Deep-learning-based image generation and classification have significantly advanced computer-aided diagnosis. However, the scarcity of high-quality labeled tongue image datasets for colorectal cancer (CRC) diagnosing tasks severely limits model generalization. While conventional data augmentation fails to mimic complex clinically relevant pathological variations, generative data augmentation offers a promising alternative. This work proposes DTMG-Net, an improved DDIM generative augmentation framework. We embed the DeepSeekMoE module into the denoising U-Net to realize adaptive modeling across diffusion timesteps of multi-scale pathological features. A multi-scale dilated attention (MSDA) block is introduced in the bottleneck to jointly capture global tongue structures and subtle local lesion textures. Moreover, a joint loss combining standard noise estimation and intra-sample diversity regularization is designed to alleviate mode collapse and expand the visual diversity of synthetic tongue samples. We adopt FID and IS to quantitatively evaluate the quality of generated images, and perform downstream classification experiments on four mainstream backbone networks to verify the effectiveness of synthetic data. Compared with VAE, DCGAN, PNDM and original DDIM, DTMG-Net achieves lower FID values of 73.83 and 58.99 for CRC and HC tongue samples respectively. Although these values remain relatively high in absolute terms, our method attains the minimal distribution discrepancy among all compared generative models under the small-sample tongue image setting. Ablation experiments indicate that DeepSeekMoE and MSDA both contribute to improved generation performance, and the proposed diversity constraint elevates IS without obvious FID degradation. In downstream classification tasks, training sets augmented with DTMG-Net synthetic images achieve generally higher numerical AUC, F1-score and Accuracy on WideResNet, ResNet50, MedMamba and ViT among the compared augmentation strategies. The proposed DTMG-Net can generate high-diversity, high-fidelity tongue samples to effectively expand small-scale datasets. The intra-sample diversity loss balances reconstruction quality and visual richness of generated content. Without complex preprocessing operations, this diffusion-based generative augmentation scheme achieves generally the highest numerical classification performance among the compared strategies on various classification backbones, supporting the potential of combining improved MoE and multi-scale attention modules for small-sample tongue image augmentation tasks.
The role of sex in Alzheimer’s disease (AD) is gaining increasing attention. However, the extent to which the same AD pathological burden affects the brain differently between sex remains unknown. Gait is increasingly recognized as a reflection of brain function and is also considered a potential indicator for evaluating early AD, yet the role of sex differences in the relationship between beta-amyloid (Aβ) pathology and gait remains unclear. We conducted a cross-sectional study of 253 cognitively normal or mildly impaired participants (96 men, 157 women) aged 55–85 years. All subjects underwent gait testing, plasma Aβ42 and Aβ40 quantification, neuropsychological evaluation, and MRI scanning. Participants were stratified by sex and plasma Aβ42/40 ratio. The group differences in gait parameters and gray matter structure were explored. Aβ(+) females exhibited faster gait speeds and showed a negative correlation between right CA1 head volume and swing velocity-average-right (r = − 0.51, p = 0.003), swing velocity-minimum-right (r = − 0.36, p = 0.046) and step speed-minimum-right (r = − 0.36, p = 0.042). However, there was no significant correlation between gait indicators and neuroimaging parameters in the Aβ(-) females. Aβ(+) males exhibited left hippocampal body atrophy (p = 0.040) without gait changes. The impact of Aβ on gait was specific to females, which might be attributed to enlargement of the right hippocampal head.
