The Solid Harmonic Wavelet Bispectrum in 2D provides a multi-scale, rotation- and translation-covariant representation that preserves relative phase and captures higher-order interactions between wavelet responses. This representation encodes rich structural information in a data-efficient and interpretable form. Applications across texture classification, medical imaging, galaxy merger regression, and image reconstruction demonstrate that phase-sensitive, cross-scale interactions enhance discriminative power, model complex dependencies, and retain sufficient information for accurate reconstructions. By embedding roto-translation invariance and preserving relative phase, the operator captures structural features often lost in conventional scattering methods, enabling robust performance in low-data regimes. Cross-scale and higher-order interactions further enrich the representation, allowing nonlinear dependencies between features to be encoded without learning. Results show competitive or superior performance compared to deep learning models in tasks where symmetries and structural cues dominate, highlighting the potential of phase-sensitive, symmetry-aware wavelet representations as a versatile tool for signal and image analysis.

The concomitant development and evolution of lung computed tomography (CT) and artificial intelligence (AI) have made non-invasive lung imaging a key component of clinical care of patients. However, the scarcity of labeled CT data and the limited generative capacity of existing models have constrained their clinical utility. Here, we present LCTfound, a large-scale vision foundation model designed to overcome these limitations. Trained on a multi-center dataset comprising 105,184 CT scans, LCTfound leverages diffusion-based pretraining and joint encoding of imaging and clinical information to support 8 tasks, including CT enhancement, virtual computed tomography angiography (CTA), sparse-view reconstruction, lesion segmentation, diagnosis, prognosis, cancer pathological response prediction, and three-dimensional surgical navigation. In comprehensive multicenter evaluations, LCTfound consistently outperforms leading baseline models, delivering a unified, broadly deployable solution that both augments clinical decision-making and elevates CT image quality across diverse practice settings. LCTfound establishes a scalable foundation for next-generation clinical imaging intelligence, uniting large AI model with precision healthcare.

Heart failure with preserved ejection fraction (HFpEF) is a complex, age-related cardiovascular disease with limited treatment options, partly due to a poor understanding of underlying mechanisms, lack of robust preclinical models and diagnostic tools with limited specificity. Traditional cardiac magnetic resonance imaging (MRI) protocols and analysis in preclinical research are time-consuming, and manual analysis methods are prone to high inter-observer variability (correlation coefficient of 0.79 for left ventricular (LV) ejection fraction between observers). To accelerate and standardize phenotyping, we optimized a comprehensive non-contrast cardiac MRI protocol for high throughput, enabling acquisition of a stack of 12 short-axis slices in approximately seven minutes. This time-efficiency allowed us to add additional sequences, including cine-Arterial Spin Labeling (ASL) for myocardial perfusion mapping and dobutamine stress testing, allowing for a comprehensive cardiac exam. We developed a deep learning approach utilizing 3D Convolutional Neural Networks (CNNs) for fully automated segmentation and quantification of cardiac function. We validated this comprehensive pipeline in two multifactorial mouse models of HFpEF, combining diet-induced obesity (DIO) or high fat diet (HFD) and the hypertensive agent, deoxycorticosterone pivalate (DOCP). Our approach illustrated high technical sensitivity by detecting significant myocardial perfusion reduction in both the DIO (p = 0.02) and DIO + DOCP (p = 0.03) groups compared to control, along with subtle diastolic abnormalities, even in the absence of overt changes in ejection fraction. The CNN demonstrated high accuracy and reproducibility, achieving a mean Dice similarity coefficient greater than 0.9 for segmentation and Intraclass Correlation Coefficients (ICC) exceeding 0.95 for key left ventricular functional parameters (volumes and mass) compared to expert consensus reads. This optimized protocol and automated analysis pipeline provides a valuable tool for preclinical cardiovascular research, enabling efficient and reliable assessment of cardiac remodeling and contributing to a deeper understanding of HFpEF progression.

Medical image segmentation plays an important role in various clinical applications, but existing deep learning models face trade-offs between efficiency and accuracy. Convolutional Neural Networks (CNNs) capture local details well but miss the global context, whereas transformers handle the global context but at a high computational cost. Recently, State Space Sequence Models (SSMs) have shown potential for capturing long-range dependencies with linear complexity, but their direct use in medical image segmentation remains limited due to incompatibility with image structures and autoregressive assumptions. To overcome these challenges, we propose SAMA-UNet, a novel U-shaped architecture that introduces two key innovations. First, the Self-Adaptive Mamba-like Aggregated Attention (SAMA) block adaptively integrates local and global features through dynamic attention weighting, enabling an efficient representation of complex anatomical patterns. Second, the causal resonance multi-scale module (CR-MSM) improves encoder-decoder interactions by adjusting feature resolution and causal dependencies across scales, enhancing the semantic alignment between low- and high-level features. Extensive experiments on MRI, CT, and endoscopy datasets demonstrate that SAMA-UNet consistently outperforms CNN, Transformer, and Mamba-based methods. It achieves 85.38% DSC and 87.82% NSD on BTCV, 92.16% and 96.54% on ACDC, 67.14% and 68.70% on EndoVis17, and 84.06% and 88.47% on ATLAS23, establishing new benchmarks across modalities. These results confirm the effectiveness of SAMA-UNet in combining efficiency with accuracy, making it a promising solution for real-world clinical segmentation tasks. The source code is available on https://github.com/sqbqamar/SAMA-UNet.

