Interpretability can improve the safety, transparency and trust of AI models, which is especially important in healthcare applications where decisions often carry significant consequences. Mechanistic interpretability, particularly through the use of sparse autoencoders (SAEs), offers a promising approach for uncovering human-interpretable features within large transformer-based models. In this study, we apply Matryoshka-SAE to the radiology-specialised multimodal large language model, MAIRA-2, to interpret its internal representations. Using large-scale automated interpretability of the SAE features, we identify a range of clinically relevant concepts - including medical devices (e.g., line and tube placements, pacemaker presence), pathologies such as pleural effusion and cardiomegaly, longitudinal changes and textual features. We further examine the influence of these features on model behaviour through steering, demonstrating directional control over generations with mixed success. Our results reveal practical and methodological challenges, yet they offer initial insights into the internal concepts learned by MAIRA-2 - marking a step toward deeper mechanistic understanding and interpretability of a radiology-adapted multimodal large language model, and paving the way for improved model transparency. We release the trained SAEs and interpretations: https://huggingface.co/microsoft/maira-2-sae.

BackgroundSubjective cognitive decline (SCD) is recognized as an early phase in the progression of Alzheimer's disease (AD).ObjectiveTo explore the abnormal patterns of morphological and functional connectivity coupling (MC-FC coupling) and their potential diagnostic significance in SCD.MethodsThe data of 52 individuals with SCD and 51 age-gender-education matched healthy controls (HC) who underwent resting-state functional magnetic resonance imaging and high-resolution 3D T₁-weighted imaging were retrieved to build the MC and FC of gray matter. Support vector machine (SVM) methods were used for differentiating between SCD and HC.ResultsSCD individuals exhibited MC-FC decoupling in the frontoparietal network compared with HC (p = 0.002, 5000 permutations). Using these adjusted MC-FC coupling metrics, SVM analysis achieved 74.76% accuracy, 64.71% sensitivity, and 92.31% specificity (p < 0.001, 5000 permutations). Additionally, the stronger MC-FC coupling of the left inferior temporal gyrus (r = 0.294, p = 0.034) and right posterior cingulate gyrus (r = 0.372, p = 0.007) in SCD individuals was positively correlated with subjective memory complaint performance.ConclusionsThe findings of this study provide insight into the idiosyncratic feature of brain organization underlying SCD from the prospective of MC-FC coupling and highlight the potential of MC-FC coupling for the identification of the preclinical stage of AD.

Conservative treatment remains a viable option for selected patients with ectopic pregnancy (EP), but failure may lead to rupture and serious complications. Currently, serum β-hCG is the main predictor for treatment outcomes, yet its accuracy is limited. This study aimed to develop and validate a predictive model that integrates radiomic features derived from super-resolution (SR) ultrasound images with clinical biomarkers to improve risk stratification. A total of 228 patients with EP receiving conservative treatment were retrospectively included, with 169 classified as treatment success and 59 as failure. SR images were generated using a deep learning-based generative adversarial network (GAN). Radiomic features were extracted from both normal-resolution (NR) and SR ultrasound images. Features with intraclass correlation coefficient (ICC) ≥ 0.75 were retained after intra- and inter-observer evaluation. Feature selection involved statistical testing and Least Absolute Shrinkage and Selection Operator (LASSO) regression. Random forest algorithms were used to construct NR and SR models. A clinical model based on serum β-hCG was also developed. The Clin-SR model was constructed by fusing SR radiomics with β-hCG values. Model performance was evaluated using area under the curve (AUC), calibration, and decision curve analysis (DCA). An independent temporal validation cohort (n = 40; 20 failures, 20 successes) was used to validation of the nomogram derived from the Clin-SR model. The SR model significantly outperformed the NR model in the test cohort (AUC: 0.791 ± 0.015 vs. 0.629 ± 0.083). In a representative iteration, the Clin-SR fusion model achieved an AUC of 0.870 ± 0.015, with good calibration and net clinical benefit, suggesting reliable performance in predicting conservative treatment failure. In the independent validation cohort, the nomogram demonstrated good generalizability with an AUC of 0.808 and consistent calibration across risk thresholds. Key contributing radiomic features included Gray Level Variance and Voxel Volume, reflecting lesion heterogeneity and size. The Clin-SR model, which integrates deep learning-enhanced SR ultrasound radiomics with serum β-hCG, offers a robust and non-invasive tool for predicting conservative treatment failure in ectopic pregnancy. This multimodal approach enhances early risk stratification and supports personalized clinical decision-making, potentially reducing overtreatment and emergency interventions.