Patients with temporal gliomas often experience severe cognitive impairment after surgery, despite the attention given to the protection of cognitive structures during surgery. Connectomics provides a unified, network-based view dedicated to mapping and understanding the structural and functional topology of the human brain. This study aims to elucidate the characteristics of temporal glioma invasion into brain networks (including the salience network, central executive network, default mode network, dorsal attention network, and ventral attention network) and summarize the clinical effectiveness and practical experience in protecting cognitive function under the guidance of connectomics. First, connectomics was utilized to construct personalized brain functional networks for each patient, enabling a retrospective analysis of the invasion characteristics of 53 temporal gliomas for both functional structures and brain networks. The therapeutic effects of connectomics-optimized surgical programs (the experimental group) and traditional surgical strategies (the control group) on patients’ brain structures and cognitive functions were subsequently compared in 29 patients with temporal gliomas and 18 patients with left temporal gliomas. In 52/53 patients (98.11
Accurate risk stratification for non-muscle-invasive bladder cancer (NMIBC) recurrence is essential for personalized treatment. This study aimed to develop and validate a contrast-enhanced computed tomography (CT)-based model for predicting postoperative NMIBC recurrence within two years. We retrospectively enrolled 192 patients with NMIBC from three centers between June 2017 and May 2023. Center 1 formed the training set (n = 139; 74 recurrence events), and Centers 2 and 3 formed the test set (n = 53; 24 recurrence events). Super-resolution reconstruction based on a generative adversarial network (GAN) was used to enhance contrast-enhanced CT images, from which radiomics and deep-learning (DL) features were extracted. Clinical predictors were evaluated using univariable and multivariable logistic regression. Fusion Model C (FMC), integrating radiomics, DL, and clinical features, was developed exclusively in the training set and evaluated in the test set. Model interpretation used Shapley additive explanations (SHAP), and performance was assessed using the area under the receiver operating characteristic curve (AUC), calibration curves, Brier scores, and decision-curve analysis. FMC outperformed the radiomics model (test set AUC = 0.754), achieving AUCs of 0.917 (95
This study aimed to evaluate changes in computed tomography-derived fractional flow reserve (CT-FFR) and qualitative/quantitative plaque characteristics before and after transcatheter aortic valve replacement (TAVR), and assess the correlation between CT-FFR changes and plaque progression/regression. We retrospectively analyzed coronary computed tomography angiography (CCTA) data from patients with coronary stenosis (30–90
Habitat imaging analysis has recently been applied to head and neck cancer (HNC) imaging research for various clinical applications. This review systematically evaluated current evidence on its performance in various HNC applications, highlighting its advantages and limitations. A systematic search of PubMed, Web of Science, and Scopus was conducted for articles from database inception to 30 January 2026, using terms related to habitat analysis, HNC, and medical imaging. Patient numbers, tumour origin, imaging modality, clustering features, habitat methods, subregion numbers, clinical applications, and performance of the models were extracted from the included articles. Twenty-five original articles published between 2019 and 2026 were eligible for analysis, of which 10 had external validation datasets. Nasopharyngeal carcinoma was the most studied HNC (n = 9); CT was the most used modality (n = 9); and K-means clustering was the predominant habitat method (n = 20). Habitat imaging analysis performed consistently well and outperformed conventional radiomics or clinical models in predicting pathological/immunohistochemical (IHC) biomarkers, cancer characterisation and nodal status, with areas under the curve (AUCs) of > 0.8, but showed variable results for treatment response prediction. Combining the multiple models slightly improved performances in most studies. Habitat imaging analysis showed potentials in various applications in HNC, performing consistently well in predicting pathological/IHC biomarker, tumour characterisation and nodal status, but not in treatment response prediction. Combining multiple models slightly improved performance, raising questions about whether such complex integrations offer clinical value. Current evidence remains heterogeneous, and so prospectively conducted and externally validated studies are required to clarify their practical utility.
