Objective To construct a machine learning diagnostic model for autism spectrum disorder (ASD) by integrating multiple resting-state functional magnetic resonance imaging (rs-fMRI) biomarkers, thereby supporting accurate diagnosis of ASD. Methods The rs-fMRI data of 1,420 subjects aged 5-18 years from the Autism Brain Imaging Data Exchange (ABIDE) database were included. Six rs-fMRI biomarkers were calculated, including resting-state functional connectivity (FC), amplitude of low-frequency fluctuation (ALFF), fractional amplitude of low-frequency fluctuation (fALFF), regional homogeneity (ReHo), degree centrality (DC), and voxel-mirrored homotopic connectivity (VMHC). Differential variables were screened using one-way analysis of variance (ANOVA) (P<0.05). The variance inflation factor (VIF) was applied to reduce multicollinearity, and forward feature selection was used to determine an important feature subset. Five machine learning models were constructed, including logistic regression (LR), extreme gradient boosting (XGBoost), light gradient boosting machine (LightGBM), support vector machine (SVM), and passive-aggressive classifier (PAC). Model performance was evaluated using the ABIDEⅠ as the training set and the ABIDE Ⅱ as the external validation set. Receiver operating characteristic (ROC) curves were plotted, and model performance was assessed using the area under the curve (AUC), accuracy, precision, and F1 score. The stability of the model was evaluated through robustness analysis. Results A total of 185 important features were identified, including 170 FC features, 6 DC features, 5 ALFF features, 3 fALFF features, and 1 VMHC feature. Among the five machine learning models integrating multiple rs-fMRI biomarker features, the LR model showed the highest diagnostic efficacy (in the training set, the AUC was 85.97%, the accuracy was 79.59%, the precision was 79.82%, and the F1 score was 79.52%; in the external validation set, the AUC was 85.14%, the accuracy was 77.97%, the precision was 78.01%, and the F1 score was 77.98%). Robustness analysis confirmed small fluctuations in model performance and excellent generalization stability. Conclusion The machine learning model integrating multiple rs-fMRI biomarkers can serve as an effective neuroimaging diagnostic tool for ASD and may provide objective technical support for early ASD screening.
Objective To investigate differences in two-dimensional amplitude of low-frequency fluctuation (2D-ALFF) among emotional symptom subtypes of functional constipation (FC) using resting-state functional MRI (rs-fMRI), and to identify neuroimaging biomarkers reflecting emotional status. Methods Seventy-one patients with FC and 42 healthy controls were prospectively enrolled. Clinical data, Self-rating Anxiety Scale (SAS) scores, Self-rating Depression Scale (SDS) scores, and rs-fMRI data were collected. Based on SAS and SDS scores, patients with FC were classified by K-means clustering into an FC with anxiety/depressive status group (FCAD, n=39) and an FC without anxiety/depressive status group (FCNAD, n=32). The DPABISurf toolkit was used to preprocess the rs-fMRI data and generate 2D-ALFF maps for both cerebral hemispheres. Brain regions with significant intergroup differences in 2D-ALFF were identified, and selected as the seed region for subsequent whole-brain resting-state functional connectivity (RSFC) analysis. Intergroup differences in 2D-ALFF and RSFC among the three groups were assessed using one-way analysis of covariance, with Monte Carlo simulation correction (vertex-level P<0.001, cluster-level P<0.025/hemisphere, 10 000 iterations). Post hoc pairwise comparisons were performed using LSD-t tests with Bonferroni correction (P<0.016). Receiver operating characteristic (ROC) curves were used to evaluate the performance of differential 2D-ALFF values in discriminating FCAD from FCNAD. Partial correlation analyses were conducted to examine correlations between brain functional parameters and scale scores, with Bonferroni correction (P<0.001 8). Results Significant intergroup differences in 2D-ALFF were observed in the right limbic network temporal pole area 4 (Limbic_TempPole_4) and the right default mode network parietal area 3 (Default_Par_3) (P<0.025). The 2D-ALFF value of the right Default_Par_3 showed high diagnostic performance for distinguishing FCAD from FCNAD, with an area under the curve (AUC) of 0.830. Using this region as the seed, RSFC analysis revealed significant intergroup differences (P<0.025) involving regions in the default mode, sensorimotor, frontoparietal, and ventral attention networks. In addition, the 2D-ALFF value of the right Default_Par_3 was negatively correlated with SAS score (r= -0.391, P=0.000 7). Conclusion Abnormal 2D-ALFF regions in patients with FCAD are predominantly located in the right (non-dominant) hemisphere. The 2D-ALFF value of the right Default_Par_3 has high diagnostic value for differentiating FCAD from FCNAD, shows heterogeneous functional connectivity with multiple brain networks, and is correlated with emotional status. It may serve as a neuroimaging biomarker for identifying potential individualized therapeutic targets in patients with FCAD.