CT-based opportunistic screening using artificial intelligence finds a high prevalence (43%) of osteoporosis in CT scans obtained for planning of transcatheter aortic valve replacement. Thus, opportunistic screening may be a cost-effective way to assess osteoporosis in high-risk populations. Osteoporosis is an underdiagnosed condition associated with fractures and frailty, but may be detected in routine computed tomography (CT) scans. Volumetric bone mineral density (vBMD) was measured in clinical routine thoraco-abdominal CT scans of 207 patients for planning of transcatheter aortic valve replacement (TAVR) using an artificial intelligence (AI)-based algorithm. 43% of patients had osteoporosis (vBMD < 80 mg/cm³ L1-L3) and were elderly (83.0 {interquartile range [IQR]: 78.0-85.5} vs. 79.0 {IQR: 71.8-84.0} years, p < 0.001), more often female (55.1 vs. 28.8%, p < 0.001), and had a higher Society of Thoracic Surgeon's score for mortality (3.0 {IQR:1.8-4.6} vs. 2.1 {IQR: 1.4-3.2}%, p < 0.001). In addition to lumbar vBMD (58.2 ± 14.7 vs. 106 ± 21.4 mg/cm³, p < 0.001), thoracic vBMD (79.5 ± 17.9 vs. 127.4 ± 26.0 mg/cm³, p < 0.001) was also significantly reduced in these patients and showed high diagnostic accuracy for osteoporosis assessment (area under curve: 0.96, p < 0.001). Osteoporotic patients were significantly more often at risk for falls (40.4 vs. 22.9%, p = 0.007) and required help in activities of daily life (ADL) more frequently (48.3 vs. 33.1%, p = 0.026), while direct-to-home discharges were fewer (88.8 vs. 96.6%, p = 0.026). In-hospital bleeding complications (3.4 vs. 5.1%), stroke (1.1 vs. 2.5%), and death (1.1 vs. 0.8%) were equally low, while in-hospital device success was equally high (94.4 vs. 94.9%, p > 0.05 for all comparisons). However, one-year probability of survival was significantly lower (84.0 vs. 98.2%, log-rank p < 0.01). Applying an AI-based algorithm to TAVR planning CT scans can reveal a high rate of 43% patients having osteoporosis. Osteoporosis may represent a marker related to frailty and worsened outcome in TAVR patients.

Recent developments in the registration of histology and micro-computed tomography (µCT) have broadened the perspective of pathological applications such as virtual histology based on µCT. This topic remains challenging because of the low image quality of soft tissue CT. Additionally, soft tissue samples usually deform during the histology slide preparation, making it difficult to correlate the structures between the histology slide and µCT. In this work, we propose a novel 2D-3D multi-modal deformable image registration method. The method utilizes an initial global 2D-3D registration using an ML-based differentiable similarity measure. The registration is then finalized by an analytical out-of-plane deformation refinement. The method is evaluated on datasets acquired from tonsil and tumor tissues. µCTs of both phase-contrast and conventional absorption modalities are investigated. The registration results from the proposed method are compared with those from intensity- and keypoint-based methods. The comparison is conducted using both visual and fiducial-based evaluations. The proposed method demonstrates superior performance compared to the other two methods.

