Artificial intelligence (AI) has enabled accurate and fast plaque quantification from coronary computed tomography angiography (CCTA). However, AI detects any coronary plaque in up to 97% of patients. To avoid overdiagnosis, a plaque burden safety cut-off for future coronary events is needed. Percent atheroma volume (PAV) was quantified with AI-guided quantitative computed tomography in a blinded fashion. Safety cut-off derivation was performed in the Turku CCTA registry (Finland), and pre-defined as ≥90% sensitivity for acute coronary syndrome (ACS). External validation was performed in the Amsterdam CCTA registry (the Netherlands). In the derivation cohort, 100/2271 (4.4%) patients experienced ACS (median follow-up 6.9 years). A threshold of PAV ≥ 2.6% was derived with 90.0% sensitivity and negative predictive value (NPV) of 99.0%. In the validation cohort 27/568 (4.8%) experienced ACS (median follow-up 6.7 years) with PAV ≥ 2.6% showing 92.6% sensitivity and 99.0% NPV for ACS. In the derivation cohort, 45.2% of patients had PAV < 2.6 vs. 4.3% with PAV 0% (no plaque) (P < 0.001) (validation cohort: 34.3% PAV < 2.6 vs. 2.6% PAV 0%; P < 0.001). Patients with PAV ≥ 2.6% had higher adjusted ACS rates in the derivation [Hazard ratio (HR) 4.65, 95% confidence interval (CI) 2.33-9.28, P < 0.001] and validation cohort (HR 7.31, 95% CI 1.62-33.08, P = 0.010), respectively. This study suggests that PAV up to 2.6% quantified by AI is associated with low-ACS risk in two independent patient cohorts. This cut-off may be helpful for clinical application of AI-guided CCTA analysis, which detects any plaque in up to 96-97% of patients.

Coronary computed tomography angiography (CCTA) and magnetic resonance imaging (MRI) can predict periprocedural myocardial injury (PMI) after percutaneous coronary intervention (PCI). We aimed to investigate whether integrating MRI with CCTA, using the latest imaging and quantitative techniques, improves PMI prediction and to explore a potential hybrid CCTA-MRI strategy. This prospective, multi-centre study conducted coronary atherosclerosis T1-weighted characterization MRI for patients scheduled for elective PCI for an atherosclerotic lesion detected on CCTA without prior revascularization. PMI was defined as post-PCI troponin-T > 5× the upper reference limit. Using deep learning-enabled software, volumes of total plaque, calcified plaque, non-calcified plaque (NCP), and low-attenuation plaque (LAP; < 30 Hounsfield units) were quantified on CCTA. In non-contrast T1-weighted MRI, high-intensity plaque (HIP) volume was quantified as voxels with signal intensity exceeding that of the myocardium, weighted by their respective intensities. Of the 132 lesions from 120 patients, 43 resulted in PMI. In the CCTA-only strategy, LAP volume (P = 0.012) and NCP volume (P = 0.016) were independently associated with PMI. In integrating MRI with CCTA, LAP volume (P = 0.029), and HIP volume (P = 0.024) emerged as independent predictors. MRI integration with CCTA achieved a higher C-statistic value than CCTA alone (0.880 vs. 0.738; P = 0.004). A hybrid CCTA-MRI strategy, incorporating MRI for lesions with intermediate PMI risk based on CCTA, maintained superior diagnostic accuracy over the CCTA-only strategy (0.803 vs. 0.705; P = 0.028). Integrating MRI with CCTA improves PMI prediction compared with CCTA alone.

Thyroid ultrasound (US) is an essential tool for detecting and characterizing thyroid nodules. In this study, we propose an innovative approach to enhance thyroid nodule assessment by integrating Doppler US images with grayscale US images through weakly supervised data augmentation networks (WSDAN). Our method reduces background noise by replacing inefficient augmentation strategies, such as random cropping, with an advanced technique guided by bounding boxes derived from Doppler US images. This targeted augmentation significantly improves model performance in both classification and localization of thyroid nodules. The training dataset comprises 1288 paired grayscale and Doppler US images, with an additional 190 pairs used for three-fold cross-validation. To evaluate the model's efficacy, we tested it on a separate set of 190 grayscale US images. Compared to five state-of-the-art models and the original WSDAN, our Enhanced WSDAN model achieved superior performance. For classification, it reached an accuracy of 91%. For localization, it achieved Dice and Jaccard indices of 75% and 87%, respectively, demonstrating its potential as a valuable clinical tool.