To develop and validate a machine learning model based on clinical variables and multi-lesion, multiphase contrast-enhanced CT radiomics for predicting progression-free survival (PFS) in patients with unresectable colorectal liver metastases (CRLM). This retrospective cohort study included 159 patients with unresectable CRLM who received first-line chemotherapy. Patients were divided into a training cohort (n = 127) and an internal hold-out set (n = 32) using stratified sampling. All patients underwent contrast-enhanced CT before treatment. For each patient, all measurable liver metastases were segmented separately on arterial-phase (A) and portal venous-phase (V) images, and radiomic features were extracted. A feature selection pipeline based on the Criteria Importance Through Intercriteria Correlation (CRITIC) objective weighting method and multi-criteria decision making (MCDM) was applied. Multiple machine learning classifiers were trained on unimodal data, and the base learners were integrated using a performance-weighted soft-voting ensemble strategy. Model performance was evaluated using the area under the receiver operating characteristic curve (AUC), accuracy, sensitivity, specificity, calibration curves, and decision curve analysis (DCA). Model interpretability was assessed using Shapley additive explanations (SHAP). The trimodal arterial-phase, portal venous-phase, and clinical-variable ensemble model (AVC) showed encouraging discriminative performance in the internal hold-out set, with an AUC of 0.792 (95
The aim of this study was to evaluate the value of combining delta-radiomics and clinical characteristics in predicting the risk of EGFR T790M resistance mutation in patients with non-small cell lung cancer (NSCLC) prior to first-line treatment with epidermal growth factor receptor (EGFR) tyrosine kinase inhibitors (EGFR-TKIs). This retrospective study included non-small cell lung cancer (NSCLC) patients whose disease progressed after first-line treatment with an EGFR tyrosine kinase inhibitor (TKI) between January 2013 and September 2019. The patients were randomly assigned to the training and test cohorts. Four predictive models for the acquisition of the T790M mutation were developed on the basis of clinical characteristics, non-contrast-enhanced computed tomography (NECT) radiomics features, delta-radiomics features, and a combination of clinical and delta-radiomics features. The optimal model was selected using the area under the receiver operating characteristic curve (AUC) and decision curve analysis (DCA). Additionally, the Shapley additive explanations (SHAP) interpretability framework was applied to visualize and interpret the model’s decision-making process. A total of 285 patients were included in this study (213 in the training set and 72 in the test set). In the test set, the AUCs for the clinical model, NECT radiomics model, delta-radiomics model, and combined model were 0.70, 0.63, 0.75, and 0.82, respectively. The corresponding accuracies were 66.7
Directly measured waist- and hip-based indices are less well studied than body mass index (BMI) or CT-derived size measures in coronary CT angiography (CCTA). We examined whether waist-to-hip ratio (WHR) was associated with scanner-reported dose metrics and acquisition extent. This retrospective observational study analyzed anthropometric measurements recorded during the pre-CCTA clinical workflow and retrospectively extracted scanner dose reports from 134 consecutive adults. Cardiac CTDIvol, cardiac-helical dose-length product (DLP), total examination DLP, and CCTA scan length were evaluated. Primary multivariable models included body surface area (BSA), WHR scaled per 0.1-unit increase, age, sex, pre-scan heart rate, and beta-blocker administration. A primary path-based statistical decomposition of the WHR–cardiac DLP association through scan length was adjusted for BSA, age, and sex and used 10,000 bootstrap resamples; a sensitivity decomposition additionally adjusted for pre-scan heart rate and beta-blocker exposure. Among 134 participants, 81 (60.4
