Urinary tract dilation (UTD) is a frequent problem in infants. Automated and objective classification of UTD from renal ultrasounds would streamline their interpretations. To develop and evaluate the performance of different deep learning models in predicting UTD classifications from renal ultrasound images. We searched our image archive to identify renal ultrasounds performed in infants ≤ 3-months-old for the clinical indications of prenatal UTD and urinary tract infection (9/2023-8/2024). An expert pediatric uroradiologist provided the ground truth UTD labels for representative sagittal sonographic renal images. Three different deep learning models trained with cross-entropy loss were adapted with four-fold cross-validation experiments to determine the overall performance. Our curated database included 492 right and 487 left renal ultrasounds (mean age ± standard deviation = 1.2 ± 0.1 months for both cohorts, with 341 boys/151 girls and 339 boys/148 girls, respectively). The model prediction accuracies for the right and left kidneys were 88.7% (95% confidence interval [CI], [85.8%, 91.5%]) and 80.5% (95% CI, [77.6%, 82.9%]), with weighted kappa scores of 0.90 (95% CI, [0.88, 0.91]) and 0.87 (95% CI, [0.82, 0.92]), respectively. When predictions were binarized into mild (normal/P1) and severe (UTD P2/P3) dilation, accuracies of the right and left kidneys increased to 96.3% (95% CI, [94.9%, 97.8%]) and 91.3% (95% CI, [88.5%, 94.2%]), but agreements decreased to 0.78 (95% CI, [0.73, 0.82]) and 0.75 (95% CI, [0.68, 0.82]), respectively. Deep learning models demonstrated high accuracy and agreement in classifying UTD from infant renal ultrasounds, supporting their potential as decision-support tools in clinical workflows.

To develop a multi-modality machine learning-based radiomics model utilizing Magnetic Resonance Imaging (MRI), Ultrasound (US), and Mammography (MMG) for the differentiation of benign and malignant breast nodules. This study retrospectively collected data from 204 patients across three hospitals, including MRI, US, and MMG imaging data along with confirmed pathological diagnoses. Lesions on 2D US, 2D MMG, and 3D MRI images were selected to outline the areas of interest, which were then automatically expanded outward by 3 mm, 5 mm, and 8 mm to extract radiomic features within and around the tumor. ANOVA, the maximum correlation minimum redundancy (mRMR) algorithm, and the least absolute shrinkage and selection operator (LASSO) were used to select features for breast cancer diagnosis through logistic regression analysis. The performance of the radiomics models was evaluated using receiver operating characteristic (ROC) curve analysis, curves decision curve analysis (DCA), and calibration curves. Among the various radiomics models tested, the MRI_US_MMG multi-modality logistic regression model with 5 mm peritumoral features demonstrated the best performance. In the test cohort, this model achieved an AUC of 0.905(95% confidence interval [CI]: 0.805-1). These results suggest that the inclusion of peritumoral features, specifically at a 5 mm expansion, significantly enhanced the diagnostic efficiency of the multi-modality radiomics model in differentiating benign from malignant breast nodules. The multi-modality radiomics model based on MRI, ultrasound, and mammography can predict benign and malignant breast lesions.

Aberrant functional connectivity (FC) between brain networks has been indicated closely associated with bipolar disorder (BD). However, the previous findings of specific brain network connectivity patterns have been inconsistent, and the clinical utility of FCs for predicting treatment outcomes in bipolar depression was underexplored. To identify robust neuro-biomarkers of bipolar depression, a connectome-based analysis was conducted on resting-state functional MRI (rs-fMRI) data of 580 bipolar depression patients and 116 healthy controls (HCs). A subsample of 148 patients underwent a 4-week quetiapine treatment with post-treatment clinical assessment. Adopting machine learning, a predictive model based on pre-treatment brain connectome was then constructed to predict treatment response and identify the efficacy-specific networks. Distinct brain network connectivity patterns were observed in bipolar depression compared to HCs. Elevated intra-network connectivity was identified within the default mode network (DMN), sensorimotor network (SMN), and subcortical network (SC); and as to the inter-network connectivity, increased FCs were between the DMN, SMN and frontoparietal (FPN), ventral attention network (VAN), and decreased FCs were between the SC and cortical networks, especially the DMN and FPN. And the global network topology analyses revealed decreased global efficiency and increased characteristic path length in BD compared to HC. Further, the support vector regression model successfully predicted the efficacy of quetiapine treatment, as indicated by a high correspondence between predicted and actual HAMD reduction ratio values (r_(df=147)=0.4493, p = 2*10^-4). The identified efficacy-specific networks primarily encompassed FCs between the SMN and SC, and between the FPN, DMN, and VAN. These identified networks further predicted treatment response with r = 0.3940 in the subsequent validation with an independent cohort (n = 43). These findings presented the characteristic aberrant patterns of brain network connectome in bipolar depression and demonstrated the predictive potential of pre-treatment network connectome for quetiapine response. Promisingly, the identified connectivity networks may serve as functional targets for future precise treatments for bipolar depression.