B-ultrasound results are widely used in early pregnancy loss (EPL) prediction, but there are inevitable intra-observer and inter-observer errors in B-ultrasound results especially in early pregnancy, which lead to inconsistent assessment of embryonic status, and thus affect the judgment of EPL. To address this, we need a rapid and accurate model to predict pregnancy loss in the first trimester. This study aimed to construct an artificial intelligence model to automatically extract biometric parameters from ultrasound videos of early embryos and predict pregnancy loss. This can effectively eliminate the measurement error of B-ultrasound results, accurately predict EPL, and provide decision support for doctors with relatively little clinical experience. A total of 630 ultrasound videos from women with early singleton pregnancies of gestational age between 6 and 10 weeks were used for training. A two-stage artificial intelligence model was established. First, some biometric parameters such as gestational sac areas (GSA), yolk sac diameter (YSD), crown rump length (CRL) and fetal heart rate (FHR), were extract from ultrasound videos by a deep neural network named A3F-net, which is a modified neural network based on U-Net designed by ourselves. Then an ensemble learning model predicted pregnancy loss risk based on these features. Dice, IOU and Precision were used to evaluate the measurement results, and sensitivity, AUC etc. were used to evaluate the predict results. The fetal heart rate was compared with those measured by doctors, and the accuracy of results was compared with other AI models. In the biometric features measurement stage, the precision of GSA, YSD and CRL of A3F-net were 98.64%, 96.94% and 92.83%, it was the highest compared to other 2 models. Bland-Altman analysis did not show systematic deviations between doctors and AI. The mean and standard deviation of the mean relative error between doctors and the AI model was 0.060 ± 0.057. In the EPL prediction stage, the ensemble learning models demonstrated excellent performance, with CatBoost being the best-performing model, achieving a precision of 98.0% and an AUC of 0.969 (95% CI: 0.962-0.975). In this study, a hybrid AI model to predict EPL was established. First, a deep neural network automatically measured the biometric parameters from ultrasound video to ensure the consistency and accuracy of the measurements, then a machine learning model predicted EPL risk to support doctors making decisions. The use of our established AI model in EPL prediction has the potential to assist physicians in making more accurate and timely clinical decision in clinical application.

In high energy density physics (HEDP) and inertial confinement fusion (ICF), predictive modeling is complicated by uncertainty in parameters that characterize various aspects of the modeled system, such as those characterizing material properties, equation of state (EOS), opacities, and initial conditions. Typically, however, these parameters are not directly observable. What is observed instead is a time sequence of radiographic projections using X-rays. In this work, we define a set of sparse hydrodynamic features derived from the outgoing shock profile and outer material edge, which can be obtained from radiographic measurements, to directly infer such parameters. Our machine learning (ML)-based methodology involves a pipeline of two architectures, a radiograph-to-features network (R2FNet) and a features-to-parameters network (F2PNet), that are trained independently and later combined to approximate a posterior distribution for the parameters from radiographs. We show that the machine learning architectures are able to accurately infer initial conditions and EOS parameters, and that the estimated parameters can be used in a hydrodynamics code to obtain density fields, shocks, and material interfaces that satisfy thermodynamic and hydrodynamic consistency. Finally, we demonstrate that features resulting from an unknown EOS model can be successfully mapped onto parameters of a chosen analytical EOS model, implying that network predictions are learning physics, with a degree of invariance to the underlying choice of EOS model. To the best of our knowledge, our framework is the first demonstration of recovering both thermodynamic and hydrodynamic consistent density fields from noisy radiographs.

This study aims to explore the role of intra- and peri-tumoral radiomics features in tumor risk prediction, with a particular focus on the impact of peri-tumoral characteristics on the tumor microenvironment. A total of 133 patients, including 128 with thymomas and 5 with thymic carcinomas, were ultimately enrolled in this study. Based on the high- and low-risk classification, the cohort was divided into a training set (n = 93) and a testing set (n = 40) for subsequent analysis.Based on imaging data from these 133 patients, multiple radiomics prediction models integrating intra-tumoral and peritumoral features were developed. The data were sourced from patients treated at the Affiliated Hospital of Guangdong Medical University between 2015 and 2023, with all imaging obtained through preoperative CT scans. Radiomics feature extraction involved three primary categories: first-order features, shape features, and high-order features. Initially, the tumor's region of interest (ROI) was manually delineated using ITK-SNAP software. A custom Python algorithm was then used to automatically expand the peri-tumoral area, extracting features within 1 mm, 2 mm, and 3 mm zones surrounding the tumor. Additionally, considering the multimodal nature of the imaging data, image fusion techniques were incorporated to further enhance the model's ability to capture the tumor microenvironment. To build the radiomics models, selected features were first standardized using z-scores. Initial feature selection was performed using a t-test (p < 0.05), followed by Spearman correlation analysis to remove redundancy by retaining only one feature from each pair with a correlation coefficient ≥ 0.90. Subsequently, hierarchical clustering and the LASSO algorithm were applied to identify the most predictive features. These selected features were then used to train machine learning models, which were optimized on the training dataset and assessed for predictive performance. To further evaluate the effectiveness of these models, various statistical methods were applied, including DeLong's test, NRI, and IDI, to compare predictive differences among models. Decision curve analysis (DCA) was also conducted to assess the clinical applicability of the models. The results indicate that the IntraPeri1mm model performed the best, achieving an AUC of 0.837, with sensitivity and specificity at 0.846 and 0.84, respectively, significantly outperforming other models. SHAP value analysis identified several key features, such as peri_log_sigma_2_0_mm 3D_firstorder RootMeanSquared and intra_wavelet_LLL_firstorder Skewness, which made substantial contributions to the model's predictive accuracy. NRI and IDI analyses further confirmed the model's superior clinical applicability, and the DCA curve demonstrated robust performance across different thresholds. DeLong's test highlighted the statistical significance of the IntraPeri1mm model, underscoring its potential utility in radiomics research. Overall, this study provides a new perspective on tumor risk assessment, highlighting the importance of peri-tumoral features in the analysis of the tumor microenvironment. It aims to offer valuable insights for the development of personalized treatment plans. Not applicable.

