The development of foundation models for brain MRI depends critically on the scale, diversity, and consistency of available data, yet systematic assessments of these factors remain scarce. In this study, we analyze 54 publicly accessible brain MRI datasets encompassing over 538,031 to provide a structured, multi-level overview tailored to foundation model development. At the dataset level, we characterize modality composition, disease coverage, and dataset scale, revealing strong imbalances between large healthy cohorts and smaller clinical populations. At the image level, we quantify voxel spacing, orientation, and intensity distributions across 15 representative datasets, demonstrating substantial heterogeneity that can influence representation learning. We then perform a quantitative evaluation of preprocessing variability, examining how intensity normalization, bias field correction, skull stripping, spatial registration, and interpolation alter voxel statistics and geometry. While these steps improve within-dataset consistency, residual differences persist between datasets. Finally, feature-space case study using a 3D DenseNet121 shows measurable residual covariate shift after standardized preprocessing, confirming that harmonization alone cannot eliminate inter-dataset bias. Together, these analyses provide a unified characterization of variability in public brain MRI resources and emphasize the need for preprocessing-aware and domain-adaptive strategies in the design of generalizable brain MRI foundation models.

This paper presents the FuzzyDistillViT-MobileNet model, a novel approach for lung cancer (LC) classification, leveraging dynamic fuzzy logic-driven knowledge distillation (KD) to address uncertainty and complexity in disease diagnosis. Unlike traditional models that rely on static KD with fixed weights, our method dynamically adjusts the distillation weight using fuzzy logic, enabling the student model to focus on high-confidence regions while reducing attention to ambiguous areas. This dynamic adjustment improves the model ability to handle varying uncertainty levels across different regions of LC images. We employ the Vision Transformer (ViT-B32) as the instructor model, which effectively transfers knowledge to the student model, MobileNet, enhancing the student generalization capabilities. The training process is further optimized using a dynamic wait adjustment mechanism that adapts the training procedure for improved convergence and performance. To enhance image quality, we introduce pixel-level image fusion improvement techniques such as Gamma correction and Histogram Equalization. The processed images (Pix1 and Pix2) are fused using a wavelet-based fusion method to improve image resolution and feature preservation. This fusion method uses the wavedec2 function to standardize images to a 224x224 resolution, decompose them into multi-scale frequency components, and recursively average coefficients at each level for better feature representation. To address computational efficiency, Genetic Algorithm (GA) is used to select the most suitable pre-trained student model from a pool of 12 candidates, balancing model performance with computational cost. The model is evaluated on two datasets, including LC25000 histopathological images (99.16% accuracy) and IQOTH/NCCD CT-scan images (99.54% accuracy), demonstrating robustness across different imaging domains.

Osteochondrosis encompasses a heterogenous group of pathologies affecting endochondral ossification in the growing child and adolescent. The cause of each disease is multifactorial, though many are often related to overuse injury, and may be epiphyseal, physeal, or apophyseal. Identification and treatment of this group of disorders is complex, thus this review aims to briefly describe common pathologies, their management, and highlight novel developments within the field. Machine learning as well as advanced diagnostic tools for more precise evaluation and prognostication of osteochondroses have been studied including perfusion MRI in Legg-Calvé-Perthes disease. Novel treatments include leukocyte-rich platelet-rich plasma (LR-PRP), which offer promising improvements in pain and function in Osgood-Schlatter disease. Surgical technique studies have begun to examine optimal operative management of Freiberg's disease. The osteochondroses are an often-self-limiting spectrum of pathologies affecting the physis in children and adolescents that may be managed with conservative treatment, though some require surgical intervention. Advances in imaging, prognostication tools, and treatment modalities support earlier and accurate diagnoses, as well as better informed treatment decisions.

