Hailong Li

Background: Endotracheal tubes (ETTs) are critical life-support devices for mechanically ventilated pediatric patients, yet automated ETT assessment on pediatric chest radiographs (CXRs) remains limited. Objective: To develop and evaluate a two-stage deep learning pipeline for automated detection and localization of ETTs on pediatric CXRs. Materials and methods: This retrospective study included 1,000 pediatric CXRs (476 ETT-positive, 524 ETT-negative) acquired in 2021 at a single institution. ETT segmentation masks and distal tip coordinates were annotated by trained analysts and verified by pediatric radiologists. A two-stage pipeline consisting of a ResNet classification model followed by a U-Net segmentation model was developed for ETT detection and localization. Performance was evaluated on a held-out test set using multiple metrics, including the area under the receiver operating characteristic curve (AUROC) and the mean absolute error (MAE), with 95% confidence intervals (CI). Inter-observer variability was assessed as a reference for localization performance. Results: Inter-observer variability for ETT tip localization was 2.01 mm MAE on the held-out test set. The pipeline achieved an AUROC of 0.994 (95% CI 0.986, 1.000) for ETT detection. For localization, the pipeline achieved a MAE of 6.59 mm (95% CI 5.19, 8.21 mm). Incorporating the classification stage substantially reduced false-positive segmentations from 17 to 3 among ETT-negative CXRs compared with the standalone segmentation model. Conclusion: A two-stage deep learning pipeline demonstrated high performance for automated ETT detection and promising performance for tip localization on pediatric CXRs in this single-center pilot study. Further evaluation in larger and external pediatric cohorts is needed to assess generalizability.

The accurate placement of lines, drains, and airways (LDAs) is essential for the safe management of critically ill patients, as malpositioned medical devices can result in severe complications ranging from ineffective treatment to life-threatening injury. Chest radiography (CXR) is the primary imaging modality for confirming the placement of many LDAs; however, the growing volume and complexity of bedside CXRs have created increasing demand for rapid, reliable interpretation. Recent advances in artificial intelligence (AI) have enabled automated detection, localization, and position assessment of LDA devices, supporting opportunities to improve clinical workflow and patient safety. This review summarizes recent developments in AI algorithms for automated LDA assessment on CXRs, emphasizing their clinical applications, technical approaches, performance, and limitations. The review begins with clinical characteristics, radiographic appearance, and placement criteria of common LDA categories, followed by a survey of AI methods for presence detection, device localization, and position classification. Although some models have achieved performance approaching that of radiologists, most do not perform at the level needed for clinical deployment. Future research should prioritize multicenter validation, standardized annotation protocols, pediatric-specific datasets, and anatomically informed models capable of simultaneously evaluating multiple LDA devices. Advances in these areas will facilitate the integration of AI-assisted LDA assessment into clinical decision support and ultimately improve patient care.

Background: Anti-tumor necrosis factor (TNF) medical therapy is a major advancement for Crohn’s disease (CD) management, yet many patients fail to respond. This study aims to develop and evaluate a multi-modal ensemble model to predict anti-TNF treatment response in pediatric patients with CD using pre-treatment magnetic resonance enterography (MRE) and non-imaging clinical data. Methods: This retrospective study included 92 pediatric CD patients (median [IQR] age, 14.6 [12.7, 16.6] years; 65.2% male; 37 responders and 55 non-responders) who underwent MRE within 3 months before initiating anti-TNF therapy between 2009 and 2021. Responders achieved mucosal healing within 36 months; non-responders did not, underwent surgery, or changed therapy. Four pre-treatment feature sets were used: (1) radiologist MRE assessment, (2) radiomic features from bowel regions, (3) deep learning features from representative bowel images, and (4) non-imaging clinical features. A 2-stage stacking ensemble model was trained and evaluated. The area under the receiver operating characteristic curve (AUROC) was the primary performance metric. Results: The multi-modal ensemble model integrating all MRE-based features achieved an AUROC of 0.771 [95% confidence intervals (CI): 0.745, 0.797], outperforming models using individual feature types (radiologist assessment: 0.706 [0.691, 0.721]; radiomic: 0.710 [0.691, 0.729]; deep learning: 0.724 [0.700, 0.749]). Combining MRE and clinical features (0.739 [0.696, 0.781]) achieved the highest AUROC of 0.817 [0.793, 0.842], though not significantly (P = .10). Radiomic features reflecting intensity distribution and texture heterogeneity were most predictive of treatment response. Conclusion: A multi-modal ensemble model combining MRE-based radiologist assessment, radiomic, deep learning, and clinical features accurately predicted anti-TNF treatment response in pediatric CD patients.