Objective To evaluate the predictive value of the low-density lipoprotein cholesterol/high-density lipoprotein cholesterol (LDL-C/HDL-C) ratio combined with the pericoronary fat attenuation index (FAI) for coronary in-stent restenosis (ISR) in patients with type 2 diabetes mellitus (T2DM) after percutaneous coronary intervention (PCI). Methods A total of 185 patients with T2DM who underwent PCI, received preoperative coronary computed tomography angiography (CCTA), and underwent follow-up invasive coronary angiography (ICA) after PCI were retrospectively enrolled, involving 216 target lesions. Patients were divided into ISR (n=44) and non-ISR (n=141) groups. FAI at the target lesion site was measured, and the LDL-C/HDL-C ratio was calculated. Maximally selected Log-rank test was used to determine the optimal cutoff values for FAI and the LDL-C/HDL-C ratio. Event-free survival was analyzed using the Kaplan-Meier method, and survival differences among subgroups were compared using Log-rank test. Cox regression analysis was used to identify risk factors for ISR and to construct three ISR prediction models: Model 1 was based on clinical risk factors, Model 2 added the LDL-C/HDL-C ratio to Model 1, and Model 3 added FAI to Model 2. Model performance was evaluated using Harrell’s C statistic (C-index), continuous net reclassification improvement (NRI), integrated discrimination improvement (IDI), and time-dependent receiver operating characteristic (ROC) curves with area under the curve (AUC) values at 1, 3, and 5 years. The DeLong test was used to compare C-indices among models, and bootstrap resampling (B=1,000) was used to compare NRI and IDI values between models. Results The median follow-up time was 19.88 (12.30, 40.87) months, and the ISR occurred in 23.78% of the poctients. The optimal cutoff values for predicting ISR were -67.64 HU for FAI and 2.06 for the LDL-C/HDL-C ratio. The risk of ISR was significantly higher in the double-positive group (both parameters above their respective cutoff values) than in the double-negative or single-positive groups (Log-rank test, P<0.001). Multivariate analysis showed that hypertension (HR=2.01), higher FAI (HR=1.08), and a higher LDL-C/HDL-C ratio (HR=1.40) were independent risk factors for ISR in patients with T2DM (all P<0.05). Model 3 showed significantly better predictive performance than Model 2 (C-index, 0.81 vs. 0.78; Z=3.07, P=0.048; NRI=0.78; IDI=0.67; both P=0.001). The AUCs of Model 3 at 1, 3, and 5 years were 0.91, 0.76, and 0.89, respectively. Conclusion Adding the LDL-C/HDL-C ratio and FAI to conventional clinical risk factors significantly improves the prediction of ISR, providing incremental prognostic value for ISR risk assessment in patients with T2DM after PCI.
