Multimodal Radiomics and Deep Learning Integration for Bone Health Assessment in Postmenopausal Women via Dental Radiographs: Development of an Interpretable Nomogram

ObjectiveThis study aims to create a reliable framework for grading esophageal cancer. The framework combines feature extraction, deep learning with attention mechanisms, and radiomics to ensure accuracy, interpretability, and practical use in tumor analysis.Materials and methodsThis retrospective study used data from 2,560 esophageal cancer patients across multiple clinical centers, collected from 2018 to 2023. The dataset included CT scan images and clinical information, representing a variety of cancer grades and types. Standardized CT imaging protocols were followed, and experienced radiologists manually segmented the tumor regions. Only high-quality data were used in the study. A total of 215 radiomic features were extracted using the SERA platform. The study used two deep learning models—DenseNet121 and EfficientNet-B0—enhanced with attention mechanisms to improve accuracy. A combined classification approach used both radiomic and deep learning features, and machine learning models like Random Forest, XGBoost, and CatBoost were applied. These models were validated with strict training and testing procedures to ensure effective cancer grading.ResultsThis study analyzed the reliability and performance of radiomic and deep learning features for grading esophageal cancer. Radiomic features were classified into four reliability levels based on their ICC (Intraclass Correlation) values. Most of the features had excellent (ICC > 0.90) or good (0.75 < ICC ≤ 0.90) reliability. Deep learning features extracted from DenseNet121 and EfficientNet-B0 were also categorized, and some of them showed poor reliability. The machine learning models, including XGBoost and CatBoost, were tested for their ability to grade cancer. XGBoost with Recursive Feature Elimination (RFE) gave the best results for radiomic features, with an AUC (Area Under the Curve) of 91.36%. For deep learning features, XGBoost with Principal Component Analysis (PCA) gave the best results using DenseNet121, while CatBoost with RFE performed best with EfficientNet-B0, achieving an AUC of 94.20%. Combining radiomic and deep features led to significant improvements, with XGBoost achieving the highest AUC of 96.70%, accuracy of 96.71%, and sensitivity of 95.44%. The combination of both DenseNet121 and EfficientNet-B0 models in ensemble models achieved the best overall performance, with an AUC of 95.14% and accuracy of 94.88%.ConclusionsThis study improves esophageal cancer grading by combining radiomics and deep learning. It enhances diagnostic accuracy, reproducibility, and interpretability, while also helping in personalized treatment planning through better tumor characterization.Clinical trial numberNot applicable.

This study aimed to create a reliable method for preoperative grading of meningiomas by combining radiomic features and deep learning-based features extracted using a 3D autoencoder. The goal was to utilize the strengths of both handcrafted radiomic features and deep learning features to improve accuracy and reproducibility across different MRI protocols. The study included 3,523 patients with histologically confirmed meningiomas, consisting of 1,900 low-grade (Grade I) and 1,623 high-grade (Grades II and III) cases. Radiomic features were extracted from T1-contrast-enhanced and T2-weighted MRI scans using the Standardized Environment for Radiomics Analysis (SERA). Deep learning features were obtained from the bottleneck layer of a 3D autoencoder integrated with attention mechanisms. Feature selection was performed using Principal Component Analysis (PCA) and Analysis of Variance (ANOVA). Classification was done using machine learning models like XGBoost, CatBoost, and stacking ensembles. Reproducibility was evaluated using the Intraclass Correlation Coefficient (ICC), and batch effects were harmonized with the ComBat method. Performance was assessed based on accuracy, sensitivity, and the area under the receiver operating characteristic curve (AUC). For T1-contrast-enhanced images, combining radiomic and deep learning features provided the highest AUC of 95.85% and accuracy of 95.18%, outperforming models using either feature type alone. T2-weighted images showed slightly lower performance, with the best AUC of 94.12% and accuracy of 93.14%. Deep learning features performed better than radiomic features alone, demonstrating their strength in capturing complex spatial patterns. The end-to-end 3D autoencoder with T1-contrast images achieved an AUC of 92.15%, accuracy of 91.14%, and sensitivity of 92.48%, surpassing T2-weighted imaging models. Reproducibility analysis showed high reliability (ICC > 0.75) for 127 out of 215 features, ensuring consistent performance across multi-center datasets. The proposed framework effectively integrates radiomic and deep learning features to provide a robust, non-invasive, and reproducible approach for meningioma grading. Future research should validate this framework in real-world clinical settings and explore adding clinical parameters to enhance its prognostic value.

