Background High-resolution magnetic resonance imaging (HR-MRI) provides a non-invasive, radiation-free approach for evaluating stenosis caused by carotid atherosclerosis. However, manual recognition is time-consuming and inter-observer variability. We propose a novel architecture for automated segmentation and stenosis evaluation of extracranial carotid arteries by HR-MRI in comparison with digital subtraction angiography (DSA). Methods The 641 stenotic arteries from 422 patients retrospectively collected from three tertiary hospitals were divided into a training-validation set (372 patients, 545 lesions) and an independent test set (50 patients, 96 lesions). An external validation set (89 patients, 168 lesions) was collected from the fourth tertiary hospital. Results The architecture demonstrated high consistency with manual segmentation and DSA diagnostic criteria, with mean Dice similarity coefficients of 0.97 ± 0.01, 0.96 ± 0.01, and stenosis evaluation accuracies of 0.88, 0.86 on the independent test and external validation set, respectively. Conclusion Thus, the proposed architecture achieved accuracy comparable to manual segmentation by physicians and demonstrated high consistency with DSA diagnostic criteria. By shortening diagnostic time and minimizing inter-observer variability, the proposed architecture is promising to offer a reliable, efficient, and intelligent tool for diagnosing head and neck atherosclerotic disease and assessing stroke risk.

Automatic segmentation and volume analysis of the hippocampus are crucial for the diagnosis of Alzheimer’s disease (AD). However, existing segmentation models face significant challenges due to the hippocampus’s irregular shape, blurred boundaries, and substantial inter-individual variability. This study aims to develop a robust segmentation method and improve the accuracy and reliability of hippocampal analysis. We propose a novel hybrid segmentation network (NU-Mamba) to segment hippocampus in the coronal T1 MRI. It features a dual-layer nested U-shape structure that integrates Mamba and CNN blocks, capturing both intra-stage global contextual information and inter-stage multi-level information. A Mamba block based on the predefined adjustment strategy is proposed to enhance the capability for long-range spatial modeling. Furthermore, we propose an Uncertainty Attention Module (UAM) that captures low-confidence regions within the saliency map and guides the model to focus on semantically ambiguous boundaries. Extensive experiments conducted on multi-center clinical and public datasets demonstrate the superior performance of NU-Mamba, achieving DSC: 0.900, IoU: 0.813, HD95: 1.234 mm, and ASD: 0.243 mm. Additionally, hippocampal volume analysis and Bland–Altman analysis demonstrated good agreement between NU-Mamba and ground-truth volumes. NU-Mamba offers accurate and robust hippocampal segmentation in coronal T1 MRI. These results highlight NU-Mamba as a valuable tool for aiding AD imaging analysis and diagnosis.

Background:Mitochondrial encephalomyopathy with lactic acidosis and stroke-like episodes (MELAS) is a rare maternally inherited disease. Cognitive impairment is one of the main clinical manifestations in MELAS patients, however, the underlying brain network mechanism of cognitive impairment is not entirely clear. The "triple network model" provides a common framework for understanding cognitive impairment in core neurocognitive networks, yet little is known about the dynamic functional connectivity (dFC) of MELAS patients in the triple network. Therefore, this study aimed to investigate the characteristics of dFC within the triple network in MELAS patients to better understand the neural network mechanisms underlying their cognitive impairment. Methods:This cross-sectional study analyzed data from an ongoing prospective cohort study of genetically confirmed MELAS patients. Thirty patients at the acute stage (MELAS-acute group), 30 patients at the chronic stage (MELAS-chronic group), and 30 healthy control volunteers (HC group) were included in this study. The triple network was confirmed using a group spatial independent component analysis (ICA), and dFC was analyzed using a sliding window approach (SWA) and k-means clustering algorithm. In addition, we explored the correlations between temporal properties of dFC states and volumes of stroke-like lesions (SLLs). Results:The intrinsic brain functional connectivity (FC) within the triple network was clustered into four states. The results revealed distinct FC states, characterized by varying patterns of inter-network coupling. State 4, characterized by the weakest FC across all networks, was the most prevalent state in all participants. State 2 exhibited the strongest positive default mode network (DMN)-central executive network (CEN) coupling but negative salience network (SN)-DMN/CEN integration. State 3 was characterized by weaker positive DMN-left CEN (lCEN) coupling and weaker negative SN-DMN/CEN integration than state 2. State 1 demonstrated stronger positive DMN-CEN coupling and stronger positive SN-DMN/CEN integration than state 3. We found that MELAS patients spent more time in states with weaker FC. Specifically, the MELAS-acute group had a lower recurrence fraction (RF) in state 1 (P=0.0229) and shorter mean dwell time (MDT) (P=0.0414) but higher RF (P=0.008) and longer MDT (P=0.0162) in state 3 compared with MELAS-chronic group. And that MELAS-chronic group had lower RF (P=0.0141) and shorter MDT (P=0.0137) in state 3 but higher RF (P=0.0499) in state 4 compared with HC group. And MELAS-chronic group switched less frequently across states compared with HC group (P=0.0347, Dunn's correction). Conclusions:This study revealed abnormal temporal properties of dFC states within the triple network in MELAS patients, providing novel insights for understanding neural network mechanisms of their cognitive impairment.