Attention-deficit/hyperactivity disorder (ADHD) is a common neurodevelopmental disorder that affects behavior, attention, and learning. Current diagnoses rely heavily on subjective assessments, underscoring the need for objective imaging-based methods. This study aims to explore whether structural connectivity networks derived from MRI can reveal alterations associated with ADHD and support data-driven understanding. We collected brain MRI data from 947 individuals (aged 7-26 years; 590 males, 356 females, 1 unspecified) across eight centers, sourced from the Neuro Bureau ADHD-200 preprocessed dataset. Transformer-based deep learning models were used to learn relationships between different brain regions and construct structural connectivity networks. To prepare input for the model, each region was transformed into a standardized data sequence using four different strategies. The strength of connectivity between brain regions was then measured to identify structural differences related to ADHD. Five-fold cross-validation and statistical analyses were used to evaluate model robustness and group differences, respectively. Here we show that the proposed method performs well in distinguishing ADHD individuals from healthy controls, with accuracy reaching 71.9 percent and an area under curve of 0.74. The structural networks also reveal significant differences in connectivity patterns (paired t-test: P = 0.81 × 10^-6), particularly involving regions responsible for motor and executive function. Notably, the importance rankings of several brain regions, including the thalamus and caudate, differ markedly between groups. This study shows that ADHD may be associated with connectivity alterations in multiple brain regions. Our findings suggest that brain structural connectivity networks built using Transformer-based methods offer a promising tool for both diagnosis and further research into brain structure.

Sparse-view computed tomography (CT) holds promise for reducing radiation exposure and enabling novel system designs. Traditional reconstruction algorithms, including Filtered Backprojection (FBP) and Model-Based Iterative Reconstruction (MBIR), often produce artifacts in sparse-view data. Deep Learning Reconstruction (DLR) offers potential improvements, but task-based evaluations of DLR in sparse-view CT remain limited. This study employs an Artificial Intelligence (AI) observer to evaluate the diagnostic accuracy of FBP, MBIR, and DLR for intracranial hemorrhage detection and classification, offering a cost-effective alternative to human radiologist studies. A public brain CT dataset with labeled intracranial hemorrhages was used to train an AI observer model. Sparse-view CT data were simulated, with reconstructions performed using FBP, MBIR, and DLR. Reconstruction quality was assessed using metrics such as Peak Signal-to-Noise Ratio (PSNR), Structural Similarity Index Measure (SSIM), and Learned Perceptual Image Patch Similarity (LPIPS). Diagnostic utility was evaluated using Receiver Operating Characteristic (ROC) analysis and Area Under the Curve (AUC) values for One-vs-Rest and One-vs-One classification tasks. DLR outperformed FBP and MBIR in all quality metrics, demonstrating reduced noise, improved structural similarity, and fewer artifacts. The AI observer achieved the highest classification accuracy with DLR, while FBP surpassed MBIR in task-based accuracy despite inferior image quality metrics, emphasizing the value of task-based evaluations. DLR provides an effective balance of artifact reduction and anatomical detail in sparse-view CT brain imaging. This proof-of-concept study highlights AI observer models as a viable, cost-effective alternative for evaluating CT reconstruction techniques.