PurposeTo evaluate convolutional neural network (CNN) model training strategies that optimize the performance of calcaneus fracture detection on radiographs at different image resolutions. Materials and MethodsThis retrospective study included foot radiographs from a single hospital between 2015 and 2022 for a total of 1,775 x-ray series (551 fractures; 1,224 without) and was split into training (70%), validation (15%), and testing (15%). ImageNet pre-trained ResNet models were fine-tuned on the dataset. Three training strategies were evaluated: 1) single size: trained exclusively on 128x128, 256x256, 512x512, 640x640, or 900x900 radiographs (5 model sets); 2) curriculum learning: trained exclusively on 128x128 radiographs then exclusively on 256x256, then 512x512, then 640x640, and finally on 900x900 (5 model sets); and 3) multi-scale augmentation: trained on x-ray images resized along continuous dimensions between 128x128 to 900x900 (1 model set). Inference time and training time were compared. ResultsMulti-scale augmentation trained models achieved the highest average area under the Receiver Operating Characteristic curve of 0.938 [95% CI: 0.936 - 0.939] for a single model across image resolutions compared to the other strategies without prolonging training or inference time. Using the optimal model sets, curriculum learning had the highest sensitivity on in-distribution low-resolution images (85.4% to 90.1%) and on out-of-distribution high-resolution images (78.2% to 89.2%). However, curriculum learning models took significantly longer to train (11.8 [IQR: 11.1-16.4] hours; P<.001). ConclusioWhile 512x512 images worked well for fracture identification, curriculum learning and multi-scale augmentation training strategies algorithmically improved model robustness towards different image resolutions without requiring additional annotated data. Summary statementDifferent deep learning training strategies affect performance in detecting calcaneus fractures on radiographs across in- and out-of-distribution image resolutions, with a multi-scale augmentation strategy conferring the greatest overall performance improvement in a single model. Key pointsO_LITraining strategies addressing differences in radiograph image resolution (or pixel dimensions) could improve deep learning performance. C_LIO_LIThe highest average performance across different image resolutions in a single model was achieved by multi-scale augmentation, where the sampled training dataset is uniformly resized between square resolutions of 128x128 to 900x900. C_LIO_LICompared to model training on a single image resolution, sequentially training on increasingly higher resolution images up to 900x900 (i.e., curriculum learning) resulted in higher fracture detection performance on images resolutions between 128x128 and 2048x2048. C_LI

Carotid artery plaques, a hallmark of atherosclerosis, are key risk indicators for ischemic stroke, a major global health burden with 101 million cases and 6.65 million deaths in 2019. Early ultrasound detection is vital but hindered by manual analysis limitations. Machine learning (ML) offers a promising solution for automated plaque detection, yet its comparative performance is underexplored. This systematic review and meta-analysis evaluates ML models for carotid plaque detection using ultrasound. We searched PubMed, Scopus, Embase, Web of Science, and ProQuest for studies on ML-based carotid plaque detection with ultrasound, following PRISMA guidelines. Eligible studies reported diagnostic metrics and used a reference standard. Data on study characteristics, ML models, and performance were extracted, with risk of bias assessed via PROBAST+AI. Pooled sensitivity, specificity, AUROC were calculated using STATA 18 with MIDAS and METADTA modules. Of ten studies, eight were meta-analyzed (200-19,751 patients) Best models showed a pooled sensitivity 0.94 (95% CI: 0.88-0.97), specificity 0.95 (95% CI: 0.86-0.98), AUROC 0.98 (95% CI: 0.97-0.99), and DOR 302 (95% CI: 54-1684), with high heterogeneity (I² = 90%) and no publication bias. ML models show promise in carotid plaque detection, supporting potential clinical integration for stroke prevention, though high heterogeneity and potential bias highlight the need for standardized validation.