Brain tumours (BTs) are severe neurological disorders. They affect more than 308,000 people each year worldwide. The mortality rate is over 251,000 deaths annually (IARC, 2020 reports). Detecting BTs is complex because they vary in nature. Early diagnosis is essential for better survival rates. The study presents a new system for detecting BTs. It combines deep (DL) learning and machine (ML) learning techniques. The system uses advanced models like Inception-V3, ResNet-50, and VGG-16 for feature extraction, and for dimensional reduction, it uses the PCA model. It also employs ensemble methods such as Stacking, k-NN, Gradient Boosting, AdaBoost, Multi-Layer Perceptron (MLP), and Support Vector Machines for classification and predicts the BTs using MRI scans. The MRI scans were resized to 224 × 224 pixels, and pixel intensities were normalized to a [0,1] scale. We apply the Gaussian filter for stability. We use the Keras Image Data Generator for image augmentation. It applied methods like zooming and ± 10% brightness adjustments. The dataset has 5,712 MRI scans. These scans are classified into four groups: Meningioma, No-Tumor, Glioma, and Pituitary. A tenfold cross-validation method helps check if the model is reliable. Deep transfer (TL) learning and ensemble ML models work well together. They showed excellent results in detecting BTs. The stacking ensemble model achieved the highest accuracy across all feature extraction methods, with ResNet-50 features reduced by PCA (500), producing an accuracy of 0.957, 95% CI: 0.948-0.966; AUC: 0.996, 95% CI: 0.989-0.998, significantly outperforming baselines (p < 0.01). Neural networks and gradient-boosting models also show strong performance. The stacking model is robust and reliable. This method is useful for medical applications. Future studies will focus on using multi-modal imaging. This will help improve diagnostic accuracy. The research improves early detection of brain tumors.

Multiparametric MRI (mpMRI) has become an essential tool in the detection of prostate cancer (PCa) and can help many men avoid unnecessary biopsies. However, interpreting prostate mpMRI remains subjective, labor-intensive, and more complex compared to traditional transrectal ultrasound. These challenges will likely grow as MRI is increasingly adopted for PCa screening and diagnosis. This development has sparked interest in non-invasive artificial intelligence (AI) support, as larger and better-labeled datasets now enable deep-learning (DL) models to address important tasks in the prostate MRI workflow. Specifically, DL classification networks can be trained to differentiate between benign tissue and PCa, identify non-clinically significant disease versus clinically significant disease, and predict high-grade cancer at both the lesion and patient levels. This review focuses on the integration of DL classification networks with mpMRI for PCa assessment, examining key network architectures and strategies, the impact of different MRI sequence inputs on model performance, and the added value of incorporating domain knowledge and clinical information into MRI-based DL classifiers. It also highlights reported comparisons between DL models and the Prostate Imaging Reporting and Data System (PI-RADS) for PCa diagnosis and the potential of AI-assisted predictions, alongside ongoing efforts to improve model explainability and interpretability to support clinical trust and adoption. It further discusses the potential role of DL-based computer-aided diagnosis systems in improving the prostate MRI reporting workflow while addressing current limitations and future outlooks to facilitate better clinical integration of these systems. Evidence Level: N/A. Technical Efficacy: Stage 2.