Objective To investigate the feasibility and clinical value of a mammography-based radiomics logistic regression (LR) classifier model combined with clinical and mammographic features for predicting axillary lymph node (ALN) metastasis. Methods A total of 355 female patients with breast cancer who underwent mammography at two hospitals and were confirmed by surgical pathology were retrospectively enrolled. Patients from one hospital (n=279) were randomly divided into a training set (n=195) and an internal validation set (n=84) at a 7∶3 ratio, and patients collected during the same period from the other hospital (n=76) served as an external test set. According to postoperative pathology of the ALNs, all patients were assigned to the ALN metastasis group (n=142) and non-metastasis group (n=213). Radiomic features were extracted from mammographic images, followed by dimensionality reduction and feature selection. An LR model based on radiomic features was constructed, and the radiomics score (Radscore) was calculated from the output probability. Laboratory results, imaging T stage, and mammographic features including maximum mass diameter, shape, margin, and density were analyzed to identify clinical and mammographic indicators with statistically significant differences. Multivariate logistic regression was used to identify independent predictors of ALN metastasis, which were combined with Radscore to build a combined model. Receiver operating characteristic (ROC) curves were used to evaluate predictive performance. The DeLong test was used to compare AUCs among models. Calibration curves were used to assess goodness of fit for the better-performing models, and decision curve analysis was performed to evaluate clinical net benefit. Results Ten optimal radiomic features were selected to construct the LR model and calculate the Radscore. CA153, imaging T stage, and maximum mass diameter were independent predictors of ALN metastasis and were used to construct the clinical model, mammographic feature model, and combined model with the Radscore, respectively. The combined model achieved the highest AUC and sensitivity for predicting ALN metastasis (DeLong test, all P<0.001). The combined model and the LR model showed good agreement between their predictions of ALN metastasis and the actual clinical observations, and both demonstrated a high clinical net benefit. Conclusion The mammography-based radiomics LR classifier model and the combined model can accurately predict ALN metastasis before surgery and may support individualized diagnosis and treatment decision-making.
Objective To investigate gestational age-dependent changes in multiple ultrasound parameters of normal fetal lung maturation during the third trimester of pregnancy, assess the strength of their associations with gestational age, and establish normal reference ranges for these parameters. Methods A total of 300 pregnant women at 30-41+6 weeks of gestation were retrospectively enrolled. According to gestational age, participants were divided into six groups: 30-31+6 weeks, 32-33+6 weeks, 34-35+6 weeks, 36-37+6 weeks, 38-39+6 weeks, and 40-41+6 weeks, with 50 cases in each group. Two-dimensional ultrasound was used to measure fetal lung maturation parameters, including the diameter of the main pulmonary artery (MPA), right lung area, and right lung perimeter. Color doppler ultrasound was employed to obtain MPA flow velocity waveform parameters, including the pulsatility index (PI), resistance index (RI), acceleration time (AT), ejection time (ET), and AT/ET ratio. One-way ANOVA was used to compare the differences in ultrasound parameters among gestational age groups. Pearson correlation analysis was employed to examine the relationships between ultrasound parameters and gestational age, and regression curves were plotted. The 95% reference range of the ultrasound parameters at different stages of the third trimester was determined. Results Significant differences were observed in PI, AT, AT/ET, MPA diameter, right lung area, and right lung perimeter among the gestational-age groups (all P<0.05), whereas no significant differences were found for RI or ET (both P>0.05). Fetal PI showed a very strong negative correlation with gestational age (r=-0.811, P<0.001). AT, AT/ET, MPA diameter, right lung area, and right lung perimeter demonstrated strong positive correlations with gestational age (r=0.658, 0.675, 0.663, 0.761, and 0.759, respectively; all P<0.001). Further research findings indicate that with increasing gestational age, PI shows an overall decreasing trend, while AT, AT/ET, MPA diameter, right lung area, and right lung circumference show an overall increasing trend. Conclusion Ultrasound multi-parameter assessment can effectively reflect fetal lung development in late pregnancy. This study preliminarily explored the changing patterns of multiple ultrasound parameters for fetal lung development in the third trimester of pregnancy, which may provide an objective imaging basis for clinical non-invasive assessment of fetal lung development.