Gliomas and brain metastases (BMs) on MRI pose significant diagnostic challenges for radiologists. This study aims to develop a multi-task model and a computer-aided diagnosis (CAD) system for the detection and diagnosis of gliomas and BMs. This study enrolled 3909 participants from seven centers, and developed a brain tumor segmentation and classification network (BTSC-Net) and BTSC-CAD with visualization of tumor masks. For detection, BTSC-Net achieved a Dice coefficient of 0.888 and 0.872 on the internal and external test sets, respectively. For diagnosis, BTSC-Net achieved AUCs of 0.941 and 0.933 on the internal and external test sets, respectively. With BTSC-CAD assistance, junior radiologists achieved mean AUC improvements of 4.8% (P < 0.05) for detection and 17.3% (P < 0.001) for diagnosis, along with an average reduction of 64.75 s in reading time. BTSC-CAD significantly improved radiologists' diagnostic accuracy and efficiency.

To address the clinical challenge of preoperative microvascular invasion (MVI) grading in hepatocellular carcinoma (HCC), this paper proposes a cross-channel attention fine-grained Network (CRAF-Net) using dynamic contrast-enhanced magnetic resonance imaging (DCE-MRI). Current methods often suffer from fragmented representations and suboptimal channel weighting, obscuring key features and reducing accuracy. CRAF-Net addresses these challenges through three innovations: (1) a Dense-NeXt encoder that enhances 3D hierarchical feature reuse; (2) a cross-channel attention (CCA) Module that adaptively weights multi-phase MRI sequences to emphasize tumor microenvironment (TME) biomarkers; and (3) a dual-branch disentangled network with a multi-classifier fusion Module that separates morphological MVI risk assessment from severity grading, alleviating inter-class ambiguities. Evaluated on a primary cohort of 472 HCC cases, CRAF-Net achieves an 84.8

BACKGROUND AND PURPOSE: Hierarchy is a fundamental principle of network organization in the human brain. Functional gradient introduces a new perspective in identifying hierarchy alterations by capturing major axes of functional connectivity in low-dimensional space. However, whether this gradient structure is disrupted in patients with mitochondrial encephalomyopathy, lactic acidosis, and stroke-like episodes (MELAS) and how this disruption is modulated by the gene-expression profiles remain unknown. MATERIALS AND METHODS: Thirty-one patients with MELAS at the acute stage (MELAS-acute) and 31 healthy controls underwent resting-state functional MR imaging. On the basis of whole-brain voxelwise functional connectivity patterns, functional gradient values were generated and group-averaged gradient values were further extracted and compared from the global-to-voxel level. In combination with the Allen Human Brain Atlas, we then assessed the spatial correlations between MELAS-related gradient alterations and gene-expression profiles. RESULTS: Relative to the healthy controls, patients with MELAS-acute exhibited global alterations in the principal gradients, including reduced gradient range and gradient variation. In addition, patients showed lower gradient values in the default mode network but higher values in the ventral attention network and the sensorimotor network at the network and voxel levels. Furthermore, we established a link between MELAS-acute-related principal gradient and gene-expression profiles, with 2 gene sets mainly enriched in mitochondrion, neuron, glutamatergic synapse, and ATPase activity. CONCLUSIONS: These results highlight the connectome gradient alterations in patients with MELAS at the acute stage and their linkage with gene-expression profiles, providing insight into the neurobiological basis of functional alterations during the acute stroke-like episodes stage in MELAS.

