Traditional Image Quality Assessment (IQA) has primarily aimed to quantify perceptual quality in terms of technical degradations such as noise, blur, or compression artifacts. However, in image rendering, the key factor influencing perceived quality is not the presence of degradations but the manner in which color processing algorithms are applied, as they directly shape the overall aesthetic appearance of the image. To date, the quantitative evaluation of how rendering methods affect image quality has been insufficiently addressed. In this work, we introduce Image Rendering Quality Assessment (IRQA) as a new problem setting within IQA and present REPID, a benchmark designed for its study. REPID contains 30,000 edited images and preference annotations collected from 13,648 voters, resulting in an over 2.5 million unique votes. Based on REPID, we investigate content-dependent render preferences and the influence of rendering parameters, and further explore applications such as aesthetic preference prediction (including personalization), render ranking, and benchmarking of aesthetic evaluation methods. We further perform an extensive experimental comparison of traditional IQA metrics, handcrafted features, deep learning approaches, and foundation-model embeddings. On the REPID benchmark, IRQA-specific models achieve up to 40% better precision than conventional distortion-oriented IQA methods.
Accurate and rapid identification of quarantine-significant tephritids is critical to global agricultural biosecurity, but the application of deep learning is limited by the lack of large public image datasets. We present Tephritid26, a multi-angle image dataset of 26 tephritid species to address this gap. The dataset includes 38,081 images from 1,473 specimens across seven genera and two subfamilies, assembled through a global collaborative effort to source these regulated species. Specimens were mounted using a novel protocol combining varied thoracic attachment points and pin angles, and a rotational imaging setup then systematically captured each specimen from multiple perspectives to mimic real inspection conditions. The dataset is formatted for machine learning workflows. To demonstrate its utility, we trained deep learning models for species identification. ResNet-50, ConvNeXt-B, Vit-Small and Swin-Tiny all attained high species-level accuracy (Macro-Averaged F1-score > 96.75). Gradient-weighted Class Activation Mapping confirmed that the models focused on taxonomically informative morphological regions. This dataset serves as a benchmark for developing automated identification tools in phytosanitary applications.
The Kotrupi landslide area has remained active since the 1970s, with a major landslide occurring on 13th August 2017. Since then, the site has experienced repeated reactivations. This study integrates UAV mapping; satellite image analysis; field investigations; and numerical simulation, to evaluate the landslide reactivation and slope stability. TanDEM-X (10 m) and UAV derived DEMs (Digital Elevation Model) were used for establish the pre and post event boundary conditions for stability assessment. Seven representative profiles were selected to characterize the deformation regime and analyze the reactivation potential zones. The Factor of Safety (FoS) was estimated using the Limit Equilibrium Method (LEM) and maximum displacement values inferred using a Finite Element Model (FEM). The results indicate that the right flank exhibits the lowest FoS values, ranging between 0.35 and 0.5 making it highly susceptible to reactivation. In contrast, the left and central portions are comparable stable, as these portions have relatively higher FoS. However, all seven profiles have FoS values lower than 1, indicating overall slope instability. Satellite image analysis further conformed the progressive reactivation if the right flank, whereas the left flank remained comparatively stable over time. Extensive field surveys were conducted to collect geological, geotechnical, and hydrological information. The dataset consists of rock orientation, fault mapping, joint planes, tension crack development, rock type, and hydrological data. Temporal satellite image analysis confirmed continued enlargement of the affected zone and identified significant reactivation events during the monsoon periods of 2021 and 2022. The findings reveal that the Kotrupi landslide is progressively expanding, particularly toward the right flank, with widening observed in the crown area. The reactivation and expansion is primarily controlled by unfavorable rock orientation, presence of thrust (Main Boundary Thrust), tectonic activity; development of extensive joint planes, and tension cracks, all of which reduce the strength of the rock mass and soil during prolonged rainfall. The integrated methodology used in this study provides valuable insight into landslide reactivation mechanisms and helps identify areas susceptible to future slope failure. These findings can support hazard mitigation and risk reduction strategies for local communities and government agencies.