Automated radiology report generation (RRG) aims to produce detailed textual reports from clinical imaging, such as computed tomography (CT) scans, to improve the accuracy and efficiency of diagnosis and provision of management advice. RRG is complicated by two key challenges: (1) inherent complexity in extracting relevant information from imaging data under resource constraints, and (2) difficulty in objectively evaluating discrepancies between model-generated and expert-written reports. To address these challenges, we propose $\mu^2$LLM, a $\underline{\textbf{mu}}$ltiscale $\underline{\textbf{mu}}$ltimodal large language models for RRG tasks. The novel ${\mu}^2$Tokenizer, as an intermediate layer, integrates multi-modal features from the multiscale visual tokenizer and the text tokenizer, then enhances report generation quality through direct preference optimization (DPO), guided by GREEN-RedLlama. Experimental results on four large CT image-report medical datasets demonstrate that our method outperforms existing approaches, highlighting the potential of our fine-tuned $\mu^2$LLMs on limited data for RRG tasks. At the same time, for prompt engineering, we introduce a five-stage, LLM-driven pipeline that converts routine CT reports into paired visual-question-answer triples and citation-linked reasoning narratives, creating a scalable, high-quality supervisory corpus for explainable multimodal radiology LLM. All code, datasets, and models will be publicly available in our official repository. https://github.com/Siyou-Li/u2Tokenizer

Accurate interpretation of multi-view radiographs is crucial for diagnosing fractures, muscular injuries, and other anomalies. While significant advances have been made in AI-based analysis of single images, current methods often struggle to establish robust correspondences between different X-ray views, an essential capability for precise clinical evaluations. In this work, we present a novel self-supervised pipeline that eliminates the need for manual annotation by automatically generating a many-to-many correspondence matrix between synthetic X-ray views. This is achieved using digitally reconstructed radiographs (DRR), which are automatically derived from unannotated CT volumes. Our approach incorporates a transformer-based training phase to accurately predict correspondences across two or more X-ray views. Furthermore, we demonstrate that learning correspondences among synthetic X-ray views can be leveraged as a pretraining strategy to enhance automatic multi-view fracture detection on real data. Extensive evaluations on both synthetic and real X-ray datasets show that incorporating correspondences improves performance in multi-view fracture classification.

Current artificial intelligence models for medical imaging are predominantly single modality and single disease. Attempts to create multimodal and multi-disease models have resulted in inconsistent clinical accuracy. Furthermore, training these models typically requires large, labour-intensive, well-labelled datasets. We developed MerMED-FM, a state-of-the-art multimodal, multi-specialty foundation model trained using self-supervised learning and a memory module. MerMED-FM was trained on 3.3 million medical images from over ten specialties and seven modalities, including computed tomography (CT), chest X-rays (CXR), ultrasound (US), pathology patches, color fundus photography (CFP), optical coherence tomography (OCT) and dermatology images. MerMED-FM was evaluated across multiple diseases and compared against existing foundational models. Strong performance was achieved across all modalities, with AUROCs of 0.988 (OCT); 0.982 (pathology); 0.951 (US); 0.943 (CT); 0.931 (skin); 0.894 (CFP); 0.858 (CXR). MerMED-FM has the potential to be a highly adaptable, versatile, cross-specialty foundation model that enables robust medical imaging interpretation across diverse medical disciplines.

In the medical domain, acquiring large datasets is challenging due to both accessibility issues and stringent privacy regulations. Consequently, data availability and privacy protection are major obstacles to applying machine learning in medical imaging. To address this, our study proposes the Med-LSDM (Latent Semantic Diffusion Model), which operates directly in the 3D domain and leverages de-identified semantic maps to generate synthetic data as a method of privacy preservation and data augmentation. Unlike many existing methods that focus on generating 2D slices, Med-LSDM is designed specifically for 3D semantic image synthesis, making it well-suited for applications requiring full volumetric data. Med-LSDM incorporates a guiding mechanism that controls the 3D image generation process by applying a diffusion model within the latent space of a pre-trained VQ-GAN. By operating in the compressed latent space, the model significantly reduces computational complexity while still preserving critical 3D spatial details. Our approach demonstrates strong performance in 3D semantic medical image synthesis, achieving a 3D-FID score of 0.0054 on the conditional Duke Breast dataset and similar Dice scores (0.70964) to those of real images (0.71496). These results demonstrate that the synthetic data from our model have a small domain gap with real data and are useful for data augmentation.