The Carotid Plaque Reporting and Data System (Plaque-RADS) is a new quantitative grading system for plaque stability. Herein, high-resolution magnetic resonance imaging (HRMRI)-based Plaque-RADS(HRMRI-Plaque-RADS) classification was performed to characterize the different classifications in patients with cerebral infarction. HRMRI images from patients with acute cerebral infarction (ACI) of the anterior cerebral circulation were collected retrospectively, and carotid plaques were categorized into the Plaque-RADS 1–4 subtypes. Plaque-RADS 1 and 2 were considered low-risk, and Plaque-RADS 3 and 4 as high-risk. After contrast agent injection, high-risk plaques were further classified as enhanced or non-enhanced. A mixed-effects logistic regression model was applied to identify relationships between plaque types and infarction. This study included 98 patients (age 65.6 ± 9.6 years; 88.8

Owing to a large amount of multi-modal data in modern medical systems, such as medical images and reports, Medical Vision-Language Pre-training (Med-VLP) has demonstrated incredible achievements in coarse-grained downstream tasks (i.e., medical classification, retrieval, and visual question answering). However, the problem of transferring knowledge learned from Med-VLP to fine-grained multi-organ segmentation tasks has barely been investigated. Multi-organ segmentation is challenging mainly due to the lack of large-scale fully annotated datasets and the wide variation in the shape and size of the same organ between individuals with different diseases. In this paper, we propose a novel pre-training fine-tuning framework for Multi-Organ Segmentation by harnessing Medical repOrt Supervision (MOSMOS). Specifically, we first introduce global contrastive learning to maximally align the medical image-report pairs in the pre-training stage. To remedy the granularity discrepancy, we further leverage multi-label recognition to implicitly learn the semantic correspondence between image pixels and organ tags. More importantly, our pre-trained models can be transferred to any segmentation model by introducing the pixel-tag attention maps. Different network settings, i.e., 2D U-Net and 3D UNETR, are utilized to validate the generalization. We have extensively evaluated our approach using different diseases and modalities on BTCV, AMOS, MMWHS, and BRATS datasets. Experimental results in various settings demonstrate the effectiveness of our framework. This framework can serve as the foundation to facilitate future research on automatic annotation tasks under the supervision of medical reports.

Rationale and Objectives: Diagnosis of carotid plaques from head and neck CT angiography (CTA) scans is typically time-consuming and labor-intensive, leading to limited studies and unpleasant results in this area. The objective of this study is to develop a deep-learning-based model for detection and segmentation of carotid plaques using CTA images. Materials and Methods: CTA images from 1061 patients (765 male; 296 female) with 4048 carotid plaques were included and split into a 75% training-validation set and a 25% independent test set. We built a workflow involving three modified deep learning networks: a plain U-Net for coarse artery segmentation, an Attention U-Net for fine artery segmentation, a dual-channel-input ConvNeXt-based U-Net architecture for plaque segmentation, and post-processing to refine predictions and eliminate false positives. The models were trained on the training-validation set using five-fold cross-validation and further evaluated on the independent test set using comprehensive metrics for segmentation and plaque detection. Results: The proposed workflow was evaluated in the independent test set (261 patients with 902 carotid plaques) and achieved a mean dice similarity coefficient (DSC) of 0.91 +/- 0.04 in artery segmentation, and 0.75 +/- 0.14/0.67 +/- 0.15 in plaque segmentation per artery/patient. The model detected 95.5% (861/902) plaques, including 96.6% (423/438), 95.3% (307/322), and 92.3% (131/142) of calcified, mixed, and soft plaques, with less than one (0.63 +/- 0.93) false positive plaque per patient on average. Conclusion: This study developed an automatic detection and segmentation deep learning-based CAP-Net for carotid plaques using CTA, which yielded promising results in identifying and delineating plaques.