To propose a multi-parametric ultrasound imaging-based deep learning method for accurately classifying metastatic and non-metastatic axillary lymph nodes in breast cancer patients. The proposed method integrates the conventional ultrasound B-mode imaging with shear wave elastography and color Doppler images of 174 patients to train a transfer learning-based network comprising pretrained MobileNetv2 with a custom shallow head consisting of a convolutional neural network with mixed pooling, weighted sum mixed pooling and squeeze-and-excite attention mechanisms for the first time in the context of ALN classification. The proposed method was evaluated using five-fold cross-validation, achieving a mean classification accuracy of 0.91, specificity of 0.91, sensitivity of 0.93, F1 score of 0.93, area under the precision-recall curve of 0.94, and a cross-validated AUC (cvAUC) of 0.92. A network ablation study confirmed the robustness of the model, with relatively narrow 95% confidence intervals (CIs) for cvAUC. Comparative analysis showed that the proposed network (Acc: 0.91) outperformed state-of-the-art deep learning models (Acc: 0.67-0.88) for ALN classification and exhibited narrower CIs, highlighting its relative stability. Additionally, results demonstrated that multi-parametric imaging significantly enhanced classification performance, reducing the 95% CI width by nearly half compared to uni-parametric data, further supporting the method's robustness and reliability. The integration of multi-parametric ultrasound imaging with deep learning network can remarkably improve the classification of metastatic and non-metastatic ALNs in breast cancer patients.

This study developed a multiomics model combining radiomics, pathomics, and temporal imaging to predict major pathological response in patients with locally advanced non-small cell lung cancer (NSCLC) undergoing neoadjuvant immunochemotherapy. A retrospective, multicenter study was conducted, enrolling 271 patients with stage IB-III NSCLC who received neoadjuvant immunochemotherapy. High-resolution CT images were enhanced using a generative adversarial network-based super-resolution technique. Radiomics features were extracted from multi-sequence CT scans at multiple time points, while pathomics features were derived from whole-slide imaging of surgical specimens. A transformer-based attention mechanism was used to integrate radiomics, pathomics, and temporal imaging data. The model was trained and validated on data from one center and tested on external cohorts. Performance was evaluated using area under the curve (AUC), net reclassification improvement, integrated discrimination improvement, and decision curve analysis. The Trans-Model demonstrated superior predictive performance, achieving an AUC of 0.858 (95% CI 0.783 to 0.933) in the external test cohort. It outperformed Rad-Model (AUC: 0.839) and Patho-Model (AUC: 0.753). The Trans-Model effectively stratified patients by survival outcomes, with major pathological response (MPR)-positive patients exhibiting significantly improved 3-year overall survival (87.3% vs 76.1%, p=0.034) and 5-year progression-free survival (45.8% vs 34.7%, p=0.033) compared with MPR-negative patients. Decision curve analysis confirmed the model's clinical utility across a wide range of threshold probabilities. The multiomics model, integrating multi-temporal, multi-sequence data with attention-based feature fusion, improves MPR prediction in patients with NSCLC receiving neoadjuvant immunochemotherapy, enabling personalized treatment by identifying responders and optimizing outcomes.

Thrombosis, the formation of blood clots within blood vessels, poses serious health risks including pulmonary embolism and post-thrombotic syndrome. Ultrasound (US) imaging is a widely used, non-invasive diagnostic tool owing to its real-time capability and safety profile; however, its effectiveness is often limited by operator dependency and variability in interpretation. This scoping review investigates how deep learning (DL) techniques have been applied to enhance thrombosis detection and risk assessment using US imaging across venous, arterial, and cardiac contexts. A comprehensive literature search was conducted in PubMed and Scopus following PRISMA-ScR methodology, targeting studies that used DL models for thrombus detection, classification, segmentation, or risk prediction in conjunction with vascular US modalities such as B-mode, Doppler, intravascular ultrasound (IVUS), and transesophageal echocardiography (TEE). Out of 233 records initially identified, 22 studies met the eligibility criteria. The most frequently used models included convolutional neural networks (CNNs), U-Net, Residual Neural Networks (ResNet), and Artificial Neural Networks (ANNs). DL models mainly aided deep vein thrombosis (DVT) diagnosis by evaluating vein compressibility and supporting point-of-care ultrasound (POCUS) imaging. Arterial thrombosis applications focused on plaque segmentation and vessel reconstruction using IVUS, while cardiac studies employed TEE to differentiate thrombi from tumours. Studies often reported high sensitivity, specificity, accuracy, and area under the curve (AUC), frequently outperforming traditional rule-based or manual interpretation methods, although considerable variability in datasets and validation approaches was observed. Overall, DL-enhanced US imaging shows great promise for improving diagnostic precision and clinical decision-making in thrombosis care. Future research should prioritize model interpretability, real-world integration, and the development of standardized, publicly accessible datasets.