Objective To investigate the clinical value of a model integrating MRI-based tumoral and peritumoral radiomics features with clinical factors for differentiating benign from malignant soft tissue tumors (STTs). Methods Patients with pathologically confirmed STTs from two medical centers were retrospectively enrolled. Patients from Center 1 (n=295) were randomly divided into a training set (n=206) and an internal validation set (n=89) at a ratio of 7∶3, and patients from Center 2 (n=106) served as the external test set. Tumoral and peritumoral radiomics features were extracted from fat-suppressed T2-weighted imaging (FS-T2WI) and contrast-enhanced T1-weighted imaging (CE-T1WI). A multilayer perceptron (MLP) algorithm was used to construct tumor, peritumoral, and tumor-peritumoral radiomics models. SHapley Additive exPlanations (SHAP) analysis was used to interpret the contribution of each feature to model differentiation. A combined model was developed by integrating the best-performing radiomics model with the selected clinical factors. Receiver operating characteristic (ROC) curves and the area under the curve (AUC) were used to evaluate model performance. Calibration curves were used to assess agreement between predicted probabilities and observed outcomes, and decision curve analysis (DCA) was used to evaluate clinical net benefit across different threshold probabilities. Results The tumor-peritumoral radiomics model constructed using 8 tumoral features and 9 peritumoral features showed the best performance among the three radiomics models, with AUCs of 0.941, 0.876, and 0.836 in the training, internal validation, and external test sets, respectively. After this model was combined with the clinical factors of age and maximum tumor diameter, the combined model showed further improvement, with AUCs of 0.969, 0.902, and 0.851 in the training, internal validation, and external test sets, respectively. SHAP further revealed the contribution of each feature to the model’s output. DCA and calibration curves demonstrated that the combined model had favorable clinical net benefit and good calibration. Conclusion The combined model integrating tumor-peritumoral radiomics features with clinical factors provides a noninvasive and relatively accurate method for differentiating benign from malignant STTs and has high clinical value.
Objective To evaluate the predictive value of an MRI-based deep learning model for knee osteoarthritis (KOA). Methods This retrospective study included 305 patients who underwent knee MRI because of knee pain, morning stiffness, or related symptoms, including 174 males and 131 females, with a mean age of (62.35±6.10) years. The patients were randomly divided into a training set (n=214) and a validation set (n=91) at a ratio of 7∶3. According to the diagnostic criteria for KOA, the training set was divided into a KOA group (n=128) and a non-KOA group (n=86), while the validation set was divided into the KOA group (n=55) and the non-KOA group (n=36). Meniscal MRI radiomics features were extracted using Pyradiomics 3.0.1. Least absolute shrinkage and selection operator (LASSO) regression and 10-fold cross-validation were used to select the optimal subset of quantitative meniscal features for predicting KOA in the training set. Random forest (RF), support vector machine (SVM), and extreme gradient boosting (XGBoost) prediction models were then constructed. The receiver operating characteristic (ROC) curve was used to evaluate the performance of the predictive model, and its area under the curve (AUC), sensitivity, and specificity were calculated. The DeLong test was employed to compare the differences in AUC values among different predictive models. Calibration curves and decision curve analysis were used to assess the calibration and clinical utility of the predictive model. Results Nine optimal radiomics features for predicting KOA were selected, including 2 first-order statistical features, 4 texture features, 1 shape feature, and 2 derived features from wavelet transform. The ROC curve results showed that the AUC values of the XGBoost prediction model in both the training set and the validation set were higher than those of the RF and SVM prediction models (all P<0.05). Calibration curves showed good agreement between predicted probabilities and observed outcomes for the RF, SVM, and XGBoost models in both the training and validation sets. Decision curve analysis showed that the XGBoost model achieved greater net benefit across a relatively wide range of threshold probabilities (training set, 0.05-0.82; validation set, 0.03-1.00). Conclusion The XGBoost prediction model based on automatic MRI meniscus segmentation demonstrated good diagnostic performance and may serve as a reliable quantitative tool for assisting early diagnosis and clinical decision-making in KOA.