Whole-heart coronary magnetic resonance angiography (CMRA) enables noninvasive and accurate detection of coronary artery stenosis. Nevertheless, the visual interpretation of CMRA is constrained by the observer's experience, necessitating substantial training. The purposes of this study were to develop a deep learning (DL) algorithm using a deep convolutional neural network to accurately detect significant coronary artery stenosis in CMRA and to investigate the effectiveness of this DL algorithm as a tool for assisting in accurate detection of coronary artery stenosis. Nine hundred and fifty-one coronary segments from 75 patients who underwent both CMRA and invasive coronary angiography (ICA) were studied. Significant stenosis was defined as a reduction in luminal diameter of >50% on quantitative ICA. A DL algorithm was proposed to classify CMRA segments into those with and without significant stenosis. A 4-fold cross-validation method was used to train and test the DL algorithm. An observer study was then conducted using 40 segments with stenosis and 40 segments without stenosis. Three radiology experts and 3 radiology trainees independently rated the likelihood of the presence of stenosis in each coronary segment with a continuous scale from 0 to 1, first without the support of the DL algorithm, then using the DL algorithm. Significant stenosis was observed in 84 (8.8%) of the 951 coronary segments. Using the DL algorithm trained by the 4-fold cross-validation method, the area under the receiver operating characteristic curve (AUC) for the detection of segments with significant coronary artery stenosis was 0.890, with 83.3% sensitivity, 83.6% specificity and 83.6% accuracy. In the observer study, the average AUC of trainees was significantly improved using the DL algorithm (0.898) compared to that without the algorithm (0.821, p<0.001). The average AUC of experts tended to be higher with the DL algorithm (0.897), but not significantly different from that without the algorithm (0.879, p=0.082). We developed a DL algorithm offering high diagnostic accuracy for detecting significant coronary artery stenosis on CMRA. Our proposed DL algorithm appears to be an effective tool for assisting inexperienced observers to accurately detect coronary artery stenosis in whole-heart CMRA.

Artificial intelligence (AI) is increasingly integrated into clinical workflows. The performance of AI in production can diverge from initial evaluations. Post-deployment monitoring (PDM) remains a challenging ingredient of ongoing quality assurance once AI is deployed in clinical production. To develop and evaluate a PDM framework that uses large language models (LLMs) for free-text classification of radiology reports, and human oversight. We demonstrate its application to monitor a commercially vended pulmonary embolism (PE) detection AI (CVPED). We retrospectively analyzed 11,999 CT pulmonary angiography (CTPA) studies performed between 04/30/2023-06/17/2024. Ground truth was determined by combining LLM-based radiology-report classification and the CVPED outputs, with human review of discrepancies. We simulated a daily monitoring framework to track discrepancies between CVPED and the LLM. Drift was defined when discrepancy rate exceeded a fixed 95% confidence interval (CI) for seven consecutive days. The CI and the optimal retrospective assessment period were determined from a stable dataset with consistent performance. We simulated drift by systematically altering CVPED or LLM sensitivity and specificity, and we modeled an approach to detect data shifts. We incorporated a human-in-the-loop selective alerting framework for continuous prospective evaluation and to investigate potential for incremental detection. Of 11,999 CTPAs, 1,285 (10.7%) had PE. Overall, 373 (3.1%) had discrepant classifications between CVPED and LLM. Among 111 CVPED-positive and LLM-negative cases, 29 would have triggered an alert due to the radiologist not interacting with CVPED. Of those, 24 were CVPED false-positives, one was an LLM false-negative, and the framework ultimately identified four true-alerts for incremental PE cases. The optimal retrospective assessment period for drift detection was determined to be two months. A 2-3% decline in model specificity caused a 2-3-fold increase in discrepancies, while a 10% drop in sensitivity was required to produce a similar effect. For example, a 2.5% drop in LLM specificity led to a 1.7-fold increase in CVPED-negative-LLM-positive discrepancies, which would have taken 22 days to detect using the proposed framework. A PDM framework combining LLM-based free-text classification with a human-in-the-loop alerting system can continuously track an image-based AI's performance, alert for performance drift, and provide incremental clinical value.