Lung ultrasound (LUS) is an essential tool for diagnosing lung diseases. However, its effectiveness is often limited by its reproducibility, making interpretation challenging for clinicians. LUS diagnosis typically relies on subjective assessments of pleural line and vertical artifacts. To address this limitation, we introduce a novel quantitative approach aimed at reducing the need to rely on human operators (HOs) for LUS data assessment (i.e., improving the reproducibility). In the first phase of our study, we propose a hybrid method that integrates motion estimation and K-means clustering for automated segmentation of LUS images. The technique utilizes K-means clustering to identify pleural line based on intensity variations, while motion estimation detects vertical artifacts by analyzing motion vectors between consecutive frames. Rather than employing a conventional learning-based classification model, we develop an interpretable scoring framework that assigns scores to individual video frames according to standard scoring criteria. A threshold-based approach is then applied to aggregate frame-level scores, determining the final score for each video. We evaluated our method on a clinical dataset comprising 420 neonatal LUS videos from 70 patients, with annotations provided by three HOs. When using the majority vote among HOs as the reference standard, our method achieved a video-level accuracy of 0.72. For cases with full agreement among HOs, accuracy improved to 0.77. These results demonstrate that our approach offers comparable or superior performance to state-of-the-art deep learning (DL)-based methods in terms of scoring consistency, while reducing the need for a huge training dataset.

Peritoneal metastases (PM) significantly impact treatment options and prognosis of patients with cancer. Early detection and accurate evaluation are essential for guiding clinical decisions. This systematic review and meta-analysis aimed to provide a comprehensive overview and evaluate the performance of Artificial Intelligence (AI) and radiomics models for diagnosis and prognosis of PM on imaging. A systematic search of PubMed, Embase, and the Cochrane Library was conducted for studies published up to July 2024 that evaluated AI or radiomics models analyzing imaging data for diagnosing or predicting prognosis in PM. Data were extracted, and if more than 3 studies evaluated the same endpoint and reported true/false positive and negative values, a meta-analysis was conducted to obtain pooled area under the curve (AUC), sensitivity, and specificity. Bias was assessed using the PROBAST + AI tool. This review included 24 studies, of which 18 evaluated PM presence, 2 assessed PM severity (low versus high Peritoneal Cancer Index (PCI)), and 4 focused on prognosis or treatment efficacy. Meta-analysis of 13 studies evaluating PM presence revealed a pooled AUC of 0.84, sensitivity of 0.75, and specificity of 0.80. Subgroup analysis indicated comparable performance for 2D and 3D imaging data, and lower performance for models detecting occult PM compared to all PM presentations. Incorporating clinical factors into AI and radiomics models improved performance. AI and radiomics models demonstrated promising performance outcomes for PM evaluation on imaging, showing potential to aid in diagnosis and prognosis prediction. However, large validation studies are needed to evaluate their effects in clinical practice.