This study aims to develop a robust and clinically applicable framework for preoperative grading of meningiomas using T1-contrast-enhanced and T2-weighted MRI images. The approach integrates radiomic feature extraction, attention-guided deep learning models, and reproducibility assessment to achieve high diagnostic accuracy, model interpretability, and clinical reliability. We analyzed MRI scans from 2546 patients with histopathologically confirmed meningiomas (1560 low-grade, 986 high-grade). High-quality T1-contrast and T2-weighted images were preprocessed through harmonization, normalization, resizing, and augmentation. Tumor segmentation was performed using ITK-SNAP, and inter-rater reliability of radiomic features was evaluated using the intraclass correlation coefficient (ICC). Radiomic features were extracted via the SERA software, while deep features were derived from pre-trained models (ResNet50 and EfficientNet-B0), with attention mechanisms enhancing focus on tumor-relevant regions. Feature fusion and dimensionality reduction were conducted using PCA and LASSO. Ensemble models employing Random Forest, XGBoost, and LightGBM were implemented to optimize classification performance using both radiomic and deep features. Reproducibility analysis showed that 52% of radiomic features demonstrated excellent reliability (ICC > 0.90). Deep features from EfficientNet-B0 outperformed ResNet50, achieving AUCs of 94.12% (T1) and 93.17% (T2). Hybrid models combining radiomic and deep features further improved performance, with XGBoost reaching AUCs of 95.19% (T2) and 96.87% (T1). Ensemble models incorporating both deep architectures achieved the highest classification performance, with AUCs of 96.12% (T2) and 96.80% (T1), demonstrating superior robustness and accuracy. This work introduces a comprehensive and clinically meaningful AI framework that significantly enhances the preoperative grading of meningiomas. The model's high accuracy, interpretability, and reproducibility support its potential to inform surgical planning, reduce reliance on invasive diagnostics, and facilitate more personalized therapeutic decision-making in routine neuro-oncology practice. Not applicable.

Purpose123I-Ioflupane SPECT is an effective tool for the diagnosis and progression assessment of Parkinson’s disease (PD). Radiomics and deep learning (DL) can be used to track and analyze the underlying image texture and features to predict the Hoehn-Yahr stages (HYS) of PD. In this study, we aim to predict HYS at year 0 and year 4 after the first diagnosis with combined imaging, radiomics and DL-based features using 123I-Ioflupane SPECT images at year 0.MethodsIn this study, 161 subjects from the Parkinson’s Progressive Marker Initiative database underwent baseline 3T MRI and 123I-Ioflupane SPECT, with HYS assessment at years 0 and 4 after first diagnosis. Conventional imaging features (IF) and radiomic features (RaF) for striatum uptakes were extracted from SPECT images using MRI- and SPECT-based (SPECT-V and SPECT-T) segmentations respectively. A 2D DenseNet was used to predict HYS of PD, and simultaneously generate deep features (DF). The random forest algorithm was applied to develop models based on DF, RaF, IF and combined features to predict HYS (stage 0, 1 and 2) at year 0 and (stage 0, 1 and ≥ 2) at year 4, respectively. Model predictive accuracy and receiver operating characteristic (ROC) analysis were assessed for various prediction models.ResultsFor the diagnostic accuracy at year 0, DL (0.696) outperformed most models, except DF + IF in SPECT-V (0.704), significantly superior based on paired t-test. For year 4, accuracy of DF + RaF model in MRI-based method is the highest (0.835), significantly better than DF + IF, IF + RaF, RaF and IF models. And DL (0.820) surpassed models in both SPECT-based methods. The area under the ROC curve (AUC) highlighted DF + RaF model (0.854) in MRI-based method at year 0 and DF + RaF model (0.869) in SPECT-T method at year 4, outperforming DL models, respectively. And then, there was no significant differences between SPECT-based and MRI-based segmentation methods except for the imaging feature models.ConclusionThe combination of radiomic and deep features enhances the prediction accuracy of PD HYS compared to only radiomics or DL. This suggests the potential for further advancements in predictive model performance for PD HYS at year 0 and year 4 after first diagnosis using 123I-Ioflupane SPECT images at year 0, thereby facilitating early diagnosis and treatment for PD patients. No significant difference was observed in radiomics results obtained between MRI- and SPECT-based striatum segmentations for radiomic and deep features.