PURPOSE:Accurate preoperative grading of gliomas is critical for therapeutic planning and prognostic evaluation. We developed a noninvasive machine learning model leveraging whole-brain resting-state functional magnetic resonance imaging (rs-fMRI) biomarkers to discriminate high-grade (HGGs) and low-grade gliomas (LGGs) in the frontal lobe. METHODS:This retrospective study included 138 patients (78 LGGs, 60 HGGs) with left frontal gliomas. A total of 7134 features were extracted from the mean amplitude of low-frequency fluctuation (mALFF), mean fractional ALFF, mean percentage amplitude of fluctuation (mPerAF), mean regional homogeneity (mReHo) maps and resting-state functional connectivity (RSFC) matrix. Twelve predictive features were selected through Mann-Whitney U test, correlation analysis and least absolute shrinkage and selection operator method. The patients were stratified and randomized into the training and testing datasets with a 7:3 ratio. The logical regression, random forest, support vector machine (SVM) and adaptive boosting algorithms were used to establish models. The model performance was evaluated using area under the receiver operating characteristic curve, accuracy, sensitivity, and specificity. RESULTS:The selected 12 features included 7 RSFC features, 4 mPerAF features, and 1 mReHo feature. Based on these features, the model was established using the SVM had an optimal performance. The accuracy in the training and testing datasets was 0.957 and 0.727, respectively. The area under the receiver operating characteristic curves was 0.972 and 0.799, respectively. CONCLUSIONS:Our whole-brain rs-fMRI radiomics approach provides an objective tool for preoperative glioma stratification. The biological interpretability of selected features reflects distinct neuroplasticity patterns between LGGs and HGGs, advancing understanding of glioma-network interactions.

Background:Early infarct growth rate (EIGR) and both arterial- and tissue-level collaterals (TLCs) are strongly associated with stroke prognosis, yet the complex interplay among multi-level collateral status, EIGR, and stroke outcomes remains incompletely understood. This study aimed to comprehensively characterize collaterals at the arterial, tissue, and venous outflow (VO) levels, and to delineate the distribution of EIGR among acute stroke patients with varying clinical outcomes. Methods:Patients with acute large vessel occlusion were retrospectively recruited. Pial arterial collaterals (PACs), TLCs and VO were measured by the modified Tan (mTan) scale, hypoperfusion intensity ratio (HIR), and the cortical vein opacification score (COVES), respectively. EIGR was subsequently calculated. The imaging and clinical outcomes were measured by final infarct volume (FIV) and modified Rankin Scales (mRS) at 90 days, respectively. Mediation analysis was performed to quantify the effect of collaterals on outcomes as explained by EIGR. Results:A total of 166 patients were included, with 61 exhibiting good clinical outcomes and 105 poor outcomes. The median EIGR and FIV were 2.0 mL/h and 14.4 mL in the good clinical outcome group, and 11.3 mL/h and 126.5 mL in the poor outcome group. Patients with unfavorable collaterals developed significantly higher EIGR. The mTan score (r=-0.456, P<0.001), HIR (r=0.314, P<0.001) and COVES score (r=-0.300, P<0.001) demonstrated significant correlations with EIGR. Presentation National Institutes of Health Stroke Scale (NIHSS) score and EIGR were identified to be co-determinants of FIV and mRS. EIGR mediated 28.2%, 27.9% and 32.2% of the effects of PACs, TLCs, VO on FIV and 33.7%, 25.9% and 35.9% on mRS, respectively. Conclusions:Collaterals and EIGR were integral determinants of stroke prognosis, with EIGR functioning as a key mediator. These findings offer a more nuanced understanding of how different levels of collateral flow influence infarct evolution and stroke outcomes in acute ischemic stroke (AIS).

ABSTRACT Background Alzheimer's disease (AD) and white‐matter structural connectivity have been linked in some observational studies, although it is unknown if this is a causal relationship. The purpose of this study was to examine the impact of various white‐matter structural connectivity on AD via a two‐sample multivariate Mendelian randomization (MR) approach. Methods The genome‐wide association study (GWAS) of Wainberg et al. provided the summary data on white‐matter structural connectivity, and Bellenguez et al.’s study provided the GWAS aggregated data for AD. MR methods included inverse variance weighted, Mendelian randomization Egger, simple mode, weighted median, and weighted mode. Heterogeneity, horizontal pleiotropy, and “leave‐one‐out” analysis guaranteed the robustness of causation. Finally, reverse MR analysis was conducted on the white‐matter structural connectivity that showed positive results in the forward MR analysis. Results Among 206 white‐matter structural connections, we identified 10 connections were strongly correlated with genetic susceptibility to AD. Right‐hemisphere limbic network to thalamus white‐matter structural connectivity and Right‐hemisphere salience_ventral attention network to accumbens white‐matter structural connectivity were positively correlated with the likelihood of AD, while the remaining 8 white‐matter structural connections were negatively related with AD. None of the above 10 white‐matter structural connections have a reverse causal relationship with AD. Conclusion Our MR study reveals a certain degree of association between white‐matter structural connectivity and AD, which may provide support for future diagnosis and treatment of AD.

