METHODS

Available videos of a heel-toe tapping task were preprocessed to optimize key point extraction using markerless motion analysis. Postprocessed key point data were passed to a time series classification ensemble algorithm to classify dystonia and choreoathetosis duration and amplitude classes (scores 0, 1, 2, 3, and 4), respectively. As ground truth clinical scoring of dystonia and choreoathetosis by the Dyskinesia Impairment Scale was used. Multiclass performance metrics as well as metrics for summarized scores: absence (score 0) and presence (score 1-4) were determined.

RESULTS

Thirty-three participants were included: 29 with dyskinetic CP and 4 typically developing, age 14 years:6 months ± 5 years:15 months. The multiclass accuracy results for dystonia were 77% for duration and 68% for amplitude; for choreoathetosis 30% for duration and 38% for amplitude. The metrics for score 0 versus score 1 to 4 revealed an accuracy of 81% for dystonia duration, 77% for dystonia amplitude, 53% for choreoathetosis duration and amplitude.

CONCLUSIONS

This methodology study yielded encouraging results in distinguishing between presence and absence of dystonia, but not for choreoathetosis. A larger dataset is required for models to accurately represent distinct classes/scores. This study presents a novel methodology of automated assessment of movement disorders solely from video data.