To develop and validate a comprehensive and interpretable framework for multi-class classification of Alzheimer's disease (AD) progression stages based on hippocampal MRI, integrating radiomic, deep, and clinical features. This retrospective multi-center study included 2956 patients across four AD stages (Non-Demented, Very Mild Demented, Mild Demented, Moderate Demented). T1-weighted MRI scans were processed through a standardized pipeline involving hippocampal segmentation using four models (U-Net, nnU-Net, Swin-UNet, MedT). Radiomic features (n = 215) were extracted using the SERA platform, and deep features (n = 256) were learned using an LSTM network with attention applied to hippocampal slices. Fused features were harmonized with ComBat and filtered by ICC (≥ 0.75), followed by LASSO-based feature selection. Classification was performed using five machine learning models, including Logistic Regression (LR), Support Vector Machine (SVM), Random Forest (RF), Multilayer Perceptron (MLP), and eXtreme Gradient Boosting (XGBoost). Model interpretability was addressed using SHAP, and a nomogram and decision curve analysis (DCA) were developed. Additionally, an end-to-end 3D CNN-LSTM model and two transformer-based benchmarks (Vision Transformer, Swin Transformer) were trained for comparative evaluation. MedT achieved the best hippocampal segmentation (Dice = 92.03% external). Fused features yielded the highest classification performance with XGBoost (external accuracy = 92.8%, AUC = 94.2%). SHAP identified MMSE, hippocampal volume, and APOE ε4 as top contributors. The nomogram accurately predicted early-stage AD with clinical utility confirmed by DCA. The end-to-end model performed acceptably (AUC = 84.0%) but lagged behind the fused pipeline. Statistical tests confirmed significant performance advantages for feature fusion and MedT-based segmentation. This study demonstrates that integrating radiomics, deep learning, and clinical data from hippocampal MRI enables accurate and interpretable classification of AD stages. The proposed framework is robust, generalizable, and clinically actionable, representing a scalable solution for AD diagnostics.

Lymph-vascular space invasion prediction in cervical cancer: Exploring radiomics and deep learning multilevel features of tumor and peritumor tissue on multiparametric MRI

The preoperative prediction of the overall survival (OS) status of patients with head and neck cancer (HNC) is significant value for their individualized treatment and prognosis. This study aims to evaluate the impact of adding 3D deep learning features to radiomics models for predicting 5-year OS status. Two hundred twenty cases from The Cancer Imaging Archive public dataset were included in this study; 2212 radiomics features and 304 deep features were extracted from each case. The features were selected by univariate analysis and the least absolute shrinkage and selection operator, and then grouped into a radiomics model containing Positron Emission Tomography /Computed Tomography (PET/CT) radiomics features score, a deep model containing deep features score, and a combined model containing PET/CT radiomics features score +3D deep features score. TumorStage model was also constructed using initial patient tumor node metastasis stage to compare the performance of the combined model. A nomogram was constructed to analyze the influence of deep features on the performance of the model. The 10-fold cross-validation of the average area under the receiver operating characteristic curve and calibration curve were used to evaluate performance, and Shapley Additive exPlanations (SHAP) was developed for interpretation. The TumorStage model, radiomics model, deep model, and the combined model achieved areas under the receiver operating characteristic curve of 0.604, 0.851, 0.840, and 0.895 on the train set and 0.571, 0.849, 0.832, and 0.900 on the test set. The combined model showed better performance of predicting the 5-year OS status of HNC patients than the radiomics model and deep model. The combined model was shown to provide a favorable fit in calibration curves and be clinically useful in decision curve analysis. SHAP summary plot and SHAP The SHAP summary plot and SHAP force plot visually interpreted the influence of deep features and radiomics features on the model results. In predicting 5-year OS status in patients with HNC, 3D deep features could provide richer features for combined model, which showed outperformance compared with the radiomics model and deep model.