BackgroundCarotid plaque presence is associated with cardiovascular risk, even among asymptomatic individuals. While deep learning has shown promise for carotid plaque phenotyping in patients with advanced atherosclerosis, its application in population-based settings of asymptomatic individuals remains unexplored. MethodsWe developed a YOLOv8-based model for plaque detection using carotid ultrasound images from 19,499 participants of the population-based UK Biobank (UKB) and fine-tuned it for external validation in the BiDirect study (N = 2,105). Cox regression was used to estimate the impact of plaque presence and count on major cardiovascular events. To explore the genetic architecture of carotid atherosclerosis, we conducted a genome-wide association study (GWAS) meta-analysis of the UKB and CHARGE cohorts. Mendelian randomization (MR) assessed the effect of genetic predisposition to vascular risk factors on carotid atherosclerosis. ResultsOur model demonstrated high performance with accuracy, sensitivity, and specificity exceeding 85%, enabling identification of carotid plaques in 45% of the UKB population (aged 47-83 years). In the external BiDirect cohort, a fine-tuned model achieved 86% accuracy, 78% sensitivity, and 90% specificity. Plaque presence and count were associated with risk of major adverse cardiovascular events (MACE) over a follow-up of up to seven years, improving risk reclassification beyond the Pooled Cohort Equations. A GWAS meta-analysis of carotid plaques uncovered two novel genomic loci, with downstream analyses implicating targets of investigational drugs in advanced clinical development. Observational and MR analyses showed associations between smoking, LDL cholesterol, hypertension, and odds of carotid atherosclerosis. ConclusionsOur model offers a scalable solution for early carotid plaque detection, potentially enabling automated screening in asymptomatic individuals and improving plaque phenotyping in population-based cohorts. This approach could advance large-scale atherosclerosis research. O_FIG O_LINKSMALLFIG WIDTH=200 HEIGHT=131 SRC="FIGDIR/small/24315675v2_ufig1.gif" ALT="Figure 1"> View larger version (33K): [email protected]@27a04corg.highwire.dtl.DTLVardef@18cef18org.highwire.dtl.DTLVardef@1a53d8f_HPS_FORMAT_FIGEXP M_FIG O_FLOATNOGRAPHICAL ABSTRACT.C_FLOATNO ASCVD - Atherosclerotic Cardiovascular Disease, CVD - Cardiovascular disease, PCE - Pooled Cohort Equations, TP- true positive, FN - False Negative, FP - False Positive, TN - True Negative, GWAS - Genome-Wide Association Study. C_FIG CLINICAL PERSPECTIVECarotid ultrasound is a well-established method for assessing subclinical atherosclerosis with potential to improve cardiovascular risk assessment in asymptomatic individuals. Deep learning could automate plaque screening and enable processing of large imaging datasets, reducing the need for manual annotation. Integrating such large-scale carotid ultrasound datasets with clinical, genetic, and other relevant data can advance cardiovascular research. Prior studies applying deep learning to carotid ultrasound have focused on technical tasks-plaque classification, segmentation, and characterization-in small sample sizes of patients with advanced atherosclerosis. However, they did not assess the potential of deep learning in detecting plaques in asymptomatic individuals at the population level. We developed an efficient deep learning model for the automated detection and quantification of early carotid plaques in ultrasound imaging, primarily in asymptomatic individuals. The model demonstrated high accuracy and external validity across population-based cohort studies. Predicted plaque prevalence aligned with known cardiovascular risk factors. Importantly, predicted plaque presence and count were associated with future cardiovascular events and improved reclassification of asymptomatic individuals into clinically meaningful risk categories. Integrating our model predictions with genetic data identified two novel loci associated with carotid plaque presence--both previously linked to cardiovascular disease--highlighting the models potential for population-scale atherosclerosis research. Our model provides a scalable solution for automated carotid plaque phenotyping in ultrasound images at the population level. These findings support its use for automated screening in asymptomatic individuals and for streamlining plaque phenotyping in large cohorts, thereby advancing research on subclinical atherosclerosis in the general population.