Early identification of cognitive impairment associated with type 2 diabetes mellitus (T2DM) and elucidation of its underlying mechanisms have become major areas of research interest. MRI, including structural MRI, arterial spin labeling (ASL), diffusion tensor imaging (DTI) and its derived techniques, resting-state functional MRI (rs-fMRI), and quantitative susceptibility mapping (QSM), provides a comprehensive approach for investigating the neuropathological basis of T2DM-related cognitive impairment. These imaging modalities can characterize multiple aspects of brain alterations, including macrostructural morphology, vascular injury, white matter microstructural damage and glymphatic dysfunction, abnormal brain network connectivity, and iron deposition. This article reviews the research advances in the application of MRI to T2DM-related cognitive impairment, with the aim of providing imaging evidence for clinical diagnosis and management.
Pediatric cardiac computed tomography (CCT) is an important imaging modality for the diagnosis of cardiovascular diseases in children; however, radiation safety remains a major clinical concern. With advances in artificial intelligence, deep learning reconstruction (DLR) has been applied to low-dose cardiac imaging and has shown substantial value in reducing image noise, improving spatial resolution, and lowering radiation dose and contrast media volume. This review summarizes the technical principles and advantages of DLR, as well as recent progress in phantom and clinical studies of DLR in low-dose pediatric CCT.
Tumor-infiltrating lymphocytes (TILs) are a key component of the tumor immune microenvironment in breast cancer, and their levels are closely associated with patient prognosis, response to neoadjuvant chemotherapy, and the efficacy of immunotherapy. Conventional pathological assessment of TILs is limited by subjectivity, invasiveness, and time consumption; digital pathology combined with artificial intelligence can improve the objectivity and reproducibility of TILs assessment and provide a relatively reliable pathological reference for imaging-based models. By contrast, imaging methods can more comprehensively characterize the spatiotemporal heterogeneity of TILs distribution within tumor tissue. Emerging techniques, including radiomics, habitat imaging, deep learning, and multi-task learning, can partially overcome the limitations of conventional methods and provide new perspectives and evidence for precision diagnosis and treatment of breast cancer. This review summarizes research progress in medical imaging and artificial intelligence for assessing TILs in breast cancer.
The microsatellite instability (MSI) status of gastric cancer significantly affects treatment strategies and prognosis. Imaging modalities such as CT and MRI provide novel approaches for predicting MSI status. By analyzing quantitative imaging parameters or extracting imaging features to construct predictive models, MSI status can be non-invasively predicted preoperatively, thereby providing a basis for individualized treatment selection. This article reviews the research progresses of multimodel imaging technologies, radiomics, and deep learning in predicting MSI status in gastric cancer.
Tumor angiogenesis plays a critical role in the progression of soft tissue sarcoma (STS). MRI techniques, including dynamic contrast-enhanced MRI (DCE-MRI), intravoxel incoherent motion (IVIM) imaging, diffusion kurtosis imaging (DKI), blood oxygenation level-dependent functional MRI (BOLD-fMRI), and proton magnetic resonance spectroscopy (¹H-MRS), integrate structural and functional information and can noninvasively quantify microscopic features such as tumor microvascular perfusion and permeability, cellular density, and metabolism, thereby reflecting the angiogenic status of STS. This review summarizes the research progress in the application of MRI for evaluating angiogenesis in STS.