Mucosal thickness (MT) is a key factor influencing peri-implant soft-tissue response and outcomes in dental implantology. This ex vivo study evaluated the agreement of peri-implant MT measurements obtained using ultrasound (US) standardized with a custom probe holder, compared with transgingival probing (TP) and cone-beam computed tomography (CBCT). Porcine hemimandibles (n = 18) underwent guided implant placement. MT was measured at five standardized points using four approaches: (1) US with expert annotation, (2) US with artificial intelligence (AI)-based image segmentation, (3) CBCT, and (4) TP. US images (17 MHz) were independently annotated by two trained specialists; a deep-learning-based method was used to derive automated MT measurements. Method differences were analyzed using a linear mixed-effects model; agreement was assessed using intraclass correlation coefficients (ICCs) and Bland-Altman analysis. The overall method effect was not significant (p = 0.105). Pairwise comparisons showed no significant difference between expert-annotated US and TP (p = 0.328), whereas CBCT yielded higher MT values than TP (p = 0.035). Agreement was moderate for expert-annotated US versus TP (ICC = 0.58; 95% confidence interval (CI): 0.42-0.70) and for expert-annotated US versus AI-segmented US (ICC = 0.67; 95% CI: 0.53-0.77), but poor for expert-annotated US versus CBCT (ICC = 0.14; 95% CI: -0.05-0.33). Bland-Altman analysis showed mean differences (95% limits of agreement) of 0.08 mm (- 0.96 mm to + 1.13 mm) for expert-annotated US - TP, - 0.01 mm (- 0.68 mm to + 0.67 mm) for expert-annotated US-AI-segmented US, and - 0.30 mm (- 2.29 mm to + 1.68 mm) for expert-annotated US-CBCT. Under controlled ex vivo conditions, expert-annotated US standardized with a custom probe holder showed moderate comparative agreement with TP, while AI-segmented measurements showed moderate agreement with expert annotation. CBCT showed limited agreement with US. This integrated approach represents a proof-of-concept requiring further in vivo validation.
Neural architecture search (NAS) can improve medical image segmentation, but its practical use is limited by computational cost and instability of the discovered architectures, particularly when applied to pre-trained models and limited data. We present Shapley-guided pruning as a practical validation-guided extension of retrofit NAS for pre-trained U-Nets. Rather than defining a new NAS family, the method keeps the IAC search space and PC-DARTS-style supernet optimization, while adding iterative pruning driven by Shapley value estimates computed on held-out validation data. By progressively removing low-impact components while preserving learned architecture parameters, the approach reduces search space complexity and improves the reliability of the final discrete architecture. We evaluate the method on four public benchmarks (ACDC, BraTS, KiTS, and AMOS) in a controlled 2D slice-based, [Formula: see text], single-GPU setting. Within this protocol, the proposed approach improves or matches strong baselines in most comparisons, accelerates search by up to four times, and yields more stable operation choices across runs. The findings support Shapley-guided pruning as a practical retrofit-NAS mechanism under resource constraints, without implying direct clinical competitiveness with high-resolution or fully 3D segmentation pipelines.
Magnetic Resonance Imaging (MRI) is a non-invasive imaging method that can give detailed visualization of the cardiac structures and blood flow, which is effective in diagnosis of cardiovascular diseases (CVDs). It has been proposed that the combination of deep learning (DL) with MRI has an improved ability to automatically identify cardiovascular anomalies by identifying intricate patterns in large-scale imaging data. In this study, an Automated Cardiovascular Disease Detection framework (ACVD-RDODL) is proposed, which combines deep learning with the Red Deer Optimiser (RDO). After image enhancement methods like Wiener Filtering (WF) and Dynamic Histogram Equalization (DHE), features are extracted using radiomics. An Attention-Based Convolutional Gated Recurrent Unit (ACGRU) network is considered to ensure proper classification and RDO is used to optimize the hyperparameters and improve the performance of the progress model. As experimental testing of a benchmark cardiac MRI dataset shows, the proposed method is greater to the existing approaches in terms of classification accuracy and computational efficiency.