Differentiating Brucellar Spondylitis (BS) from Spinal Tuberculosis (STB) remains a clinical conundrum due to their overlapping radiological and clinical manifestations. Accurate early diagnosis is critical for determining appropriate antibiotic regimens and surgical interventions. This study aims to develop a non-invasive, multi-modal diagnostic framework utilizing 2.5D deep learning and Transformer mechanisms as a potential adjunct to assist differentiation. In this retrospective multicenter study, clinical and MRI data from 238 patients (169 BS, 69 STB) were analyzed, with class imbalance addressed during model development. Data from Center A (n = 198) were divided into training (n = 138) and internal validation (n = 60) sets; Center B provided an external test set (n = 40). We developed a multimodal differential diagnosis model that integrated clinical features, MRI radiomics, and a 2.5D deep learning network. Lesions were segmented (3D Slicer), radiomic features were extracted and screened, and a 2.5D model was constructed. A Transformer architecture then achieved deep fusion of multimodal features. Performance was evaluated using metrics such as the area under the curve (AUC), sensitivity, and specificity. ResNet18 demonstrated the best performance in slice-level prediction, achieving AUCs of 0.863, 0.766, and 0.791 in the training, validation, and test cohorts, respectively. The Transformer fusion model showed outstanding performance, achieving AUCs of 0.901 in the validation cohort, significantly outperforming multiple instance learning (0.812) and ensemble fusion (0.807). The final combined model yielded AUCs of 0.957 (95
Accurate segmentation of retinal blood vessels is the core technical basis for computer-aided diagnosis of multiple fundus diseases, including diabetic retinopathy and hypertensive retinopathy. Existing segmentation methods still face challenges when processing scale variations, complex structures, low contrast, and pathological interferences in fundus images, which restricts the clinical application value of related technologies. This study aims to construct a segmentation network to effectively improve the fine feature extraction and obtain more accurate and robust segmentation results. This study proposes a Wavelet-Refined Semantic-Guided Network (WRSG-Net) for retinal vessel segmentation. First, the network adopts Parametric Wavelet Refined Encoder (PWREC) to adaptively capture high-frequency vascular details via learnable wavelet decomposition, and refine features through a dual-attention mechanism to mitigate the loss of fine vessel details during downsampling. Then, an Inter-Scale Information Propagation Module (ISIPM) is introduced to establish a layer-wise information propagation flow for integrating hierarchical features, enhancing the consistency and continuity of the segmented vessels. Finally, a Semantic Guided Decoder (SGDC) is designed to perform context-aware feature aggregation under decoder semantic guidance and complete accurate pixel-level prediction, which suppresses noise from the encoder and alleviates blurring caused by decoder upsampling. (Code can be available from https://github.com/JerryYaoGl/WRSG-Net after the paper acceptance). The proposed method achieves competitive performance on all three test datasets. It obtains 85.17
Pretreatment diffusion weighted imaging (DWI) is effective to predict the response to chemotherapy for spine metastases of breast cancer, but current methods of analyzing DWI data with multiple b-values require a predefined mathematical model for curve fitting. This study proposes and validates a method of combining histogram signatures and deep learning to predict treatment response for spine metastases of breast cancer without the need of fitting MRI signals at multiple b-values into a predefined diffusion model. 37 patients with breast cancer with 191 spine metastatic lesions were included in this study. All patients underwent MRI scanning with DWI before chemotherapy and were divided into an investigating group (n = 19) and a test group (n = 18). Two radiologists created region of interest (ROI) by manually delineating lesion borders on pretreatment DWI images. DWI signals inside ROI at 10 b-values were converted into histograms and concatenated into a histogram signature. A convolutional neural network (CNN) was constructed to classify histogram signatures into a progressive disease (PD) group or a non-PD group. The predictive performance of eight fitted diffusion parameters, ADC, DDC, α, D*, Dtrue, frac, Dapp and Kapp, were compared with histogram signature by receiver operating characteristic (ROC) curves. The area under ROC curve (AUC) of PD prediction for spine metastases of breast cancer was 0.882 (95
