An interactive cascaded deep learning framework with expert refinement for accurate striatal subregion segmentation.

Silent corticotroph adenomas (SCAs) are a subtype of nonfunctioning pituitary adenomas that exhibit more aggressive behavior. However, rapid and accurate preoperative diagnostic methods are currently lacking. The purpose of this study was to examine the differences between SCA and non-SCA features and to establish radiomics models and a clinical scale for rapid and accurate prediction. A total of 260 patients (72 SCAs vs. 188 NSCAs) with nonfunctioning adenomas from Peking Union Medical College Hospital were enrolled in the study as the internal dataset. Thirty-five patients (6 SCAs vs. 29 NSCAs) from Fuzhou General Hospital were enrolled as the external dataset. Radiomics models and an SCA scale to preoperatively diagnose SCAs were established based on MR images and clinical features. There were more female patients (internal dataset: p < 0.001; external dataset: p = 0.028) and more multiple microcystic changes (internal dataset: p < 0.001; external dataset: p = 0.012) in the SCA group. MRI showed more invasiveness (higher Knosp grades, p ≤ 0.001). The radiomics model achieved AUCs of 0.931 and 0.937 in the internal and external datasets, respectively. The clinical scale achieved an AUC of 0.877 and a sensitivity of 0.952 in the internal dataset and an AUC of 0.899 and a sensitivity of 1.0 in the external dataset. Based on clinical information and imaging characteristics, the constructed radiomics model achieved high preoperative diagnostic ability. The SCA scale achieved the purpose of rapidity and practicality while ensuring sensitivity, which is conducive to simplifying clinical work.

Apart from invasive pathological examination, there is no effective method to differentiate breast diffuse large B-cell lymphoma (DLBCL) from breast invasive ductal carcinoma (IDC). In this study, we aimed to develop and validate an effective deep learning radiomics model to discriminate between DLBCL and IDC. A total of 324 breast nodules from 236 patients with baseline 18F-fluorodeoxyglucose (18F-FDG) positron emission tomography/computed tomography (PET/CT) were retrospectively analyzed. After grouping breast DLBCL and breast IDC patients, external and internal datasets were divided according to the data collected by different centers. Preprocessing was then used to process the original PET/CT images and an attention-based aggregate convolutional neural network (AACNN) model was designed. The AACNN model was trained using patches of CT or PET tumor images and optimized with an improved loss function. The final ensemble predictive model was built using distance weight voting. Finally, the model performance was evaluated and statistically verified. A total of 249 breast nodules from Fudan University Shanghai Cancer Center (FUSCC) and 75 breast nodules from Shanghai Proton and Heavy Ion Center (SPHIC) were selected as internal and external datasets, respectively. On the internal testing, our method yielded an area under the curve (AUC), accuracy (ACC), sensitivity (SEN), specificity (SPE), positive predictive value (PPV), negative predictive value (NPV), and harmonic mean of precision and sensitivity (F1) of 0.886, 83.0%, 80.9%, 85.0%, 84.8%, 81.2%, and 0.828, respectively. Meanwhile on the external testing, the results were 0.788, 71.6%, 61.4%, 84.7%, 84.0%, 62.6%, and 0.709, respectively. Our study outlines a deep learning radiomics method which can automatically, noninvasively, and accurately differentiate breast DLBCL from breast IDC, which will be more in line with the needs and strategies of precision medicine, individualized diagnosis, and treatment.

Small intracranial aneurysms are being increasingly detected while the rupture risk is not well-understood. We aimed to develop rupture-risk models of small aneurysms by combining clinical, morphologic, and hemodynamic information based on machine learning techniques and to test the models in external validation datasets. From January 2010 to December 2016, five hundred four consecutive patients with only small aneurysms (<5 mm) detected by CTA and invasive cerebral angiography (or surgery) were retrospectively enrolled and randomly split into training (81%) and internal validation (19%) sets to derive and validate the proposed machine learning models (support vector machine, random forest, logistic regression, and multilayer perceptron). Hemodynamic parameters were obtained using computational fluid dynamics simulation. External validation was performed in other hospitals to test the models. The support vector machine performed the best with areas under the curve of 0.88 (95% CI, 0.85-0.92) and 0.91 (95% CI, 0.74-0.98) in the training and internal validation datasets, respectively. Feature ranks suggested hemodynamic parameters, including stable flow pattern, concentrated inflow streams, and a small (<50%) flow-impingement zone, and the oscillatory shear index coefficient of variation, were the best predictors of aneurysm rupture. The support vector machine showed an area under the curve of 0.82 (95% CI, 0.69-0.94) in the external validation dataset, and no significant difference was found for the areas under the curve between internal and external validation datasets (P = .21). This study revealed that machine learning had a good performance in predicting the rupture status of small aneurysms in both internal and external datasets. Aneurysm hemodynamic parameters were regarded as the most important predictors.