Accurate preoperative prognosis prediction is crucial for gastric cancer (GC) treatment planning, yet existing models overlook body composition integration. This study demonstrates the potential of integrating multimodal data, including skeletal muscle (SM), adipose tissue (AT), and primary tumor computed tomography images, to improve prognostic stratification in GC patients using an entire cohort of 1862 patients. By leveraging a Vision Transformer-based deep learning approach, we developed and validated a SM-AT-Tumor-Clinical (SMAT-TC) integrated score to predict recurrence-free survival (RFS) in GC patients. The SMAT-TC score achieved a C-index of 0.966 (95% CI: 0.937-0.990), 0.890 (95% CI: 0.866-0.915), and 0.855 (95% CI: 0.829-0.881) in the training, internal validation, and external validation cohorts, respectively, outperforming the Clinical, SM, AT, Tumor, Tumor-Clinical (TC), and SM-Tumor-Clinical (SM-TC) models. The net reclassification improvement and integrated discrimination improvement confirmed the incremental value of body composition. The SMAT-TC score was an independent risk factor for recurrence. The SMAT-TC model could stratify patients into high-, medium-, and low-risk groups with distinct 3- (99.6% vs. 67.0% vs. 10.9%) and 5-year RFS rates (98.8% vs. 61.7% vs. 2.4%). Collectively, the SMAT-TC score may serve as a novel imaging biomarker for GC patients, enhancing risk stratification and guiding individualized treatment strategies.

Intracranial meningiomas(IM) are often associated with peritumoral brain edema(PTBE), visible as hyperintensities on T2/FLAIR MRI. Postoperative persisting PTBE-like changes likely represent gliosis that, in turn, contributes to surgical morbidity. Since the human eye is unable to distinguish between PTBE and gliosis on MR images, we hypothesized that radiomic analysis of preoperative peritumoral T2/FLAIR hyperintensities could distinguish preoperatively established gliosis from reversible edema. MRI of patients with gross totally resected IM were retrospectively analyzed. Preoperative and 1-year postoperative PTBE were segmented on MRI. One-year MRI were classified into two categories based on whether PTBE resolution exceeded 80% of the initial volume. RF were extracted from meningioma and PTBE regions on T1-contrast-enhanced, T2, and FLAIR MRI sequences. The dataset was split into training, validation, and test cohorts(70-10-20%). Feature reduction used correlation-based exclusion and recursive feature elimination with cross-validation. Nine ML algorithms were trained and evaluated, and best model's interpretability assessed using Shapley Additive Explanations(SHAP). 644 RF were extracted per individual from the pre and postoperative MRI of 123 operated patients. The Random Forest model utilizing 10 RF achieved the best performance (accuracy:0.91;precision:0.92;F1-score:0.92;ROC-AUC:0.94), demonstrating radiomics' utility in predicting PTBE resolution at 1-year post-surgery. SHAP analysis provided interpretability, highlighting key RF, differences between patient groups, and potential sources of algorithmic error. These results underscore the potential of radiomics and ML to accurately predict postoperative PTBE resolution, differentiating transient PTBE from persistent PTBE-like changes (gliosis). This study provides initial insights into the potential of advanced imaging and computational techniques for non-invasive preoperative assessment, which may contribute to more personalized surgical strategies.

Lung cancer remains a significant cause of mortality, with non-small cell lung cancer (NSCLC) representing most cases. Currently, clinical data based models fall short in predicting survival while more advanced deep learning based image models require vast amounts of data and are often limited to predictions based on single time points. This study uses dual time point CT scans and features derived from a foundation model to predict survival. A dataset containing 102 NSCLC patients treated with radiation therapy was used, with each patient having both pre-treatment and post-treatment CT scans. A foundation model applied to the scans generated high-dimensional feature vectors and these vectors were then further summarized. Statistical analyses, including random forest and gradient boosted survival models, were then used to predict survival. The results demonstrated that temporal changes in feature vectors, specifically the Euclidean distance and element-wise subtracted feature vectors, can offer improved prediction of survival over single-time point features and clinical data.