Colorectal cancer (CRC) ranks as the second deadliest cancer globally, impacting patients' quality of life. Colonoscopy is the primary screening method for detecting adenomas and polyps, crucial for reducing long-term CRC risk, but it misses about 30% of cases. Efforts to improve detection rates include using AI to enhance colonoscopy. This study assesses the effectiveness and accuracy of a real-time AI-assisted polyp detection system during colonoscopy. The study included 390 patients aged 40 to 75 undergoing colonoscopies for either colorectal cancer screening (risk score ≥ 4) or clinical diagnosis. Participants were randomly assigned to an experimental group using software-assisted diagnosis or a control group with physician diagnosis. The software, a medical image processing tool with B/S and MVC architecture, operates on Windows 10 (64-bit) and supports real-time image handling and lesion identification via HDMI, SDI, AV, and DVI outputs from endoscopy devices. Expert evaluations of retrospective video lesions served as the gold standard. Efficacy was assessed by polyp per colonoscopy (PPC), adenoma per colonoscopy (APC), adenoma detection rate (ADR), and polyp detection rate (PDR), while accuracy was measured using sensitivity and specificity against the gold standard. In this multicenter, randomized controlled trial, computer-aided detection (CADe) significantly improved polyp detection rates (PDR), achieving 67.18% in the CADe group versus 56.92% in the control group. The CADe group identified more polyps, especially those 5 mm or smaller (61.03% vs. 56.92%). In addition, the CADe group demonstrated higher specificity (98.44%) and sensitivity (95.19%) in the FAS dataset, and improved sensitivity (95.82% vs. 77.53%) in the PPS dataset, with both groups maintaining 100% specificity. These results suggest that the AI-assisted system enhances PDR accuracy. This real-time computer-aided polyp detection system enhances efficacy by boosting adenoma and polyp detection rates, while also achieving high accuracy with excellent sensitivity and specificity.

Abnormality detection in medical imaging is a critical task requiring both high efficiency and accuracy to support effective diagnosis. While convolutional neural networks (CNNs) and Transformer-based models are widely used, both face intrinsic challenges: CNNs have limited receptive fields, restricting their ability to capture broad contextual information, and Transformers encounter prohibitive computational costs when processing high-resolution medical images. Mamba, a recent innovation in natural language processing, has gained attention for its ability to process long sequences with linear complexity, offering a promising alternative. Building on this foundation, we present SpectMamba, the first Mamba-based architecture designed for medical image detection. A key component of SpectMamba is the Hybrid Spatial-Frequency Attention (HSFA) block, which separately learns high- and low-frequency features. This approach effectively mitigates the loss of high-frequency information caused by frequency bias and correlates frequency-domain features with spatial features, thereby enhancing the model's ability to capture global context. To further improve long-range dependencies, we propose the Visual State-Space Module (VSSM) and introduce a novel Hilbert Curve Scanning technique to strengthen spatial correlations and local dependencies, further optimizing the Mamba framework. Comprehensive experiments show that SpectMamba achieves state-of-the-art performance while being both effective and efficient across various medical image detection tasks.

Microscopic tumor cell infiltration beyond contrast-enhancing regions influences glioblastoma prognosis but remains undetectable using conventional MRI. To develop and evaluate the glioblastoma infiltrating area interactive detection framework (GIAIDF), an interactive deep-learning framework that integrates diffusion tensor imaging (DTI) biomarkers for identifying microscopic infiltration within peritumoral edema. Retrospective. A total of 73 training patients (51.13 ± 13.87 years; 47 M/26F) and 25 internal validation patients (52.82 ± 10.76 years; 14 M/11F) from Center 1; 25 external validation patients (47.29 ± 11.39 years; 16 M/9F) from Center 2; 13 prospective biopsy patients (45.62 ± 9.28 years; 8 M/5F) from Center 1. 3.0 T MRI including three-dimensional contrast-enhanced T1-weighted BRAVO sequence (repetition time = 7.8 ms, echo time = 3.0 ms, inversion time = 450 ms, slice thickness = 1 mm), three-dimensional T2-weighted fluid-attenuated inversion recovery (repetition time = 7000 ms, echo time = 120 ms, inversion time = 2000 ms, slice thickness = 1 mm), and diffusion tensor imaging (repetition time = 8500 ms, echo time = 63 ms, slice thickness = 2 mm). Histopathology of 25 stereotactic biopsy specimens served as the reference standard. Primary metrics included AUC, accuracy, sensitivity, and specificity. GIAIDF heatmaps were co-registered to biopsy trajectories using Ratio-FAcpcic (0.16-0.22) as interactive priors. ROC analysis (DeLong's method) for AUC; recall, precision, and F1 score for prediction validation. GIAIDF demonstrated recall = 0.800 ± 0.060, precision = 0.915 ± 0.057, F1 = 0.852 ± 0.044 in internal validation (n = 25) and recall = 0.778 ± 0.053, precision = 0.890 ± 0.051, F1 = 0.829 ± 0.040 in external validation (n = 25). Among 13 patients undergoing stereotactic biopsy, 25 peri-ED specimens were analyzed: 18 without tumor cell infiltration and seven with infiltration, achieving AUC = 0.929 (95% CI: 0.804-1.000), sensitivity = 0.714, specificity = 0.944, and accuracy = 0.880. Infiltrated sites showed significantly higher risk scores (0.549 ± 0.194 vs. 0.205 ± 0.175 in non-infiltrated sites, p < 0.001). This study has provided a potential tool, GIAIDF, to identify regions of GBM infiltration within areas of peri-ED based on preoperative MR images.