Level II ultrasound standard section classification is of great importance for accurate prenatal diagnosis. However, sonographers face challenges in distinguishing the lateral ventricle transverse section from the thalamus transverse section. In this study, the standard section images of second-trimester fetal level II prenatal ultrasound examinations from 1261 pregnant women are collected, and a section classification model called Medical Learnable Prompt-Hierarchical Adaptive Feature Block Contrastive Language-Image Pre-training (MedLP-HAFB-CLIP) is proposed. The model uses a meticulously designed prompt learner to generate prompts that are highly compatible with anatomical categories, effectively integrating medical knowledge. By combining fine-grained anatomical descriptions and a customized loss function, the model's discriminative ability for complex image features is significantly enhanced. Additionally, the HAFB module is proposed for multi-scale feature extraction and adaptive fusion, further improving the model's classification performance. Experimental results demonstrate that MedLP-HAFB-CLIP significantly outperforms baseline models. It exhibits excellent performance in distinguishing the lateral ventricle transverse section and the thalamus transverse section. Moreover, in a small exploratory observer study involving 50 selected images and 4 sonographers, the model correctly classified all selected cases and showed shorter reading time than unaided interpretation; however, these preliminary findings should be interpreted with caution and require validation in larger reader studies. The codes are available at https://github.com/Chenan7/MedLP-HAFB-CLIP.git. This study provides a robust and clinically valuable tool for automating the classification of fetal ultrasound standard planes. The proposed method holds significant promise for enhancing the standardization and reliability of prenatal ultrasound screening in clinical practice.
Childbearing entails complex biopsychosocial challenges that have important implications for women's reproductive autonomy. Given that childbirth involves substantial bodily changes, body image concerns are increasingly relevant to understanding women's reproductive psychology and fertility decision-making. In particular, fear of childbirth has been associated with lower fertility intentions among young, nulliparous women. Grounded in objectification theory, the present research examined whether self-objectification (internalized body surveillance rooted in appearance-based concerns) was associated with fertility intentions through fear of childbirth. Two complementary studies were conducted with Chinese nulliparous women aged 18-48 years (total N = 908): a cross-sectional survey (Study 1; n = 666, Mage = 26.46, SD = 4.23) and a randomized experiment (Study 2; n = 242, Mage = 26.16, SD = 4.89). In Study 1, self-objectification, fear of childbirth, and fertility intentions were assessed using validated measures. In Study 2, participants were randomly assigned to view either an objectifying video (designed to induce self-objectification) or a neutral control video, followed by assessments of fear of childbirth and fertility intentions. Across both studies, higher self-objectification was associated with greater fear of childbirth and lower fertility intentions. Mediation analyses indicated that fear of childbirth indirectly linked self-objectification to fertility intentions. These findings help connect objectification theory with reproductive psychology by highlighting body-image-related correlates of childbirth-related concerns. These findings also point to the potential value of social and psychological supports that promote young nulliparous women's reproductive mental health and safeguard their reproductive autonomy.