Complete revascularisation has been associated with improved short-term outcomes in patients with coronary artery disease, but whether these benefits persist long-term and can be defined non-invasively remains uncertain. We investigated the long-term prognostic impact of complete versus incomplete revascularisation determined by coronary CT angiography-derived fractional flow reserve (FFR_CT). In this prospective multicentre study, 900 patients with new-onset stable angina and at least one coronary stenosis of 30% or greater on coronary CT angiography were followed for a median of 7 years. FFR_CT values were obtained for each vessel, and patients were categorised as completely revascularised (all vessels with FFR_CT ≤0.80 revascularised), incompletely revascularised (one or more vessels with FFR_CT ≤0.80 not revascularised), or with normal physiology (all vessels with FFR_CT >0.80). Early revascularisation was defined as treatment within 90 days of the index scan. Quantitative coronary plaque burden was assessed using artificial intelligence-enabled plaque analysis. The primary endpoint was a composite of cardiovascular death or spontaneous myocardial infarction. Of 900 patients, 210 (23%) were classified as incompletely revascularised, 167 (19%) as completely revascularised and 523 (58%) as having normal physiology. The primary endpoint occurred in 34 of 210 (16.2%) incompletely revascularised patients, 13 of 167 (7.8%) completely revascularised patients and 30 of 523 (5.7%) with normal physiology. Incomplete revascularisation was associated with higher risk compared with complete revascularisation (HR 2.33, 95% CI 1.23 to 4.42; p=0.01) and normal physiology (HR 3.54, 95% CI 2.16 to 5.81; p<0.001). These associations remained significant after adjustment for total plaque burden, and the risk difference persisted beyond 3 years of follow-up. Complete revascularisation defined by CT-derived fractional flow reserve was associated with a sustained reduction in cardiovascular death and spontaneous myocardial infarction over 7 years, supporting its potential role as a non-invasive tool to guide revascularisation strategies in stable coronary artery disease.

Accurate applicator reconstruction is a critical step in 3D image-guided brachytherapy (3D-IGBT) for cervical cancer, directly influencing tumor control and organ-at-risk sparing. This systematic review evaluates the accuracy, efficiency, and clinical impact of applicator reconstruction methods, focusing on AI's potential to overcome existing limitations. Following PRISMA guidelines, 23 studies from MEDLINE, PubMed, Scopus, Embase, Lilacs and Web of Science (up to May 2025) were analyzed. Evaluation metrics included geometric accuracy (tip error, Hausdorff distance), reconstruction time, and dosimetric parameters (D90 HR-CTV, D2cc OARs). Methods assessed spanned manual (e.g., MPR, scout-based), semi-automatic (library method, clustering algorithms), and AI-driven approaches (e.g., U-Net, Dilated-Supervised Deep U-Net, Attention-Gated networks). Special focus was placed on deep learning (DL) architectures and their ability to overcome metallic artifacts, partial-volume effects, and inter-operator variability. Manual methods exhibited significant limitations, with tip errors reaching 4.1 mm. Semi-automated approaches reduced variability (library-based methods: <0.5 mm mean deviation) but remained constrained by predefined applicator models. AI-driven workflows demonstrated superior precision, achieving submillimeter accuracy (median tip error: 0.64 mm; Dice Similarity Coefficient (DSC) > 0.89) and dosimetric consistency (D2cc deviations <3%). Notably, DL models like DSD-UNet and Attention-Gated U-Net reduced reconstruction time to <30 s per case while maintaining robustness against CT artifacts. However, challenges persist, including limited clinical validation (60% of studies used phantoms), data heterogeneity (slice thickness: 0.6-5 mm), and generalizability to novel applicator designs. AI-driven reconstruction reduces human-dependent errors and enhances efficiency, but clinical validation remains a priority. Reducing CT slice thickness (≤1.5 mm) and combining scout images to mitigate metal artifacts are recommended. Future research should focus on generalizable AI models for nonlibrary applicators and large-scale clinical validation.

Thyroid nodules are common incidental findings but only a small proportion of cases are malignant (4-6.5%) or symptomatic. Numerous follow-up examinations and invasive diagnostic procedures, such as fine-needle aspiration, fine-needle biopsies, and thyroidectomies, are performed, leading to potentially costly and time-consuming diagnostic procedures and overtreatment. Most experts and scientific societies (EFSUMB, WFUMB...) encourage the use of multiparametric ultrasound evaluation to improve the thyroid nodule characterization thanks also to the continuous technological developments with different ultrasound software (Microvascular flow imaging, Elastosonography...), contrast media (CEUS) and artificial intelligence (AI). Therefore, the recognition and proper use of new multiparametric ultrasound features of thyroid nodules are essential to minimize unnecessary interventions and guide appropriate treatments, also stimulating their use in routine clinical practice, which is the aim of our guideline. In addition, we analyze the use of MPUS in some emblematic thyroid scenarios such as cytologically indeterminate nodule, multinodular goiter and extrathyroidal extension of the malignant nodule and the usefulness of MPUS as a guide to thyroid biopsy/aspiration and in the staging of cervical lymph nodes. For each question, recommendations based on the level of evidence of the published literature and on the EFSUMB expert group's consensus are given.