ComBat harmonization has been developed to remove non-biological variations for data in multi-center research applying artificial intelligence (AI). We investigated the effectiveness of ComBat harmonization on radiomic and deep features extracted from large, multi-center abdominal MRI data. A retrospective study was conducted on T2-weighted (T2W) abdominal MRI data retrieved from individual patients with suspected or known chronic liver disease at three study sites. MRI data were acquired using systems from three manufacturers and two field strengths. Radiomic features and deep features were extracted using the PyRadiomics pipeline and a Swin Transformer. ComBat was used to harmonize radiomic and deep features across different manufacturers and field strengths. Student’s t-test, ANOVA test, and Cohen’s F score were applied to assess the difference in individual features before and after ComBat harmonization. Between two field strengths, 76.7%, 52.9%, and 26.7% of radiomic features, and 89.0%, 56.5%, and 0.1% of deep features from three manufacturers were significantly different. Among the three manufacturers, 90.1% and 75.0% of radiomic features and 89.3% and 84.1% of deep features from two field strengths were significantly different. After ComBat harmonization, there were no significant differences in radiomic and deep features among manufacturers or field strengths based on t-tests or ANOVA tests. Reduced Cohen’s F scores were consistently observed after ComBat harmonization. ComBat harmonization effectively harmonizes radiomic and deep features by removing the non-biological variations due to system manufacturers and/or field strengths in large multi-center clinical abdominal MRI datasets.

Objective: To develop a preoperative differentiation model for colorectal mucinous adenocarcinoma and non-mucinous adenocarcinoma using a combination of contrast-enhanced CT radiomics and deep learning methods. Methods: This is a retrospective cohort study. Clinical data of colorectal cancer patients confirmed by postoperative pathological examination were retrospectively collected from January 2016 to December 2023 at Shanghai General Hospital Affiliated to Shanghai Jiao Tong University School of Medicine (Center 1, n=220) and the First Affiliated Hospital of Bengbu Medical University (Center 2, n=51). Among them, there were 108 patients diagnosed with mucinous adenocarcinoma, including 55 males and 53 females, with an age of (68.4±12.2) years (range: 38 to 96 years); and 163 patients diagnosed with non-mucinous adenocarcinoma, including 96 males and 67 females, with an age of (67.9±11.0) years (range: 43 to 94 years). The cases from Center 1 were divided into a training set (n=156) and an internal validation set (n=64) using stratified random sampling in a 7∶3 ratio, and the cases from Center 2 were used as an independent external validation set (n=51). Three-dimensional tumor volume of interest was manually segmented on venous-phase contrast-enhanced CT images. Radiomics features were extracted using PyRadiomics, and deep learning features were extracted using the ResNet-18 network. The two sets of features were then combined to form a joint feature set. The consistency of manual segmentation was assessed using the intraclass correlation coefficient. Feature dimensionality reduction was performed using the Mann-Whitney U test and the least absolute shrinkage and selection operator regression. Six machine learning algorithms were used to construct models based on radiomics features, deep learning features, and combined features, including support vector machine, logistic regression, random forest, extreme gradient boosting, k-nearest neighbors, and decision tree. The discriminative performance of each model was evaluated using receiver operating characteristic curves, the area under the curve (AUC), DeLong test, and decision curve analysis. Results: After feature selection, 22 features with the most discriminative value were finally retained, among which 12 were traditional radiomics features and 10 were deep learning features. In the internal validation set, the Random Forest algorithm based on the combined features model achieved the best performance (AUC=0.938, 95%CI: 0.875 to 0.984), which was superior to the single-modality radiomics feature model (AUC=0.817, 95%CI: 0.702 to 0.913,P=0.048) and the deep learning feature model (AUC=0.832, 95%CI: 0.727 to 0.926,P=0.087); in the independent external validation set, the Random Forest algorithm with the combined features model maintained the highest discriminative performance (AUC=0.891, 95%CI: 0.791 to 0.969), which was superior to the single-modality radiomics feature model (AUC=0.770, 95%CI: 0.636 to 0.890,P=0.045) and the deep learning feature model (AUC=0.799, 95%CI: 0.652 to 0.911,P=0.169). Conclusion: The combined model based on radiomics and deep learning features from venous-phase enhanced CT demonstrates good performance in the preoperative differentiation of colorectal mucinous from non-mucinous adenocarcinoma.