Chlamydia psittaci pneumonia (CPP) is a rare but potentially severe zoonotic disease.This study aimed to characterize the computed tomography(CT)findings of CPP and evaluate the utility of CT severity scores for predicting intensive care unit(ICU) admission. We retrospectively analyzed patients diagnosed with CPP between January 2022 and September 2025.Patients were divided into ICU and non-ICU groups based on disease severity and ICU admission.Demographic, clinical, laboratory, and radiological data were collected.Two radiologists independently reviewed CT images to record imaging features and calculate two scores: the chest CT score(CTS, range 0-25 based on lobar involvement)and the chest CT severity score(CTSS, range 0-40 based on 20 lung segments).Multivariable logistic regression and receiver operating characteristic(ROC)curve analyses were performed to identify predictors of ICU admission.We further compared the predictive performance of CTS with that of the CURB - 65 (confusion, urea, respiratory rate, blood pressure, age ≥ 65 years) score, a widely - utilized clinical severity assessment tool for community - acquired pneumonia. The CURB - 65 score was calculated for each patient upon admission. A total of 69 patients were included(45 male,24 female; mean age 61.6 ± 12.5years).A history of poultry or bird exposure was reported in 65 patients(94.2%).The ICU group comprised 22 patients(31.9%).Among survivors(n = 65),follow-up showed complete clinical recovery, and repeat CT imaging demonstrated complete resolution or marked improvement of interstitial abnormalities in those with available follow-up scans, supporting the reversible acute nature of these changes.In the ICU group, both the CTS (9.9 ± 6.4 vs. 5.9 ± 3.4, P = 0.011) and CTSS (13.5 ± 9.0 vs. 9.0 ± 4.8, P = 0.035) were significantly elevated. ROC analysis indicated that the area under the curve was 0.710 for CTS (cut - off 7.5; sensitivity 63.6%; specificity 76.6%) and 0.657 for CTSS. The CURB - 65 score yielded an area under the curve (AUC) of 0.633 (95% confidence interval [CI]: 0.476-0.790, P = 0.076). The DeLong test showed no statistically significant difference between the AUCs of the CTS and CURB - 65 (P = 0.452). Notably, only the CTS reached statistical significance in predicting intensive care unit (ICU) admission, implying that computed tomography (CT) - based severity assessment might offer prognostic information that cannot be captured solely by clinical score. This study provides a detailed imaging characterization of CPP.CT severity scores, particularly the CTS, are independently associated with ICU admission and may serve as adjunctive tools for early risk stratification.The reversible nature of interstitial changes on follow-up supports their acute inflammatory origin.Compared with CURB-65,the CTS offered a numerically higher AUC and provided significant prognostic information where clinical scoring alone did not.
The Bonwill triangle, defined by the mandibular incisor (MI) point and the center of right (CR) and left (CL) condyles, which provide a crucial reference for determining craniofacial symmetry and occlusion. Although three-dimensional imaging has enhanced the precision of triangle measurement, few studies have evaluated Bonwill triangle geometry in patients who have undergone orthognathic surgery (OGS). The present study assessed Bonwill triangle geometry in a Taiwanese population by comparing individuals who underwent OGS and those who did not and by analyzing the effects of sex and age on mandibular asymmetry. Cone-beam computed tomography images from 109 adults (54 in the OGS group and 55 in the non-OGS group) were retrospectively analyzed. Three side lengths of the Bonwill triangle (mandibular incisor point to left condyle (MI-CL), mandibular incisor point to right condyle (MI-CR), and right condyle to left condyle (CR-CL)) were measured using Mimics software. Group comparisons and subgroup analyses by sex and age were conducted using independent and paired t tests and Pearson correlation analysis. The OGS group exhibited greater asymmetry in the bilateral side lengths than the non-OGS group did (3.41 ± 2.35 mm vs. 1.69 ± 1.02 mm, p < 0.001), particularly the men in the group (p < 0.001). Additionally, only the men in the OGS group exhibited a negative correlation between age and bilateral side length (r = - 0.480, p = 0.034). CR-CL length did not differ significantly between the OGS and non-OGS groups. The Bonwill triangle can support preoperative mandibular asymmetryassessments. Candidates for OGS, especially men, exhibit greater skeletal asymmetry than non-OGS candidates do, underscoring a need for individualized planning. Future studies evaluating surgical type and long-term outcomes can enhance the clinical applications of the Bonwill triangle in pre-OGS assessments. This retrospective study was approved by the Institutional Review Board of China Medical University Hospital, Taichung, Taiwan (CMUH 114REC2019).