Four-dimensional (4D) ultrasound imaging is widely used in clinics for diagnostics and therapy guidance. Accurate target tracking in 4D ultrasound is crucial for autonomous therapy guidance systems, such as radiotherapy, where precise tumor localization ensures effective treatment. Supervised deep learning approaches rely on reliable ground truth, making accurate labels essential. We investigate the reliability of expert-labeled ground truth data by evaluating intra- and inter-observer variability in landmark labeling for 4D ultrasound imaging in the liver. Eight 4D liver ultrasound sequences were labeled by eight expert observers, each labeling eight landmarks three times. Intra- and inter-observer variability was quantified, and observer survey and motion analysis were conducted to determine factors influencing labeling accuracy, such as ultrasound artifacts and motion amplitude. The mean intra-observer variability ranged from <math xmlns="http://www.w3.org/1998/Math/MathML"><mrow><mn>1.58</mn> <mtext>  </mtext> <mi>mm</mi> <mo>±</mo> <mn>0.90</mn> <mtext>  </mtext> <mi>mm</mi></mrow> </math> to <math xmlns="http://www.w3.org/1998/Math/MathML"><mrow><mn>2.05</mn> <mtext>  </mtext> <mi>mm</mi> <mo>±</mo> <mn>1.22</mn> <mtext>  </mtext> <mi>mm</mi></mrow> </math> depending on the observer. The inter-observer variability for the two observer groups was <math xmlns="http://www.w3.org/1998/Math/MathML"><mrow><mn>2.68</mn> <mtext>  </mtext> <mi>mm</mi> <mo>±</mo> <mn>1.69</mn> <mtext>  </mtext> <mi>mm</mi></mrow> </math> and <math xmlns="http://www.w3.org/1998/Math/MathML"><mrow><mn>3.06</mn> <mtext>  </mtext> <mi>mm</mi> <mo>±</mo> <mn>1.74</mn> <mtext>  </mtext> <mi>mm</mi></mrow> </math> . The observer survey and motion analysis revealed that ultrasound artifacts significantly affected labeling accuracy due to limited landmark visibility, whereas motion amplitude had no measurable effect. Our measured mean landmark motion was <math xmlns="http://www.w3.org/1998/Math/MathML"><mrow><mn>11.56</mn> <mtext>  </mtext> <mi>mm</mi> <mo>±</mo> <mn>5.86</mn> <mtext>  </mtext> <mi>mm</mi></mrow> </math> . We highlight variability in expert-labeled ground truth data for 4D ultrasound imaging and identify ultrasound artifacts as a major source of labeling inaccuracies. These findings underscore the importance of addressing observer variability and artifact-related challenges to improve the reliability of ground truth data for evaluating target tracking algorithms in 4D ultrasound applications.

Medical image retrieval is essential for clinical decision-making and translational research, relying on discriminative visual representations. Yet, current methods remain fragmented, relying on separate architectures and training strategies for 2D, 3D, and video-based medical data. This modality-specific design hampers scalability and inhibits the development of unified representations. To enable unified learning, we curate a large-scale hybrid-modality dataset comprising 867,653 medical imaging samples, including 2D X-rays and ultrasounds, RGB endoscopy videos, and 3D CT scans. Leveraging this dataset, we train M3Ret, a unified visual encoder without any modality-specific customization. It successfully learns transferable representations using both generative (MAE) and contrastive (SimDINO) self-supervised learning (SSL) paradigms. Our approach sets a new state-of-the-art in zero-shot image-to-image retrieval across all individual modalities, surpassing strong baselines such as DINOv3 and the text-supervised BMC-CLIP. More remarkably, strong cross-modal alignment emerges without paired data, and the model generalizes to unseen MRI tasks, despite never observing MRI during pretraining, demonstrating the generalizability of purely visual self-supervision to unseen modalities. Comprehensive analyses further validate the scalability of our framework across model and data sizes. These findings deliver a promising signal to the medical imaging community, positioning M3Ret as a step toward foundation models for visual SSL in multimodal medical image understanding.