3D blood vessel segmentation remains a critical yet challenging task in medical image analysis. The heterogeneity of clinical imaging protocols introduces substantial domain gaps, limiting the generalizability of supervised learning methods that rely on manually annotated pixel-level labels for individual datasets. Furthermore, the large labeled volumetric datasets are difficult to collect because of data privacy issues. While diffusion models offer potential solutions by generating shareable synthetic data, existing approaches often exhibit poor alignment between synthesized volumes and their corresponding vascular structure input. To address these limitations, we propose Controllable Adversarial Diffusion Model (AVDM), which integrates adversarial supervision into the diffusion training framework. Unlike conventional methods that generate imperceptible perturbations, AVDM synthesizes adversarial instances emphasizing structural variations critical for volume synthesis. Specifically, we design a segmentation-guided discriminator that enforces both the photorealism of generated volumes and pixellevel consistency with original vessel annotations. This supervision mechanism enables high-resolution synthesis of anatomically plausible vascular structures. Experiments demonstrate that AVDM surpasses state-of-the-art methods in generative fidelity and enhances performance on downstream tasks. Our code is available at https://github.com/jdai22/AVDM.

OBJECTIVE: To characterize preoperative alterations and postoperative reorganization of gray matter volume (GMV)-based individual-level morphological networks in unilateral frontal lobe low-grade gliomas (FLGGs) and evaluate clinical relevance. MATERIALS AND METHODS: T1-weighted structural MRI (sMRI) data were acquired from 90 left FLGGs (LFLGGs) patients, 45 right FLGGs (RFLGGs) patients, and 30 healthy controls (HCs). Individual morphological networks were constructed by quantifying interregional GMV distribution similarity using Jensen-Shannon divergence (JSD). Graph-theoretical metrics and interregional connectivity were compared between patients and HCs and longitudinally assessed at 1 and 3 months postoperatively. Correlations between network features, Mini-Mental State Examination (MMSE), and MD Anderson Symptom Inventory for Brain Tumor (MDASI-BT) scores were evaluated. RESULTS: Preoperatively, unilateral FLGGs exhibited decreased network segregation and enhanced integration relative to HCs. Nodal metrics decreased in peritumoral and contralesional frontal regions but increased in temporal areas. Network-Based Statistics (NBS) revealed weakened inter-frontal connectivity and strengthened frontal-to-extra-frontal connectivity. Nodal metrics of peritumoral anterior cingulate gyrus (ACG) and contralesional superior frontal gyrus (ORBsup), along with their interconnectivity, correlated positively with MMSE and negatively with MDASI-BT scores. Postoperatively, global metrics declined transiently at 1 month but recovered by 3 months. Nodal metrics decreased acutely in periresectional regions at 1 month but showed compensatory increases in contralesional areas by 3 months. Connectivity weakened near the surgical site at 1 month but strengthened in ipsilateral distal and contralesional regions by 3 months. CONCLUSION: Unilateral FLGGs disrupted morphological networks topology, correlating with cognitive deficits and symptom severity. Postoperative reorganization demonstrated acute surgical effects followed by compensatory neural adaptation. Individual-level morphological networks provided a clinically feasible tool for monitoring network dynamics and guiding personalized interventions.

Pre-trained on large-scale datasets has profoundly promoted the development of deep learning models in medical image analysis. For medical image segmentation, collecting a large number of labeled volumetric medical images from multiple institutions is an enormous challenge due to privacy concerns. Self-supervised learning with mask image modeling (MIM) can learn general representation without annotations. Integrating MIM into FL enables collaborative learning of an efficient pre-trained model from unlabeled data, followed by fine-tuning with limited annotations. However, setting pixels as reconstruction targets in traditional MIM fails to facilitate robust representation learning due to the medical image's complexity and distinct characteristics. On the other hand, the generalization of the aggregated model in FL is also impaired under the heterogeneous data distributions among institutions. To address these issues, we proposed a novel self-supervised federated learning, which combines masked self-distillation with adaptive attention federated learning. Such incorporation enjoys two vital benefits. First, masked self-distillation sets high-quality latent representations of masked tokens as the target, improving the descriptive capability of the learned presentation rather than reconstructing low-level pixels. Second, adaptive attention aggregation with Personalized federate learning effectively captures specific-related representation from the aggregated model, thus facilitating local fine-tuning performance for target tasks. We conducted comprehensive experiments on two medical segmentation tasks using a large-scale dataset consisting of volumetric medical images from multiple institutions, demonstrating superior performance compared to existing federated self-supervised learning approaches.