To develop an interpretable machine learning model, explained using SHAP, based on imaging features of adolescent idiopathic scoliosis extracted by convolutional neural networks (CNNs), in order to predict the risk of curve progression and identify the most accurate predictive model. This study included 233 patients with adolescent idiopathic scoliosis from three medical centers. CNNs were used to extract features from full-spine coronal X-ray images taken at three follow-up points for each patient. Imaging and clinical features from center 1 were analyzed using the Boruta algorithm to identify independent predictors. Data from center 1 were divided into training (80%) and testing (20%) sets, while data from centers 2 and 3 were used as external validation sets. Six machine learning models were constructed. Receiver operating characteristic (ROC) curves were plotted, and model performance was assessed by calculating the area under the curve (AUC), accuracy, sensitivity, and specificity in the training, testing, and external validation sets. The SHAP interpreter was used to analyze the most effective model. The six models yielded AUCs ranging from 0.565 to 0.989, accuracies from 0.600 to 0.968, sensitivities from 0.625 to 1.0, and specificities from 0.571 to 0.974. The XGBoost model achieved the best performance, with an AUC of 0.896 in the external validation set. SHAP analysis identified the change in the main Cobb angle between the second and first follow-ups [Cobb1(2−1)] as the most important predictor, followed by the main Cobb angle at the second follow-up (Cobb1-2) and the change in the secondary Cobb angle [Cobb2(2−1)]. The XGBoost model demonstrated the best predictive performance in the external validation cohort, confirming its preliminary stability and generalizability. SHAP analysis indicated that Cobb1(2−1) was the most important feature for predicting scoliosis progression. This model offers a valuable tool for clinical decision-making by enabling early identification of high-risk patients and supporting early intervention strategies through automated feature extraction and interpretable analysis. The online version contains supplementary material available at 10.1186/s12891-025-08841-3.

Artificial intelligence (AI) holds significant promise for medical applications, particularly in coronary computed tomography angiography (CTA). We assessed the knowledge, attitudes, and practices (KAP) of cardiovascular health care personnel regarding coronary CTA and AI-assisted diagnosis. We conducted a cross-sectional survey from 1 July to 1 August 2024 at Tsinghua University Hospital, Beijing, China. Healthcare professionals, including both physicians and nurses, aged ≥18 years were eligible to participate. We used a structured questionnaire to collect demographic information and KAP scores. We analysed the data using correlation and regression methods, along with structural equation modelling. Among 496 participants, 58.5% were female, 52.6% held a bachelor's degree, and 40.7% worked in radiology. Mean KAP scores were 13.87 (standard deviation (SD) = 4.96, possible range = 0-20) for knowledge, 28.25 (SD = 4.35, possible range = 8-40) for attitude, and 31.67 (SD = 8.23, possible range = 10-50) for practice. Knowledge (r = 0.358; P < 0.001) and attitude positively correlated with practice (r = 0.489; P < 0.001). Multivariate logistic regression indicated that educational level, department affiliation, and job satisfaction were significant predictors of knowledge. Attitude was influenced by marital status, department, and years of experience, while practice was shaped by knowledge, attitude, departmental factors, and job satisfaction. Structural equation modelling showed that knowledge was directly affected by gender (β = -0.121; P = 0.009), workplace (β = -0.133; P = 0.004), department (β = -0.197; P < 0.001), employment status (β = -0.166; P < 0.001), and night shift frequency (β = 0.163; P < 0.001). Attitude was directly influenced by marriage (β = 0.124; P = 0.006) and job satisfaction (β = -0.528; P < 0.001). Practice was directly affected by knowledge (β = 0.389; P < 0.001), attitude (β = 0.533; P < 0.001), and gender (β = -0.092; P = 0.010). Additionally, gender (β = -0.051; P = 0.010) and marriage (β = 0.066; P = 0.007) had indirect effects on practice. Cardiovascular health care personnel exhibited suboptimal knowledge, positive attitudes, and relatively inactive practices regarding coronary CTA and AI-assisted diagnosis. Targeted educational efforts are needed to enhance knowledge and support the integration of AI into clinical workflows.