Acute aortic dissection (AD) is a life threatening condition that poses considerable challenges for timely diagnosis. Non-contrast computed tomography (CT) is frequently used to diagnose AD in certain clinical settings, but its diagnostic accuracy can vary among radiologists. This study aimed to develop and validate an interpretable YOLOv8 deep learning model based on non-contrast CT to detect AD. This retrospective study included patients from five institutions, divided into training, internal validation, and external validation cohorts. The YOLOv8 deep learning model was trained on annotated non-contrast CT images. Its performance was evaluated using area under the curve (AUC), sensitivity, specificity, and inference time compared with findings from vascular interventional radiologists, general radiologists, and radiology residents. In addition, gradient weighted class activation mapping (Grad-CAM) saliency map analysis was performed. A total of 1 138 CT scans were assessed (569 with AD, 569 controls). The YOLOv8s model achieved an AUC of 0.964 (95% confidence interval [CI] 0.939 - 0.988) in the internal validation cohort and 0.970 (95% CI 0.946 - 0.990) in the external validation cohort. In the external validation cohort, the performance of the three groups of radiologists in detecting AD was inferior to that of the YOLOv8s model. The model's sensitivity (0.976) was slightly higher than that of vascular interventional specialists (0.965; p = .18), and its specificity (0.935) was superior to that of general radiologists (0.835; p < .001). The model's inference time was 3.47 seconds, statistically significantly shorter than the radiologists' mean interpretation time of 25.32 seconds (p < .001). Grad-CAM analysis confirmed that the model focused on anatomically and clinically relevant regions, supporting its interpretability. The YOLOv8s deep learning model reliably detected AD on non-contrast CT and outperformed radiologists, particularly in time efficiency and diagnostic accuracy. Its implementation could enhance AD screening in specific settings, support clinical decision making, and improve diagnostic quality.

Thyroid nodules are a common clinical concern, with accurate diagnosis being critical for effective treatment and improved patient outcomes. Traditional ultrasound examinations rely heavily on the physician's experience, which can lead to diagnostic variability. The integration of artificial intelligence (AI) into medical imaging offers a promising solution for enhancing diagnostic accuracy and efficiency. This study aimed to evaluate the effectiveness of the You Only Look Once v. 11 (YOLOv11) model in detecting and classifying thyroid nodules through ultrasound images, with the goal of supporting real-time clinical decision-making and improving diagnostic workflows. We used the YOLOv11 model to analyze a dataset of 1,503 thyroid ultrasound images, divided into training (1,203 images), validation (150 images), and test (150 images) sets, comprising 742 benign and 778 malignant nodules. Advanced data augmentation and transfer learning techniques were applied to optimize model performance. Comparative analysis was conducted with other YOLO variants (YOLOv3 to YOLOv10) and residual network 50 (ResNet50) to assess their diagnostic capabilities. The YOLOv11 model exhibited superior performance in thyroid nodule detection as compared to other YOLO variants (from YOLOv3 to YOLOv10) and ResNet50. At an intersection over union (IoU) of 0.5, YOLOv11 achieved a precision (P) of 0.841 and recall (R) of 0.823, outperforming ResNet50's P of 0.8333 and R of 0.8025. Among the YOLO variants, YOLOv11 consistently achieved the highest P and R values. For benign nodules, YOLOv11 obtained a P of 0.835 and R of 0.833, while for malignant nodules, it reached a P of 0.846 and a R of 0.813. Within the YOLOv11 model itself, performance varied across different IoU thresholds (0.25, 0.5, 0.7, and 0.9). Lower IoU thresholds generally resulted in better performance metrics, with P and R values decreasing as the IoU threshold increased. YOLOv11 proved to be a powerful tool for thyroid nodule detection and malignancy classification, offering high P and real-time performance. These attributes are vital for dynamic ultrasound examinations and enhancing diagnostic efficiency. Future research will focus on expanding datasets and validating the model's clinical utility in real-time settings.