To address SPECT’s radioactivity, complexity, and costliness in measuring renal function, this study employs artificial intelligence (AI) with non-contrast CT to estimate single-kidney glomerular filtration rate (GFR) and split renal function (SRF). 245 patients with atrophic kidney or hydronephrosis were included from two centers (Training set: 128 patients from Center I; Test set: 117 patients from Center II). The renal parenchyma and hydronephrosis regions in non-contrast CT were automatically segmented by deep learning. Radiomic features were extracted and combined with clinical characteristics using multivariable linear regression (MLR) to obtain a radiomics-clinical-estimated GFR (rcGFR). The relative contribution of single-kidney rcGFR to overall rcGFR, the percent renal parenchymal volume, and the percent renal hydronephrosis volume were combined by MLR to generate the estimation of SRF (rcphSRF). The Pearson correlation coefficient (r), mean absolute error (MAE), and Lin’s concordance coefficient (CCC) were calculated to evaluate the correlations, differences, and agreements between estimations and SPECT-based measurements, respectively. Compared to manual segmentation, deep learning-based automatic segmentation could reduce the average segmentation time by 434.6 times to 3.4 s. Compared to single-kidney GFR measured by SPECT, the rcGFR had a significant correlation of r = 0.75 (p < 0.001), MAE of 10.66 mL/min/1.73 m2, and CCC of 0.70. Compared to SRF measured by SPECT, the rcphSRF had a significant correlation of r = 0.92 (p < 0.001), MAE of 7.87

Prostate cancer (PCa) is one of the most common malignancies in men, and accurate assessment of tumor aggressiveness is crucial for treatment planning. The Gleason score (GS) remains the gold standard for risk stratification, yet it relies on invasive biopsy, which has inherent risks and sampling errors. The aim of this study was to detect PCa and non-invasively predict the GS for the early detection and stratification of clinically significant cases. We used single-modality T2-weighted imaging (T2WI) with an automatic machine-learning (ML) approach, MLJAR. The internal dataset comprised PCa patients who underwent magnetic resonance imaging (MRI) examinations at our hospital from September 2015 to June 2022 prior to prostate biopsy, surgery, radiotherapy, and endocrine therapy and whose examinations resulted in pathological findings. An external dataset from another medical center and a public challenge dataset were used for external validation. The Kolmogorov–Smirnov curve was used to evaluate the risk-differentiation ability of the PCa detection model. The area under the receiver operating characteristic curve (AUC) was calculated with confidence intervals to compare the model performance. The internal MRI dataset included 198 non-PCa and 291 PCa patients with histopathological results obtained through biopsy or surgery. External and public challenge datasets included 45 and 68 PCa patients, respectively. AUC for PCa detection in the internal-testing cohort (n = 147, PCa = 78) was 0.99. For GS prediction, AUCs were GS = 3 + 3 (0.97), GS = 3 + 4 (0.97), GS = 3 + 5 (1.0), GS = 4 + 3 (0.87), GS = 4 + 4 (0.91), GS = 4 + 5 (0.95), GS = 5 + 4 (1.0), and GS = 5 + 5 (0.99) in the internal-testing cohort (PCa = 88); GS = 3 + 3 (0.95), GS = 3 + 4 (0.76); GS = 3 + 5 (0.77), GS = 4 + 3 (0.88), GS = 4 + 4 (0.82), GS = 4 + 5 (0.87), GS = 5 + 4 (0.95), and GS = 5 + 5 (0.85) in the external-testing cohort (PCa = 45); and GS = 3 + 4 (0.89), GS = 4 + 3 (0.75), GS = 4 + 4 (0.65), and GS = 4 + 5 (0.91) in the public challenge cohort (PCa = 68). This multi-center study shows that an auto-ML model using only T2WI can accurately detect PCa and predict Gleason scores non-invasively, offering potential to reduce biopsy reliance and improve early risk stratification. These results warrant further validation and exploration for integration into clinical workflows.