This study aimed to develop a computed tomography urography (CTU)-based radiomics model for preoperative prediction of human epidermal growth factor receptor 2 (HER2) status in bladder cancer (BCa). This single-center retrospective study enrolled 190 patients with pathologically confirmed BCa who underwent preoperative CTU between October 2019 and September 2024. The cohort was randomly divided into training (n = 133) and test (n = 57) cohorts. Radiomics models were developed using Least Absolute Shrinkage and Selection Operator (LASSO) regression combined with five machine learning methods: logistic regression (LR), support vector machine (SVM), k-nearest neighbors (KNN), extreme gradient boosting (XGBoost), and random forest (RF). A clinical-radiomics model was subsequently established by integrating radiomics features with clinical parameters. Model performance was assessed using the area under the curve (AUC), calibration curves and Decision curve analysis (DCA). Among 190 patients (HER2-positive: 62.1
Although the relationship between coronary stenosis and myocardial perfusion is well established, the association between quantitative plaque parameters and myocardial perfusion remains poorly understood. This study investigated the association between quantified plaque parameters from coronary computed tomography angiography (CCTA) and myocardial flow reserve (MFR) derived from CT myocardial perfusion imaging (CT MPI). From April 2019 to October 2023, consecutive patients with suspicion of stable coronary artery disease (CAD) underwent CCTA and CT MPI. CCTA and CT MPI were performed with a third-generation dual-source CT scanner. Of the 399 patients, the median (interquartile range) age was 59.0 (53.0, 65.5) years, and 295 (73.9
Conventional MRI may show limited or equivocal changes in the early period after stereotactic radiosurgery (SRS) for brain metastases. This study evaluated whether early dynamic contrast-enhanced MRI (DCE-MRI) changes at 4–8 weeks after SRS are associated with concurrent early radiologic response in brain metastases. This prospective, exploratory, lesion-level study included 17 brain metastases from 12 consecutive patients treated with single-session Gamma Knife SRS. DCE-MRI was performed before treatment and at 4–8 weeks after SRS using a 1.5-T MRI system. Quantitative parameters (Ktrans, Ve, Kep, and initial area under the concentration–time curve [iAUC]) were derived using the Tofts pharmacokinetic model, and semiquantitative parameters (upslope and initial gradient) were obtained from signal-intensity curves. Response was classified according to RANO-BM criteria. Pre-to-post changes, between-group differences in absolute change, and exploratory associations with response category were analysed using Wilcoxon signed-rank, Mann–Whitney U, and Spearman correlation tests, respectively. Benjamini–Hochberg false discovery rate correction was applied within each test family. Six lesions (35.3
To compare the visibility of the mandibular canal (MC) at three mesial-distal levels adjacent to the mandibular third molar (M3M) using standard-dose and four low-dose cone-beam computed tomography (CBCT) protocols, and evaluate their impact on assessing the spatial relationship between the MC and M3M. A total of 100 CBCT scans were obtained from 20 dry mandibles with M3Ms, using one standard-dose, normal-resolution protocol (0.2 mm voxel; DAP: 322 mGy×cm2) and four low-dose protocols with varying resolutions (0.15-0.4 mm voxel; DAP: 23-131 mGy×cm2). For each scan, MC visibility at the distal, middle, and mesial thirds of the M3M was rated on a 4-point scale. The spatial relationship between the MC and M3M apex was evaluated both qualitatively (non-contact, contact, or penetration) and quantitatively (the MC-M3M distance). Pairwise comparisons between the standard-dose and each low-dose protocol were conducted using the Wilcoxon signed-rank test. The MC visibility at all levels was high (≥3.4/4) for all protocols except the low-dose, low-resolution protocol (voxel size: 0.4 mm; DAP: 23 mGy×cm2). Significantly lower MC visibility was observed only with the low-dose, low-resolution protocol compared to the standard-dose protocol (p = 0.006-0.026). No significant differences were found in the qualitative and quantitative assessments of the spatial relationship between the MC and M3M apex among any of the protocol pairs. Although the lowest-dose, lowest-resolution CBCT reduced MC visibility, no significant differences were observed between standard-dose and low-dose protocols in assessing the MC-M3M apex relationship. Notably, the lack of significant difference should not be interpreted as evidence of equivalence, as the study used subjective image assessments without an independent reference standard. Low-dose CBCT may be considered in selected cases, but protocol selection should remain indication-oriented, patient-specific, and device-specific.