To develop and validate DeepMocor, a deep learning-based method for motion-compensated four-dimensional magnetic resonance fingerprinting (4D-MRF) reconstruction to accelerate conventional 4D-MRF reconstruction, enabling more efficient clinical treatment planning. This prospective study enrolled 19 hepatocellular carcinoma patients (mean age, 62 years; 14 males) between June 2021 and October 2024. Abdominal free-breathing raw k-space data were acquired using a 3T MRI scanner. DeepMocor involves motion field initialization, motion field refinement, and final 4D-MRF reconstruction. A three-fold cross-validation strategy was employed for training and testing. Performance was evaluated against two alternatives (Stage-I&III-only; Stage-III-only) in terms of image quality, tissue property accuracy, tumor-to-tissue contrast, and tumor motion measurement. Image quality was assessed by Peak Signal-to-Noise Ratio (PSNR) and Structural Similarity Index (SSIM). Tissue property accuracy was evaluated by Mean Absolute Percentage Error (MAPE). Tumor-to-tissue contrast was quantified by Contrast-to-Noise Ratio (CNR) of the tumor region and the surrounding area. Tumor motion tracking was assessed by Average Motion Discrepancy (AMD) and Pearson Correlation Coefficients (PCC) in the superior-inferior (SI) and anterior-posterior (AP) directions. The Wilcoxon signed rank test was used for comparison with P < 0.05. For T1 maps, DeepMocor demonstrates PSNR of 25.49 ± 1.30, SSIM of 0.84 ± 0.03, MAPE of 3.5%-5.9%, and CNR of 6.14 ± 3.54. For T2 maps, DeepMocor achieves PSNR of 25.57 ± 1.24, SSIM of 0.88 ± 0.02, MAPE of 3.1%-15.8%, and CNR of 8.42 ± 13.72. DeepMocor achieves AMD of 0.62 ± 0.86 mm with PCC of 0.96 ± 0.07 in the SI direction and AMD of 0.32 ± 0.37 mm with PCC of 0.94 ± 0.06 in the AP direction. DeepMocor shows superior performance across most metrics compared to Stage-III-only and a subset of metrics compared to Stage-I&III-only significantly. The proposed DeepMocor method enables a 24-fold acceleration compared to the conventional reference method, highlighting its potential for liver radiotherapy planning.

To evaluate whether quantitative diffusion-weighted magnetic resonance imaging (DW-MRI) derived apparent diffusion coefficient (ADC) parametric values of nodal disease can identify nodal recurrence (NR) risk and distinguish human papillomavirus-associated oropharyngeal cancer (HPV-OPC) patients suitable for radiation de-escalation following primary tumor resection. This secondary analysis of a prospective hypoxia-guided radiation de-escalation trial included 94 of 158 HPV-OPC patients who underwent serial MRI scans and pre- and intra-treatment ¹⁸F-FMISO PET hypoxia imaging. Patients with persistent hypoxia (pre- and intra-treatment hypoxia) received standard 70Gy chemoradiation, while those with baseline normoxia or resolved hypoxia received de-escalated 30Gy treatment. All NR occurred in the 30Gy arm. Nodal volume, mean ADC and distribution parameters were quantified and correlated with hypoxia status and clinical outcomes. Random forest modeling assessed multiparametric MRI features for predicting treatment assignment. Nodal recurrences exhibited larger intra-treatment volume (weeks 2-4, p<0.05) and mean ADC (weeks 3-4, p<0.05) compared to non-recurrent nodes. Baseline hypoxic tumors were significantly larger than normoxic tumors (p<0.001). Analysis of an expanded panel of quantitative MRI (qMRI) features for prediction of hypoxia-based treatment assignment identified ADC skewness as significantly different between 30Gy and 70Gy arms (p<0.05 pre-treatment and week 1). Machine learning models incorporating multiple qMRI features achieved moderate predictive performance, with week 1 qMRI features performing best (AUC=0.67). Intra-treatment MRI features correlated with NR in de-escalated HPV-OPC patients, while ADC skewness corresponded with hypoxia-based treatment assignment. These widely available and contrast-free imaging biomarkers warrant further exploration for guiding safe treatment adaptation in precision radiotherapy.