PURPOSE: To develop and validate a multimodal deep learning framework that integrates clinical metadata with [18F]FDG PET/CT imaging to resolve overlapping metabolic phenotypes. The primary objective is the histological subtyping of non-small cell lung cancer (NSCLC), utilizing binary clinical staging (early vs. advanced) strategically as an auxiliary regularization task. METHODS: A multi-center surgical NSCLC cohort (n = 780) was partitioned into a development set (n = 675) and an independent external test set (n = 105). The framework first utilized a 3D Transformer for bounding-box-based tumor localization. Subsequently, a multi-task network employed Feature-wise Linear Modulation (FiLM) to dynamically inject clinical metadata into the visual backbone. RESULTS: For histological subtyping of adenocarcinoma versus squamous cell carcinoma, in the validation cohort, the proposed multimodal framework achieved the highest area under the receiver operating characteristic curve (AUC) of 0.894 (95% CI: 0.813-0.959), significantly outperforming the conventional radiomics baseline (AUC = 0.796, DeLong test P = 0.017) and the clinical-only baseline (AUC = 0.759, P = 0.004). On the internal test set, the multimodal model maintained an AUC of 0.832 (95% CI: 0.744-0.906), outperforming competing models numerically, though differences did not reach statistical significance (all P > 0.11). On the independent external test cohort, the multimodal framework demonstrated superior cross-center stability, maintaining an AUC of 0.787 (95% CI: 0.687-0.876). On the external cohort, the between-model AUC differences did not reach statistical significance against the clinical-only model (AUC of 0.740, P = 0.480) or the image-only model (AUC of 0.685, P = 0.082). Nevertheless, the multimodal framework achieved the highest F1-score and yielded the most optimal net clinical benefit across a wide range of threshold probabilities in decision curve analysis. For the intrinsically challenging auxiliary staging task, the unguided image-only network exhibited severe vulnerability, however, the FiLM-based multimodal mechanism effectively enhanced diagnostic capacity by employing systemic clinical priors, improving the AUC to 0.656. CONCLUSION: Combining 3D detection with an early clinico-biological fusion strategy effectively enhances NSCLC characterization on [18F]FDG PET/CT, which has the potential to mitigate the limitations of single-modality imaging in resolving diagnostically ambiguous cases characterized by overlapping [18F]FDG uptake phenotypes, thereby providing a non-invasive decision-support tool in the precision management of NSCLC.

This study aimed to evaluate the prognostic value of conventional and advanced PET metrics for predicting progression-free survival in high-risk pediatric Hodgkin lymphoma (HL), using data from the Children's Oncology Group AHOD1331 trial. Methods: This retrospective analysis included 558 patients from 150 institutions. Conventional PET metrics and radiomics features were extracted from both baseline and interim-therapy PET images. Clinical variables, including demographic characteristics and common risk factors, were also considered. A standardized outcome-modeling pipeline was developed, comprising feature selection and penalized Cox regression with elastic net regularization (CoxNet). Models were trained and evaluated using nested cross-validation stratified by institution, ensuring consistent external testing. We investigated whether conventional PET metrics added prognostic value beyond clinical variables and whether radiomics or interim PET-derived features offered further benefit. Additionally, we compared the prognostic performance of features extracted from automated deep learning (DL) segmentations against those derived from physician annotations. Model performance was assessed using the concordance index (C-index). Results: Among all types of features, the CoxNet model using conventional baseline PET metrics achieved the highest performance (C-index, 0.72 ± 0.01), significantly outperforming the clinical model (C-index, 0.65 ± 0.01). Neither the inclusion of radiomics features nor interim PET-derived features improved performance. Models using DL-generated segmentations achieved comparable prognostic accuracy (C-index, 0.72 ± 0.01) to those using physician-based segmentations. Conclusion: Quantitative PET features provided significant prognostic improvements over clinical variables. Furthermore, DL-based segmentation offered a promising approach for automating feature extraction, supporting more effective risk stratification in pediatric HL.

Foundation models are becoming increasingly effective in the medical domain, offering pre-trained models on large datasets that can be readily adapted for downstream tasks. Despite progress, fetal ultrasound images remain a challenging domain for visual-language foundation models due to their inherent complexity, often requiring substantial additional training and facing limitations due to the scarcity of paired multimodal data. To overcome these challenges, here we introduce FetalCLIP, a vision-language foundation model capable of generating universal representation of fetal ultrasound images. FetalCLIP was pre-trained using a multimodal learning approach on a diverse dataset of 210,035 fetal ultrasound images paired with text. This represents the largest paired dataset of its kind used for visual-language foundation model development to date. This unique training approach allows FetalCLIP to effectively learn the intricate anatomical features present in fetal ultrasound images, resulting in robust representations that can be used for a variety of downstream applications. In extensive benchmarking across a range of key fetal ultrasound applications, including classification, gestational age estimation, congenital heart defect (CHD) detection, and fetal structure segmentation, FetalCLIP outperformed all baselines while demonstrating remarkable generalizability and strong performance even with limited labeled data. The FetalCLIP model is publicly available at https://github.com/biomedia-mbzuai/fetalclip to support the broader scientific community.