The manual generation of radiology reports is time-consuming and error-prone, particularly for less experienced radiologists, leading to inefficiencies and delays in clinical workflows as case volumes increase. Therefore, automated radiology report generation can serve as valuable tool to support hospitals and improve efficiency. This paper presents an automated two-stage approach for radiology report generation, consisting of keyword prediction from X-ray images followed by text-to-text report generation. The publicly available IU X-ray dataset, comprising 3,851 patients with paired PA and lateral chest X-rays and corresponding reports, was divided into training (80
Pediatric physeal fractures can be difficult to assess radiographically. We characterized regional fracture-detection failure patterns of a multimodal large language model (MLLM) in confirmed pediatric physeal fractures. This retrospective, single-center, case-only study included 805 confirmed physeal fractures in children aged 0–12 years during 2025. The recorded GPT-5 API workflow was assessed in March-April 2026 using anonymized radiographs plus age, target anatomical region, and a complaint or mechanism. The reference standard was independent review by three orthopedic surgeons - one pediatric orthopedics specialist, one orthopedic trauma specialist, and one senior orthopedic surgeon, each with at least 8 years of clinical experience in orthopedics and traumatology - followed by consensus with model outputs concealed; no radiologist participated. The primary endpoint was a definite or probable fracture call in the correct anatomical region; explicit physeal recognition or correct Salter-Harris classification was not required. Salter-Harris category, anatomical region, and age were prespecified secondary, hypothesis-generating failure-pattern stratifiers. Each case was evaluated once. Because all cases were fracture positive, specificity, predictive values, overall accuracy, and receiver operating characteristic measures were not estimable. The model met the regional fracture-detection threshold in 584 of 805 cases (sensitivity, 72.5
Coronary artery calcium (CAC) is an established marker of cardiovascular risk, but CT-based CAC scoring is not universally performed. This study evaluated whether a vision-capable large language model (LLM) can classify clinically significant CAC (CAC > 100) from routine posteroanterior chest radiographs, using CT-derived Agatston score as the reference standard. This retrospective single-center diagnostic accuracy study included 349 patients who underwent both non-contrast cardiac CT for CAC scoring and chest radiography within 6 months. Patients were classified as CAC 0–100 or CAC > 100. Preprocessed chest radiographs were analyzed by GPT-5.5 in a zero-shot setting using a structured prompt. The model estimated the score of CAC > 100, assigned a binary class, and reported visible calcification findings. Diagnostic performance was assessed against CT-derived Agatston classification. Of 349 patients, 131 had CAC > 100. Using its native binary output, the LLM correctly classified 292 patients (accuracy 83.7
Although perioperative immunotherapy has shown promising results in locally advanced gastric adenocarcinoma (LAGA), the predictive biomarkers remain unclear. Our study aimed to investigate the predictive value of [18F]FDG PET/CT characteristics at baseline for the efficacy of neoadjuvant immunotherapy plus concurrent chemoradiotherapy in the post hoc analysis of the Neo-PLANET phase II trial. A total of 29 patients with LAGA receiving neoadjuvant immunotherapy plus concurrent chemoradiotherapy from the Neo-PLANET phase II trial (Zhongshan Hospital Fudan University, NCT03631615, 2018-08-13) were included in our study. Semiautomated techniques were used to analyze pre-operative [18F]FDG PET/CT images to determine primary tumor, nodal metabolic and spleen parameter scores [including max and mean adjusted for lean body mass standardized uptake values (SUV), metabolic tumour volume (MTV) and total lesional glycolysis (TLG)]. Kaplan–Meier method and log-rank test to evaluate the associations between PET/CT parameters and disease-free survival (DFS) and overall survival (OS). Pathological complete response (pCR) and major pathological response (mPR) were not related to DFS and OS in patients receiving neoadjuvant immunotherapy plus concurrent chemoradiotherapy. Interestingly, a higher SUVmean of the spleen (S-SUVmean) is positively associated with longer DFS (HR = 0.19, 95