Background: Fontan-associated liver disease (FALD) is associated with morbidity and mortality in patients with palliated single ventricle congenital heart disease. Objective: To develop machine learning models using radiomic features from T1-weighted, T2-weighted, and diffusion-weighted MRI with pertinent clinical variables to predict Fontan failure and correlates of FALD severity in patients who underwent the Fontan operation. Materials and methods: In this retrospective study of abdominal MRI examinations and clinical record data from 131 Fontan palliation patients (age range 9.1 - 53.3 years old), radiomic features from the liver and spleen were extracted using axial T1-weighted, T2-weighted fat-suppressed, and diffusion-weighted sequences. Patients were categorized by a composite clinical outcome (i.e., Fontan failure) and by correlates of FALD severity, including liver shear stiffness and portal hypertension. Support vector machine (SVM) and multivariable logistic regression models were used to perform two-class classification using radiomic features and/or clinical data. All models were trained and evaluated using five-fold cross-validation (CV). Results: The best radiomic-only model utilized T2-weighted imaging of both organs with logistic regression to predict the presence of portal hypertension, achieving an AUROC of 0.85±0.01. Clinical-only models showed inferior diagnostic accuracy with the highest AUROC of 0.70±0.08. Combining radiomic and clinical features also did not enhance performance compared to radiomic-only models, with the highest AUROC of 0.77±0.05. Ensemble modeling, which incorporated radiomics from all three MRI sequences, yielded AUROCs ranging from 0.33 to 0.72. Conclusion: Models incorporating radiomic features from abdominal MRI in Fontan circulation patients demonstrate moderate diagnostic performance for predicting Fontan failure as well as correlates of FALD severity. These models outperformed models containing only clinical electronic health record data and did not improve with ensembled radiomic and clinical data. Graphical

The integration of artificial intelligence (AI) with radiology signifies a transformative era in medicine. Vision foundation models have been adopted to enhance radiologic imaging analysis. However, the inherent complexities of 2D and 3D radiologic data present unique challenges that existing models, which are typically pretrained on general nonmedical images, do not adequately address. To bridge this gap and harness the diagnostic precision required in radiologic imaging, we introduce radiologic contrastive language–image pretraining (RadCLIP): a cross-modal vision-language foundational model that utilizes a vision-language pretraining (VLP) framework to improve radiologic image analysis. Building on the contrastive language–image pretraining (CLIP) approach, RadCLIP incorporates a slice pooling mechanism designed for volumetric image analysis and is pretrained using a large, diverse dataset of radiologic image-text pairs. This pretraining effectively aligns radiologic images with their corresponding text annotations, resulting in a robust vision backbone for radiologic imaging. Extensive experiments demonstrate RadCLIP’s superior performance in both unimodal radiologic image classification and cross-modal image-text matching, underscoring its significant promise for enhancing diagnostic accuracy and efficiency in clinical settings. Our key contributions include curating a large dataset featuring diverse radiologic 2D/3D image-text pairs, pretraining RadCLIP as a vision-language foundation model on this dataset, developing a slice pooling adapter with an attention mechanism for integrating 2D images, and conducting comprehensive evaluations of RadCLIP on various radiologic downstream tasks.

Introduction Rhabdomyosarcomas are the most common soft tissue sarcoma in children. While treatment outcomes have improved, risk-based therapy classification relies on staging and tumor subtypes for therapeutic planning. Objective This study investigated the utility of T2-weighted MR radiomics features and machine learning models in identifying the presence of distant metastasis and alveolar histological subtypes at baseline imaging in children diagnosed with rhabdomyosarcoma. Materials and methods This retrospective cross-sectional study utilized MRIs from 86 patients, 49 (median age (IQR) 59 months (37-161), alveolar subtype=15, distant metastasis=9) of whom had been imaged at outside imaging centers (training set); and 37 (median age 52 months (24-164), alveolar subtype=14, distant metastasis=8) of whom were imaged at our institution (holdout validation set). Radiomic features were extracted from T2-weighted images. We selected features that demonstrated intra-scan repeatability and used maximum relevance and minimum redundancy supervised feature selection to identify the 50 most important features. Lasso logistic regression and support vector machine (SVM) classifiers were trained to predict binary outcomes. The median of all predictions for a given patient was used as patient-level predictions. DeLong’s test compared the area under the receiver operating characteristic curves (AUC). Cut-offs obtained by maximizing the Youden index were evaluated on an external validation set, and accuracy metrics were reported. Results On the validation set, the Lasso and SVM classifiers obtained patient level AUCs of 0.76 (95% CI 0.59-0.94) and 0.73 (0.54-0.92), respectively, in predicting alveolar subtype, with the Lasso regressor obtaining 71.4% (41.9-91.6) sensitivity and 60.9% (38.5-80.3) specificity. When predicting the presence of distant metastasis, the Lasso and SVM classifier had AUCs of 0.81 (0.67-0.95) and 0.77 (0.58-0.97), respectively. There were no differences between model performance (P>0.05). A total of 12 and 18 features had nonzero coefficients in the Lasso regressors for predicting alveolar subtype and tumor metastasis, respectively. Conclusion MRI radiomics from baseline T2-weighted MRI demonstrated potential in predicting alveolar subtype and distant metastatic disease at presentation. Larger studies are needed to explore multinomial multiclass models for better prognostication of pediatric rhabdomyosarcomas. Graphical Abstract