RATIONALE AND OBJECTIVES:Diagnosis of carotid plaques from head and neck CT angiography (CTA) scans is typically time-consuming and labor-intensive, leading to limited studies and unpleasant results in this area. The objective of this study is to develop a deep-learning-based model for detection and segmentation of carotid plaques using CTA images. MATERIALS AND METHODS:CTA images from 1061 patients (765 male; 296 female) with 4048 carotid plaques were included and split into a 75% training-validation set and a 25% independent test set. We built a workflow involving three modified deep learning networks: a plain U-Net for coarse artery segmentation, an Attention U-Net for fine artery segmentation, a dual-channel-input ConvNeXt-based U-Net architecture for plaque segmentation, and post-processing to refine predictions and eliminate false positives. The models were trained on the training-validation set using five-fold cross-validation and further evaluated on the independent test set using comprehensive metrics for segmentation and plaque detection. RESULTS:The proposed workflow was evaluated in the independent test set (261 patients with 902 carotid plaques) and achieved a mean dice similarity coefficient (DSC) of 0.91±0.04 in artery segmentation, and 0.75±0.14/0.67±0.15 in plaque segmentation per artery/patient. The model detected 95.5% (861/902) plaques, including 96.6% (423/438), 95.3% (307/322), and 92.3% (131/142) of calcified, mixed, and soft plaques, with less than one (0.63±0.93) false positive plaque per patient on average. CONCLUSION:This study developed an automatic detection and segmentation deep learning-based CAP-Net for carotid plaques using CTA, which yielded promising results in identifying and delineating plaques.

BACKGROUND:This study developed a deep learning model for segmenting and classifying the amygdala-hippocampus in Alzheimer's disease (AD), using a large-scale neuroimaging dataset to improve early AD detection and intervention. METHODS:We collected 1000 healthy controls (HC) and 1000 AD patients as internal training data from 15 Chinese medical centers. The independent external validation dataset was sourced from another three centers. All subjects underwent neuroimaging and neuropsychological assessments. A semi-automated annotation pipeline was used: the amygdala-hippocampus of 200 cases in each group were manually annotated to train the U²-Net segmentation model, followed by model annotation of 800 cases with iterative refinement. The DenseNet-121 architecture was built for automated classification. The robustness of the model was evaluated using an external validation set. RESULTS:All 18 medical centers were distributed across diverse geographical regions in China. AD patients had lower MMSE/MoCA scores. Amygdala and hippocampal volumes were smaller in AD. Semi-automated annotation improved segmentation with DSC all exceeding 0.88 (P＜0.001). The final DSC of the 2000-case cohort was 0.914 in the training set and 0.896 in the testing set. The classification model achieved an AUC of 0.905. The external validation set comprised 100 cases in each group, and it can achieve an AUC of 0.835. CONCLUSION:The amygdala-hippocampus recognition precision may be improved by the deep learning-based semi-automated approach and classification model, which will help with AD evaluation, diagnosis, and clinical AI application.

To explore the alterations of gray matter volume (GMV) and structural covariant network (SCN) in unilateral frontal lobe low-grade gliomas (FLGGs). The three dimensional (3D) T1 structural images of 117 patients with unilateral FLGGs and 68 age- and sex-matched healthy controls (HCs) were enrolled. The voxel-based morphometry (VBM) analysis and graph theoretical analysis of SCN were conducted to investigate the impact of unilateral FLGGs on the brain structure. This represents the first structural MRI study integrating both voxel-level morphometric changes and network-level reorganization patterns in unilateral FLGGs. Through VBM analysis, we found that unilateral FLGGs can cause increased GMV in contralesional amygdala, calcarine, and angular gyrus, ipsilesional amygdala as well as vermis_6. The SCN of contralesional cerebrum, ipsilesional unaffected regions and cerebellum in both patients and HCs have typical small-world properties (Sigma > 1, Lambda ≈ 1 and Gamma > 1). Compared to HCs, global and nodal network metrics changed significantly in patients. The combination of VBM and SCN analysis revealed both focal GMV enlargement and topological alterations in patients with unilateral FLGGs, and provide a novel perspective of cross regional morphological collaborative changes for understanding the glioma-related neuroadaptation. These findings may suggest potential neuroimaging correlates of adaptive changes, which could inform future investigations into personalized treatment approaches. Not applicable.