Imaging modalities play a critical role in determining surgical versus conservative management for distal radius fractures. This study aimed to evaluate the impact of two-dimensional (2D) and three-dimensional (3D) computed tomography (CT) on the decision to operate in distal radius fractures and to compare their influence between AO Type B and Type C fractures. This cross-sectional, survey-based study included 97 orthopedic and traumatology specialists. Twelve distal radius fracture cases classified according to the AO system (six Type B and six Type C) were selected. Participants were sequentially presented with plain radiographs, post-reduction radiographs in cast, 2D CT images, and 3D CT reconstructions for each case. After each imaging stage, participants were asked to indicate their decision to operate (surgical or conservative). Changes in the decision to operate were statistically analyzed. Among AO Type B fractures, the addition of CT imaging to plain and post-reduction radiographs did not significantly change the decision to operate in most cases (p > 0.05). In contrast, among AO Type C fractures, the addition of 2D CT imaging significantly changed the decision to operate in favor of surgical management (p < 0.05), whereas the subsequent addition of 3D CT did not produce a further significant change (p > 0.05). For AO Type B distal radius fractures, the addition of CT imaging to plain and post-reduction radiographs had limited impact on the decision to operate. In AO Type C fractures, 2D CT imaging significantly influenced the decision to operate, whereas the subsequent addition of 3D CT did not provide an additional impact on the decision to operate. Descriptive survey study.
Colorectal cancer is one of the most prevalent malignant tumors worldwide. Early screening relies on accurate polyp detection during colonoscopy. Polyps in colonoscopic images exhibit diverse morphologies, indistinct boundaries, and low contrast. Specular reflections, fold occlusions, and imaging artifacts further complicate detection, which fail to meet the requirements of real-time clinical assistance. To address these challenges, we propose BCP-YOLO (You Only Look Once), a high-precision, relatively lightweight polyp detection framework built upon an improved YOLOv8 architecture, designed to achieve a well-balanced trade-off between detection accuracy and computational efficiency. First, to mitigate complex background interference and improve small polyp detection, a BiFormer module is integrated into the backbone network to enhance focus on salient polyp regions while suppressing noise. To alleviate boundary ambiguity, the CARAFE content-aware upsampling operator is incorporated into the feature fusion stage, to refine lesion boundaries and spatial details. PConv module is employed to optimize network efficiency, reducing computational cost while maintaining detection performance. Experimental results on the Kvasir-SEG and CVC-ClinicDB datasets demonstrate that BCP-YOLO achieves a mean average precision (mAP0.5) of 88.5% on Kvasir-SEG, representing a 3.4% improvement over the YOLOv8 baseline. Precision and recall increase by 5.5% and 1.3%, respectively. The model contains 11.7 M parameters and achieves an inference speed of 104.1 frames per second (FPS). Five-fold cross-validation on both datasets validates its strong generalization capability and robustness. The method provides a high-accuracy and deployable solution for computer-aided diagnosis in real-time colonoscopy, offering significant potential to improve the reliability and efficiency of early colorectal cancer screening.
To develop a method for accurate identification and localization of facial acupoints based on geometric and topological relationships (GTRs) among facial key points and organ features. A facial acupoint localization framework was constructed using the Google MediaPipe machine learning toolkit to extract 478 facial key points. Geometric and topological relationships between predefined acupoints and key facial landmarks were established. Based on these relationships, a rule-based mapping algorithm was designed to identify and localize facial acupoints. The method was applied to facial images collected from individuals undergoing acupuncture and physical therapy, and its localization performance was evaluated. The proposed method successfully identified and localized facial acupoints based on stable geometric and topological relationships. The approach demonstrated consistent performance across different facial structures, enabling accurate positioning of acupoints without the need for large-scale annotated datasets. The results indicate that the method is feasible and reliable for practical applications. The GTR-based approach provides an effective and efficient solution for facial acupoint identification and localization, reducing dependence on manual annotation and improving applicability in clinical and intelligent acupuncture systems.