BACKGROUND: Axillary lymph node metastasis (ALNM) is a critical prognostic factor in breast cancer. While sentinel lymph node biopsy remains the gold standard, conventional imaging relies on operator expertise with variable diagnostic accuracy. This meta-analysis evaluates the diagnostic performance of deep learning (DL) and hand-crafted radiomics (HCR) models using MRI and ultrasound (US) for ALNM prediction. METHODS: Literature search was conducted across four databases up to November 2024. Studies assessing DL or HCR models for ALNM prediction using MRI or US, with histopathological confirmation as the reference standard, were included. Diagnostic accuracy metrics were pooled using bivariate random-effects meta-analysis. Heterogeneity was assessed using Higgins I², and subgroup analyses explored its potential sources. RESULTS: Across 41 included studies, pooled sensitivity was 0.79 (95% CI: 0.74-0.84) and specificity 0.78 (95% CI: 0.75-0.81) for internal validation, with AUC 0.84. External validation demonstrated sensitivity of 0.78 and specificity of 0.74, with an AUC of 0.82. Likelihood ratio analysis (LR+ 3.0, LR- 0.33) indicated limited standalone clinical utility. Ensemble approaches combining DL and HCR showed higher diagnostic performance (AUC = 0.88 in MRI and AUC = 0.92 in US) compared to individual methods. Models incorporating both intratumoral and peritumoral regions yielded higher AUCs (0.81 vs 0.75) than intratumoral alone. CONCLUSION: AI models demonstrate moderate diagnostic accuracy but limited standalone clinical utility. These tools may serve adjunctive roles in risk stratification and treatment planning. Ensemble approaches combining DL and HCR achieve superior performance. Methodological standardization and validation across diverse populations are essential before clinical implementation. KEY POINTS: Question Can artificial intelligence models applied to US and MRI accurately predict axillary lymph-node metastasis in breast cancer? Findings Across 41 studies, AI models demonstrated good diagnostic performance for predicting nodal metastasis. DL, HCR, and combined approaches each achieved clinically meaningful accuracy, with integrated DL + HCR models showing the highest pooled diagnostic effect and lowest heterogeneity. Clinical relevance AI-enhanced imaging can assist in non-invasively stratifying nodal status and may reduce reliance on invasive procedures such as sentinel lymph-node biopsy. Consistent imaging protocols, standardized model development, and external validation are needed before clinical adoption.

Heart failure (HF) affects 11.8% of adults aged 65 and older, reducing quality of life and longevity. Preventing HF can reduce morbidity and mortality. We hypothesized that artificial intelligence (AI) applied to 24-hour single-lead electrocardiogram (ECG) data could predict the risk of HF within five years. To research this, the Technion-Leumit Holter ECG (TLHE) dataset, including 69,663 recordings from 47,729 patients, collected over 20 years, was used. Our deep learning model, DeepHHF, trained on 24-hour ECG recordings, achieved an area under the receiver operating characteristic curve of 0.80 that outperformed a model using 30-second segments and a clinical score. High-risk individuals identified by DeepHHF had a two-fold chance of hospitalization or death incidents. Explainability analysis showed DeepHHF focused on arrhythmias and heart abnormalities. This study highlights the feasibility of deep learning to model 24-hour continuous ECG data, capturing paroxysmal events essential for reliable risk prediction. Artificial intelligence applied to single-lead Holter ECG is non-invasive, inexpensive, and widely accessible, making it a promising tool for HF risk prediction.

Artificial intelligence (AI) has shown promise in dermatology, offering accurate and non-invasive diagnosis of skin cancer. While extensive research has addressed skin-tone bias, gender bias in dermatologic AI remains underexplored, potentially perpetuating diagnostic disparities. In this study, we developed LesionAttn, an algorithm designed to mitigate gender bias by directing model attention toward lesions, thereby mirroring clinicians' diagnostic focus. Combined with Pareto Frontier optimization for dual-objective model selection, LesionAttn balances gender fairness and diagnostic performance. Validated on two large-scale dermatologic datasets for binary malignancy classification, LesionAttn significantly mitigated gender bias while maintaining high diagnostic performance, outperforming existing bias-mitigation algorithms. Our study demonstrates that explicitly guiding model attention to medically essential features provides a practical approach to advance both performance and fairness in dermatologic AI. By leveraging clinical priors to bridge the gap between human expertise and algorithmic optimization, this study demonstrates a feasible pathway for developing equitable and reliable diagnostic tools.