BACKGROUND Timely identification of local failure after Stereotactic Radiosurgery (SRS) offers the opportunity for appropriate treatment modifications that may result in improved treatment outcomes, patient survival, and quality of life. Previous studies showed that the addition of either radiomics or deep learning features to clinical features increased the accuracy of the models in predicting local control (LC) of brain metastases after SRS. To date, however, no study combined both radiomics and deep learning features together with clinical features to develop machine learning algorithms to predict LC of brain metastases. In this study, we examined whether a model trained with a combination of all these features could predict LC better than models trained with only a subset of these features. MATERIAL AND METHODS Pre-treatment brain MRIs and clinical data were collected retrospectively for 129 patients at the Gamma Knife Center of Elisabeth-TweeSteden Hospital (ETZ), Tilburg, The Netherlands. The patients were split into 103 patients for training and 26 patients for testing. The segment-based radiomics features were extracted using the radiomics feature extractor of the python radiomics package. The deep learning features were extracted using a fine-tuned 3D ResNet model and then combined with the clinical and radiomics features. A Random Forest classifier was trained with the training data set and then tested with the test data set. The performance was compared across 4 different models trained with clinical features only, clinical and radiomics features, clinical and deep learning features, and clinical, radiomics and deep learning features. RESULTS The prediction model with only clinical variables demonstrated an area under the receiver operating characteristic curve (AUC) of 0.82 and an accuracy of 75.6%. The prediction model with the combination of clinical and radiomics features demonstrated an AUC of 0.880 and an accuracy of 83.3% whereas the prediction model with the combination of clinical and deep learning features demonstrated an AUC of 0.863 and an accuracy of 78.3%. The best prediction performance was associated with the model that combined the clinical, radiomics and deep learning features with an AUC of 0.886 and 87% accuracy. CONCLUSION Machine learning models built on radiomics features and deep learning features combined with patient characteristics show potential to predict LC after SRS with high accuracy. The promising findings from this study demonstrate the potential for early prediction of SRS outcome for brain metastasis prior to treatment initiation and might offer the opportunity for appropriate treatment modifications that may result in improved treatment outcomes, patient survival, and quality of life.

ObjectiveThe aim of this study was to develop a hybrid diagnostic framework integrating radiomic and deep features for accurate and reproducible detection and classification of wrist fractures using X-ray images.Materials and MethodsA total of 3,537 X-ray images, including 1,871 fracture and 1,666 non-fracture cases, were collected from three healthcare centers. Radiomic features were extracted using the PyRadiomics library, and deep features were derived from the bottleneck layer of an autoencoder. Both feature modalities underwent reliability assessment via Intraclass Correlation Coefficient (ICC) and cosine similarity. Feature selection methods, including ANOVA, Mutual Information (MI), Principal Component Analysis (PCA), and Recursive Feature Elimination (RFE), were applied to optimize the feature set. Classifiers such as XGBoost, CatBoost, Random Forest, and a Voting Classifier were used to evaluate diagnostic performance. The dataset was divided into training (70%) and testing (30%) sets, and metrics such as accuracy, sensitivity, and AUC-ROC were used for evaluation.ResultsThe combined radiomic and deep feature approach consistently outperformed standalone methods. The Voting Classifier paired with MI achieved the highest performance, with a test accuracy of 95%, sensitivity of 94%, and AUC-ROC of 96%. The end-to-end model achieved competitive results with an accuracy of 93% and AUC-ROC of 94%. SHAP analysis and t-SNE visualizations confirmed the interpretability and robustness of the selected features.ConclusionsThis hybrid framework demonstrates the potential for integrating radiomic and deep features to enhance diagnostic performance for wrist and forearm fractures, providing a reliable and interpretable solution suitable for clinical applications.