Oral squamous cell carcinoma (OSCC) is often preceded by oral potentially malignant disorders (OPMDs). Despite this known association, the transition from an OPMD to OSCC is complex, unpredictable, and non-linear, making early detection and intervention challenging for clinicians. Histopathological grading, the current standard for risk stratification, is not reliably predictive of malignant transformation (MT), and is subject to significant inter- and intra-observer variability. This scoping review evaluates emerging evidence on the integration of artificial intelligence (AI) and machine learning (ML) with molecular and histopathologic biomarkers to enable individualized risk assessment for MT. Ten retrospective studies incorporating AI/ML algorithms were analyzed, utilizing biomarkers ranging from gene expression panels, biochemical and protein-based markers like S100A7 to image-derived histomorphometric features. These models demonstrated promising predictive accuracy, with histology-derived features showing the greatest clinical feasibility. However, variability in methodologies, lack of prospective validation, and inconsistent demographic reporting limit the generalizability of the findings. This review highlights the need for multimodal biomarker integration, prospective clinical trials, and validation across broader populations. Ultimately, AI/ML-enhanced tools hold significant potential to inform personalized surveillance and treatment decisions in OPMD, but their clinical readiness requires further refinement and robust validation.
Alzheimer's disease (AD) and dementia with Lewy bodies (DLB) co-occur frequently, and growing evidence, including neuropathology, supports synergistic interplay between the diseases. We tested whether a single T1-weighted MRI scan may differentiate neuropathologically confirmed comorbid AD/DLB and AD controls using heterogeneously acquired neuroimaging. We obtained structural neuroimaging, on two groups, AD with and without DLB pathology. Convolutional neural networks are trained across dimensions. We introduce a triple-ensemble strategy consisting of majority voting schemes within a variety of plane permutations. In addition, we conduct voxel-wise statistical analyses. Here we show convolutional neural networks record a classification accuracy of 0.820 and an f1 score of 0.79 in identifying comorbid DLB/AD from AD patients. Prediction accuracy is higher proximal to date of death, while the trained model largely outperforms clinical baseline diagnosis. The slice-level performance varies depending on the sampled brain location, with sensitivity highest in the temporal lobe and specificity highest in the occipital lobe. In DLB/AD, gray matter is relatively preserved though atrophy is observed in the occipital lobe, suggesting that the comorbidity differentially affects brain loss and may accelerate it in the occipital lobe. This study demonstrates how machine learning approaches can address diverse neuroimaging data from clinical sources to differentiate neurodegenerative diseases using a true gold standard of neuropathological confirmation. The frameworks utilized here can be extended to other diseases that are frequently co-occurring and feasibly extend to single scan diagnostic clinical utility of scans already being acquired. Alzheimer’s disease (AD) and lewy body dementia (LBD) are both progressive neurodegenerative diseases that result in declines in memory and other cognitive functions. AD and LBD are known to occur together, however conventional methods that image the brain are unable to accurately show the effect of these diseases. Specialized computational methods may be able to more accurately diagnose these conditions than clinical evaluation. We used brain imaging data that is commonly obtained to see whether we could identify people with suspected AD who also had LBD. Our method identified the places in the brain whether both AD and LBD are likely to occur and was able to better diagnose in patients with confirmed disease at death that usual clinical diagnosis methods. Our method could be used to better identify people with AD and LBD and thus allow earlier appropriate treatment.