Recent advances in the treatment of acute ischemic stroke contribute to improved patient outcomes, yet the mechanisms driving long-term disease trajectory are not well-understood. Current trends in the literature emphasize the distributed disruptive impact of stroke lesions on brain network organization. While most studies use population-derived data to investigate lesion interference on healthy tissue, the potential for individualized treatment strategies remains underexplored due to a lack of availability and effective utilization of the necessary clinical imaging data. To validate the potential for individualized patient evaluation, we explored and compared the differential information in network models based on normative and individual data. We further present our novel deep learning approach providing usable and accurate estimates of individual stroke impact utilizing minimal imaging data, thus bridging the data gap hindering individualized treatment planning. We created normative and individual disconnectomes for each of 78 patients (mean age 65.1 years, 32 females) from two independent cohort studies. MRI data and Barthel Index, as a measure of activities of daily living, were collected in the acute and early sub-acute phase after stroke (baseline) and at three months post stroke incident. Disconnectomes were subsequently described using 12 network metrics, including clustering coefficient and transitivity. Metrics were first compared between disconnectomes and further utilized as features in a classifier to predict a patients disease trajectory, as defined by three months Barthel Index. We then developed a deep learning architecture based on graph convolution and trained it to predict properties of the individual disconnectomes from the normative disconnectomes. Both disconnectomes showed statistically significant differences in topology and predictive power. Normative disconnectomes included a statistically significant larger number of connections (N=604 for normative versus N=210 for individual) and agreement between network properties ranged from r2=0.01 for clustering coefficient to r2=0.8 for assortativity, highlighting the impact of disconnectome choice on subsequent analysis. To predict patient deficit severity, individual data achieved an AUC score of 0.94 compared to an AUC score of 0.85 for normative based features. Our deep learning estimates showed high correlation with individual features (mean r2=0.94) and a comparable performance with an AUC score of 0.93. We were able to show how normative data-based analysis of stroke disconnections provides limited information regarding patient recovery. In contrast, individual data provided higher prognostic precision. We presented a novel approach to curb the need for individual data while retaining most of the differential information encoding individual patient disease trajectory.

Mechanical thrombectomy has become the standard of care in patients with stroke due to large vessel occlusion (LVO). However, only 50% of successfully treated patients show a favorable outcome. We developed and evaluated interpretable deep learning models to predict functional outcomes in terms of the modified Rankin Scale score alongside individualized treatment effects (ITEs) using data of 449 LVO stroke patients from a randomized clinical trial. Besides clinical variables, we considered non-contrast CT (NCCT) and angiography (CTA) scans which were integrated using novel foundation models to make use of advanced imaging information. Clinical variables had a good predictive power for binary functional outcome prediction (AUC of 0.719 [0.666, 0.774]) which could slightly be improved when adding CTA imaging (AUC of 0.737 [0.687, 0.795]). Adding NCCT scans or a combination of NCCT and CTA scans to clinical features yielded no improvement. The most important clinical predictor for functional outcome was pre-stroke disability. While estimated ITEs were well calibrated to the average treatment effect, discriminatory ability was limited indicated by a C-for-Benefit statistic of around 0.55 in all models. In summary, the models allowed us to jointly integrate CT imaging and clinical features while achieving state-of-the-art prediction performance and ITE estimates. Yet, further research is needed to particularly improve ITE estimation.

Multimodal deep learning holds promise for improving clinical prediction by integrating diverse patient data, including text, imaging, time-series, and structured demographics. Contrastive learning facilitates this integration by producing a unified representation that can be reused across tasks, reducing the need for separate models or encoders. Although contrastive learning has seen success in vision-language domains, its use in clinical settings remains largely limited to image and text pairs. We propose the Pipeline for Contrastive Modality Evaluation and Encoding (PiCME), which systematically assesses five clinical data types from MIMIC: discharge summaries, radiology reports, chest X-rays, demographics, and time-series. We pre-train contrastive models on all 26 combinations of two to five modalities and evaluate their utility on in-hospital mortality and phenotype prediction. To address performance plateaus with more modalities, we introduce a Modality-Gated LSTM that weights each modality according to its contrastively learned importance. Our results show that contrastive models remain competitive with supervised baselines, particularly in three-modality settings. Performance declines beyond three modalities, which supervised models fail to recover. The Modality-Gated LSTM mitigates this drop, improving AUROC from 73.19% to 76.93% and AUPRC from 51.27% to 62.26% in the five-modality setting. We also compare contrastively learned modality importance scores with attribution scores and evaluate generalization across demographic subgroups, highlighting strengths in interpretability and fairness. PiCME is the first to scale contrastive learning across all modality combinations in MIMIC, offering guidance for modality selection, training strategies, and equitable clinical prediction.