This study aimed to develop a hybrid model combining radiomics and deep learning features derived from computed tomography (CT) images to classify histological subtypes of non-small cell lung cancer (NSCLC). We analyzed CT images and radiomics features from 235 patients with NSCLC, including 110 with adenocarcinoma (ADC) and 112 with squamous cell carcinoma (SCC). The dataset was split into a training set (75%) and a test set (25%). External validation was conducted using the NSCLC-Radiomics database, comprising 24 patients each with ADC and SCC. A total of 1409 radiomics and 8192 deep features underwent principal component analysis (PCA) and ℓ2,1-norm minimization for feature reduction and selection. The optimal feature sets for classification included 27 radiomics features, 20 deep features, and 55 combined features (30 deep and 25 radiomics). The average area under the receiver operating characteristic curve (AUC) for radiomics, deep, and combined features were 0.6568, 0.6689, and 0.7209, respectively, across the internal and external test sets. Corresponding average accuracies were 0.6013, 0.6376, and 0.6564. The combined model demonstrated superior performance in classifying NSCLC subtypes, achieving higher AUC and accuracy in both test datasets. These results suggest that the proposed hybrid approach could enhance the accuracy and reliability of NSCLC subtype classification.

Gliomas, the most common primary brain tumors, are classified into low-grade glioma (LGG) and high-grade glioma (HGG) based on aggressiveness. Accurate preoperative differentiation is vital for effective treatment and prognosis, but traditional methods like biopsy have limitations, such as sampling errors and procedural risks. This study introduces a comprehensive model that combines radiomics features (RFs) and deep features (DFs) from magnetic resonance imaging (MRI) scans, integrating clinical factors with advanced imaging features to enhance diagnostic precision for preoperative glioma grading. In this retrospective multi-center study [2017-2022], 582 patients underwent preoperative contrast-enhanced T1-weighted (CE-T1w) and T2-weighted fluid-attenuated inversion recovery (T2w FLAIR) MRI. The dataset, divided into 407 training and 175 testing cases, included 340 LGGs and 242 HGGs. RFs and DFs were extracted from CE-T1w images, and radiomic scores (rad-score) and deep scores (deep-score) were calculated. Additionally, a clinical model based on demographics and MRI findings (CE-T1w and T2w FLAIR imaging) was developed. A nomogram model integrating rad-score, deep-score, and clinical factors was constructed using multivariate logistic regression analysis. Decision curve analysis (DCA) was employed to evaluate the nomogram's clinical utility in distinguishing between HGGs and LGGs. The study included 582 patients (mean age: 52±14 years; 57.91% male). No significant differences in age or sex were found between the training and testing groups (P>0.05). For RFs, 73.02% of the 215 extracted features were selected based on inter-class correlation coefficients (ICCs), while for DFs, 38.27% of the 15,680 extracted features were selected. Optimal penalization coefficients lambda (λ) for RFs and DFs were determined using a five-fold cross-validation and minimal criteria process. The resulting receiver operating characteristic-area under the curve (ROC-AUC) values were 0.93 [95% confidence interval (CI): 0.91-0.94] for the training set and 0.91 (95% CI: 0.89-0.93) for the testing set. The Hosmer-Lemeshow test yielded P values of 0.619 and 0.547 for the training and testing sets, respectively, indicating satisfactory calibration. The nomogram demonstrated the highest net benefit (NB) up to a threshold of 0.7, followed by DFs and RFs. This study underscores the efficacy of integrating RFs and DFs alongside clinical data to accurately predict the pathological grading of HGGs and LGGs, offering a comprehensive approach for clinical decision-making.