Recessive dystrophic epidermolysis bullosa (RDEB) is a severe skin disease caused by loss-of-function pathogenic variants in COL7A1 encoding type VII collagen (C7). Patients with RDEB suffer since birth from skin and mucosal blistering and develop severe local and systemic complications resulting in poor prognosis. Mesenchymal stromal cells (MSCs) have demonstrated their potential to enhance wound healing and reduce skin inflammation in RDEB patients due to their anti-inflammatory properties and capacity to express C7. We aim to optimize in vitro conditioning of human bone marrow-derived MSCs (BM-MSCs) to improve their limited survival following local injection in a murine model. BM-MSCs from healthy human donors were transduced with a lentiviral vector encoding firefly luciferase and mCherry reporter proteins and then subjected to various culture conditions: monolayer on plastic or spheroid culture, either in hypoxia (5% O2) or in normoxia (21% O2). These cells were subsequently injected intradermally (ID) in immunodeficient mice and their survival was assessed by in vivo imaging. BM-MSCs populations were analyzed prior to injection by single-cell RNA sequencing (scRNAseq). Murine skin injected with BM-MSCs were sampled two months post-injection and the surviving subpopulations were characterized by spatial transcriptomic. scRNAseq analysis revealed marked variations between monolayer and spheroid conditions, which were significantly impacted by oxygen level. Although most injected cells gradually died within the first 2 months in all tested conditions, 1% of live bioluminescent cells persisted for 54 up to 61 weeks post-injection. Spatial transcriptomics data analysis demonstrated that all surviving cells, regardless of their in vitro preconditioning, retained the expression of the MSC markers THY1, ENG, and NT5E, shared features with fibroblasts, and exhibited enriched expression in genes related to extracellular matrix and collagen fibril organization, which are key processes in wound healing. Spatial transcriptomic and scRNAseq data integration suggested that surviving BM-MSCs were initially present in the injected population, regardless of their culture condition. Remarkably, none of the in vitro preconditioning strategies appeared to affect their survival capacity or functional properties following local injection. The identification and characterization of a BM-MSC subpopulation capable of long-term survival following ID injection hold promise for the development of improved cell therapy protocols for RDEB.
The need for a cost-effective, rapid, and increasingly accessible alternative to the 21-gene assay prompted this study, which developed a novel MRI-based intratumoral heterogeneity score (ITHscore) to quantify tumor heterogeneity and integrated it with radiomic and clinical features to predict the 21-gene recurrence score (RS). This retrospective study included ER+/HER2- breast cancer patients who underwent 21-gene assay and preoperative MRI at our institution (April 2017-March 2019). Patients were randomly split into training (70%) and internal test (30%) cohorts, with an external test cohort from the public Duke-Breast-Cancer-MRI dataset. Tumor volumes of interest were automatically segmented using a pre-trained Scalable and Transferable U-Net framework, followed by k-means clustering to compute the ITHscore. Predictive models for the RS were built using clinical, radiomics, and ITHscore features with the support vector machine method, and evaluated by receiver operating characteristic curves. The institutional dataset comprised 452 patients (training: 316 (187 high-risk, 129 low-risk); internal test: 136 (80 high-risk, 56 low-risk)), while the external Duke cohort included 230 patients (44 high-risk, 186 low-risk). The ITHscore was significantly elevated in high-risk patients (p < 0.001), and its incorporation into the clinical-radiomic model improved RS prediction, yielding AUCs of 0.86 for the internal test cohort and 0.82 for the external test cohort. In this exploratory study, the ITHscore, which held promise as a noninvasive and intuitive means of characterizing intratumoral heterogeneity, demonstrated a potential incremental value for predicting RS in patients with ER+/HER2- breast cancer. The MRI-derived quantification of intratumoral heterogeneity facilitates simple and quantitative assessment of tumor heterogeneity and demonstrates potential incremental value in providing a rapid, cost-effective, and accessible prediction of the recurrence score in ER+/HER2- breast cancer. There is a need for a cost-effective, rapid, and increasingly accessible alternative to the 21-gene assay for predicting recurrence risk in ER+/HER2- breast cancer. The intratumoral heterogeneity score demonstrated potential as a noninvasive, intuitive, and quantitative biomarker for characterizing intratumoral heterogeneity and predicting recurrence score.