Soft tissue tumors (STTs) comprise pathologically complex and highly heterogeneous lesions. Functional MRI (fMRI), including diffusion-weighted imaging (DWI), intravoxel incoherent motion (IVIM), and magnetic resonance spectroscopy (MRS), can quantify water diffusion, blood perfusion, and metabolic characteristics and is widely used in the diagnasis and treatment of STTs. This review summarizes the value of fMRI for investigating microscopic mechanisms such as tumor cell proliferation, angiogenesis, and metabolic reprogramming; its applications in differentiating benign from malignant STTs, preoperative staging, treatment response assessment, and prognostic prediction; and recent progress in fMRI-based radiomics and deep learning for the diagnosis and management of STTs.
Pathological remodeling of the calcified cartilage zone (CCZ) is closely associated with the initiation and progression of early osteoarthritis (OA). Ultrashort echo time (UTE) imaging overcomes the limitations of conventional MRI in capturing signals from short-T2 tissues, enabling the in vivo visualization of the CCZ. This review presents the structural characteristics of the CCZ and its pathological remodeling mechanisms during the early stages of OA. Furthermore, it summarizes recent advances in the application of UTE and its derivative sequences for morphological and quantitative evaluations of the CCZ, and discusses their clinical value in the early diagnosis, risk stratification, and treatment monitoring of OA.
Cervical ossification of the posterior longitudinal ligament (OPLL) is a chronic degenerative disease characterized by ectopic hyperplasia and ossification of the posterior longitudinal ligament. Based on imaging data from X-ray, CT, and MRI, artificial intelligence (AI) techniques have enabled the construction of various machine learning and deep learning models, substantially improving the efficiency of OPLL diagnosis and treatment. These techniques have been applied to automatic lesion detection, quantitative assessment of the ossification extent, measurement of key imaging parameters, and postoperative outcome prediction, demonstrating promising prospects for precision diagnosis and treatment of OPLL. This review summarizes the recent advances in the application of AI to automated imaging analysis and postoperative prediction in OPLL.
Objective To investigate MRI findings of the normal anatomy and injuries of the metatarsophalangeal joint (MTPJ) plantar plate. Methods This retrospective study included 28 healthy volunteers with normal MTPJ plantar plates (56 feet) and 75 patients (75 feet) with MTPJ plantar plate injuries confirmed by surgery or clinical follow-up. All volunteers and patients underwent non-contrast MRI with T1-weighted imaging (T1WI) and proton density-weighted fat-suppressed imaging (PDWI-FS) in the coronal, sagittal, and axial planes. The normal anatomy of the MTPJ plantar plate in healthy volunteers and the MRI features of MTPJ plantar plate injuries were analyzed. Results On both T1WI and PDWI-FS, the MTPJ plantar plates of the 28 healthy volunteers appeared as oval or band-like structures with homogeneous low signal intensity. Among the 75 patients with MTPJ plantar plate injuries, 60 had injuries of the first MTPJ plantar plate and 15 had injuries of the second to fifth MTPJ plantar plates (including 3 partial tears of the second MTPJ plantar plate, 7 complete tears of the second MTPJ plantar plate, 3 complete tears of the third MTPJ plantar plate, 1 complete tear of the fourth MTPJ plantar plate, and 1 complete tear of the fifth MTPJ plantar plate). Injuries of the central portion of the plantar plate, intersesamoid ligament, and metatarsosesamoid ligament manifested as discontinuity, irregular morphology, and high signal intensity on PDWI-FS. Partial tears of the sesamophalangeal ligament of the first MTPJ and of the second to fifth MTPJ plantar plates showed irregular morphology, high signal intensity, and focal fiber discontinuity. Complete tears showed full-thickness discontinuity of the ligament or plantar plate, with the tear extending through the entire ligament or plantar plate on PDWI-FS. Conclusion MRI can clearly demonstrate the imaging characteristics of MTPJ plantar plate injuries and define the location, extent, and severity of injury, thereby providing an anatomical and imaging basis for early diagnosis and accurate treatment of MTPJ plantar plate injuries.