The differentiation of autoimmune pancreatitis (AIP) and pancreatic ductal adenocarcinoma (PDAC) poses a relevant diagnostic challenge and can lead to misdiagnosis and consequently poor patient outcome. Recent studies have shown that radiomics-based models can achieve high sensitivity and specificity in predicting both entities. However, radiomic features can only capture low level representations of the input image. In contrast, convolutional neural networks (CNNs) can learn and extract more complex representations which have been used for image classification to great success. In our retrospective observational study, we performed a deep learning-based feature extraction using CT-scans of both entities and compared the predictive value against traditional radiomic features. In total, 86 patients, 44 with AIP and 42 with PDACs, were analyzed. Whole pancreas segmentation was automatically performed on CT-scans during the portal venous phase. The segmentation masks were manually checked and corrected if necessary. In total, 1411 radiomic features were extracted using PyRadiomics and 256 features (deep features) were extracted using an intermediate layer of a convolutional neural network (CNN). After feature selection and normalization, an extremely randomized trees algorithm was trained and tested using a two-fold shuffle-split cross-validation with a test sample of 20% (n = 18) to discriminate between AIP or PDAC. Feature maps were plotted and visual difference was noted. The machine learning (ML) model achieved a sensitivity, specificity, and ROC-AUC of 0.89 ± 0.11, 0.83 ± 0.06, and 0.90 ± 0.02 for the deep features and 0.72 ± 0.11, 0.78 ± 0.06, and 0.80 ± 0.01 for the radiomic features. Visualization of feature maps indicated different activation patterns for AIP and PDAC. We successfully trained a machine learning model using deep feature extraction from CT-images to differentiate between AIP and PDAC. In comparison to traditional radiomic features, deep features achieved a higher sensitivity, specificity, and ROC-AUC. Visualization of deep features could further improve the diagnostic accuracy of non-invasive differentiation of AIP and PDAC.

BackgroundCurrent preoperative imaging methods, such as ultrasound, are limited by operator dependency and suboptimal sensitivity for detecting central lymph node metastasis (CLNM). This study aimed to propose a method that integrates deep learning and radiomics to accurately predict lymph node metastasis in thyroid cancer by analyzing intra- and peri-tumoral imaging features, thereby improving the preoperative prediction accuracy.MethodsFrom July 2020 to June 2022, 405 patients diagnosed with PTC were enrolled from two centers: Center 1 (Shanghai Sixth People’s Hospital) with 294 patients divided into a training set (n=294) and an internal validation set, and Center 2 (Tongji Hospital Affiliated to Tongji University) with 111 patients as the external test set. Postoperative pathological confirmation served as the reference standard for CLNM diagnosis. A total of 1,561 radiomics features and 2,048 deep learning features were extracted from intra- and peri-tumoral regions of each ultrasound image. Feature selection was performed using analysis of variance (ANOVA) and least absolute shrinkage and selection operator (LASSO), resulting in the selection of relevant features for constructing support vector machine (SVM) models. Additionally, radiomics-deep learning fusion models were developed by combining selected radiomics and deep learning features.ResultsAmong 405 patients (mean age: 46.59±12.74 years; 68.6% female), 171 exhibited CLNM, highlighting the clinical urgency for accurate prediction. Among the 405 patients, 171 exhibited CLNM. The radiomics models demonstrated area under the curve (AUC) values of 0.760 in internal validation and 0.748 in the external test cohort. The deep learning models demonstrated improved performance with AUCs of 0.794 and 0.756 in the internal and external test sets. Notably, the highest AUC values of 0.897 (internal validation) and 0.881 (external test set) were obtained by the radiomics-deep learning fusion SVM model incorporating both intra- and peri-tumoral regions. DeLong’s test confirmed statistically significant improvements (P<0.05) of the fusion model over the intra-tumoral radiomics model (P=0.008), intra-tumoral deep learning model (P=0.005), and combined intra-tumoral radiomics-deep learning model (P=0.01). However, no significant differences were observed compared to the combined intra- and peri-tumoral deep learning model (P=0.17). Decision curve analysis indicated that the fusion model offers greater clinical utility in predicting CLNM.ConclusionsThe integration of radiomics and deep learning features significantly enhances the diagnostic performance for predicting CLNM in papillary thyroid carcinoma (PTC). The radiomics-deep learning fusion SVM model outperforms individual radiomics and deep learning models, demonstrating substantial potential for clinical application in improving surgical decision-making and patient management. The fusion model could reduce unnecessary central lymph node dissections (CLNDs) and improve surgical planning by providing personalized risk stratification.