Early and accurate detection of brain tumors is clinically valuable for improving prognosis and guiding treatment. Existing deep-learning methods for magnetic resonance imaging (MRI) brain tumor detection face three difficulties: weak texture at lesion boundaries impairs localization; heterogeneous lesion scales degrade multi-scale detection; and non-maximum suppression (NMS) post-processing limits end-to-end inference. We propose a topology-decoupled end-to-end detection framework based on boundary-preserving feature flow and inter-channel correlation (ICC) distillation. A high-capacity teacher combines a multi-gradient-flow backbone with a gather-and-distribute global fusion mechanism, capturing both pathological boundary textures and anatomical context; a lightweight student is then derived by removing the global-fusion neck while retaining the isomorphic backbone. After comparing five feature distillation methods, we adopt ICC distillation, which aligns Gram matrices of intermediate features and mitigates the background-dominated bias common in medical imaging. Across three random seeds, the ICC-distilled student reaches mAP@0.5 = [Formula: see text], surpassing the plain student ([Formula: see text]) and matching or exceeding the teacher ([Formula: see text]). On the BraTS small-lesion stratum it attains 98.7% lesion recall with a false-positive-per-image rate of 0.014. The student achieves this at low cost (6.09M parameters, 11.7 GFLOPs, 168 FPS), suiting resource-constrained clinical deployment.
Severe trauma in patients aged ≥ 65 years is increasingly frequent. Frailty and sarcopenia are major prognostic factors, yet Total Psoas Area (TPA), an objective imaging biomarker of muscle mass, is usually measured at a single time point. We investigated the relationship between frailty, initial TPA, and its post-traumatic dynamics. observational retrospective cohort study of patient's ≥ 65 years with severe trauma and ≥ 2 CT scans from the from a level 1 trauma center (2018-2023). TPA was measured on axial L3 CT slices and indexed to height2. Pre-trauma frailty was assessed using the Clinical Frailty Scale (CFS). Longitudinal TPA changes were analyzed with a multivariable GEE model, testing interactions between age and frailty. Among 76 patients, 56 were non-frail (CFS 1-3) and 20 frail (CFS ≥ 4). Initial TPA did not differ between groups. Longitudinal analysis revealed divergent trajectories: TPA decreased more sharply with age in frail patients (-17.7 mm2/m2/year) compared with non-frail (-1.1 mm2/m2/year). This significant age × frailty interaction indicates that, at younger ages, TPA is comparable or even higher in frail patients, whereas at older ages it becomes substantially lower. Time since trauma had only a modest effect on TPA evolution. In the multivariable Cox model, non-frailty and higher GCS at admission were independently associated with lower mortality, while age, ISS, ASA score, and antiplatelet therapy were not significant. TPA showed a borderline association without reaching statistical significance. The model demonstrated good discrimination (C-index 0.812). In elderly trauma patients, TPA dynamics are strongly influenced by frailty and age, rather than initial TPA alone. Early assessment of frailty and muscle mass can refine risk stratification and guide personalized rehabilitation and nutritional strategies.
Tumor budding (TB) is a critical prognostic biomarker in stage Ⅱ colorectal cancer (CRC), but postoperative histopathological assessment limits preoperative risk stratification. Dual-layer spectral detector CT (SDCT) quantitative parameters may non-invasively predict TB grade and prognosis in stage Ⅱ CRC. We retrospectively studied 153 stage II CRC patients undergoing SDCT prior to curative resection. Spectral CT parameters (iodine concentration [IC], effective atomic number [Zeff], the slope of the energy spectrum [λHU]) during arterial phase (AP) and venous phase (VP) were analyzed. TB grade was assessed by pathology; three-year disease-free survival (DFS) was the primary endpoint. Logistic regression and Cox regression identified predictive factors. SDCT parameters (A-IC, A-Zeff, A-λHU) during AP significantly differentiated high-grade TB (Bd 3) from low/intermediate (Bd 1+2) (all P <0.05). The combined model integrating these parameters achieved AUCs of 0.937 (training set) and 0.893 (validation set). The value of the combined model was significantly associated with worse three-year DFS (HR 8.811, P=0.005). SDCT quantitative parameters during AP effectively stratify TB grade and predict prognosis in stage Ⅱ CRC, offering a non-invasive preoperative tool for personalized therapy.