Artificial intelligence recognition of pathological T stage and tumor invasion in rectal cancer based on large panoramic pathological sections

PurposeBody composition analysis on abdominal CT scans is useful for opportunistic screening. It also offers prognostic insights into mortality and cardiovascular risk. However, current segmentation methods for muscle and fat often fail on quantitative CT scans used for bone densitometry. These scans are commonly used to diagnose and monitor osteoporosis. This study aims to develop an accurate segmentation method for such scans and compare its performance with existing methods.MethodsWe applied an nnU-Net framework to segment muscle, subcutaneous fat, visceral fat, and an added ‘body’ class for other non-background voxels. Training data included CT scans with bone densitometry phantoms, with segmentation annotations generated using our previous segmentation method followed by manual refinement. The proposed method was evaluated on 980 CT scans across two internal and external datasets, including 30 CT scans with phantoms in internal and external datasets (15 scans in each). Comparison was made with TotalSegmentator and our previous approach.ResultsThe proposed method achieved the highest accuracy for muscle and subcutaneous fat segmentation across all four datasets (p<0.05) and delivered comparable accuracy for visceral fat. In comparison with TotalSegmentator and the previous method, there were no false segmentations in the densitometry phantom included within the display field-of-view of the patient scan.ConclusionExperimental results showed that the proposed method improved segmentation accuracy for muscle and subcutaneous fat while maintaining high accuracy for visceral fat. Notably, segmentation accuracy was also high in the quantitative CT scans for bone densitometry. These findings highlight the potential of the method to advance body composition analysis in clinical practice.

Quantitative analysis is an important factor in radiological routine. Recently the TotalSegmentator was released, a free-to-use segmentation tool with over 104 structures included. Our aim was to add missing and enhance previously included cardiovascular (CV) structures to potentially help find new insights into diseases such as aortic aneurysms in future studies. The TotalSegmentator data set with 1613 CT scans (mean age 63.6±15.9 (SD); 675 female), was used. CT scans were selected from clinical routine including various protocols and pathologies. The data set was split in training (1472), validation (57) and testing (84). Segmentations were performed in dedicated imaging software using an iterative approach for training to reduce segmentation workload. Eleven structures were added, and segmentations of six structures were enhanced. The Dice similarity score (DICE) and the Normalized surface distance (NSD) were calculated on an internal and external data set. The external validation was performed on the Dongyang data set. The Mann Whitney U test was performed to evaluate the performance increase on the previously included structures. Median DICE [IQR] and NSD [IQR] were 0.967 [0.020] and 1.000 [0.000], respectively. DICE (p<0.001) and NSD (p<0.001) significantly increased for 5/6 structures. On evaluation using the external data set, DICE and NSD were 0.970 [0.020] and 1.000 [0.000]. Accurate segmentations and enhanced segmentations of previously included CV structures were successfully implemented. This suggests further usage in research studies while still running on conventional computers with or without a dedicated graphics processing unit.

Deep neural networks have demonstrated promising performance in screening mammography with recent studies reporting performance at or above the level of trained radiologists on internal datasets. However, it remains unclear whether the performance of these trained models is robust and replicates across external datasets. In this study, we evaluate four state-of-the-art publicly available models using four publicly available mammography datasets (CBIS-DDSM, INbreast, CMMD, OMI-DB). Where test data was available, published results were replicated. The best-performing model, which achieved an area under the ROC curve (AUC) of 0.88 on internal data from NYU, achieved here an AUC of 0.9 on the external CMMD dataset (N = 826 exams). On the larger OMI-DB dataset (N = 11,440 exams), it achieved an AUC of 0.84 but did not match the performance of individual radiologists (at a specificity of 0.92, the sensitivity was 0.97 for the radiologist and 0.53 for the network for a 1-year follow-up). The network showed higher performance for in situ cancers, as opposed to invasive cancers. Among invasive cancers, it was relatively weaker at identifying asymmetries and was relatively stronger at identifying masses. The three other trained models that we evaluated all performed poorly on external datasets. Independent validation of trained models is an essential step to ensure safe and reliable use. Future progress in AI for mammography may depend on a concerted effort to make larger datasets publicly available that span multiple clinical sites.

Workers on offshore petroleum installations might be exposed to benzene, a carcinogenic agent. Recently, a full-shift benzene exposure model was developed based on personal measurements. This study aimed to validate this exposure model by using datasets not included in the model. The exposure model was validated against an internal dataset of measurements from offshore installations owned by the same company that provided data for the model, and an external dataset from installations owned by another company. We used Tobit regression to estimate GM (geometric mean) benzene exposure overall and for individual job groups. Bias, relative bias, precision, and correlation were estimated to evaluate the agreement between measured exposures and the levels predicted by the model. Overall, the model overestimated exposure when compared to the predicted exposure level to the internal dataset with a factor of 1.7, a relative bias of 73%, a precision of 0.6, a correlation coefficient of 0.72 (p = 0.019), while the Lin’s Concordance Correlation Coefficient (CCC) was 0.53. The model underestimated exposure when compared to the external dataset with a factor of about 2, with a relative bias of −45%, a precision of 1.2, a correlation coefficient of 0.31 (p = 0.544), and a Lin’s CCC of 0.25. The exposure model overestimated benzene exposure in the internal validation dataset, while the precision and the correlation between the measured and predicted exposure levels were high. Differences in measurement strategies could be one of the reasons for the discrepancy. The exposure model agreed less with the external dataset.

BackgroundPrecise volume quantification of intracerebral hemorrhage (ICH), intraventricular hemorrhage (IVH), and perihematomal edema (PHE) is a critical parameter for guiding therapy decisions, monitoring therapeutic effects over time, and predicting patient outcomes.PurposeTo evaluate a nnU-Net-based deep learning model for automated, multilesion segmentation on non-contrast CT.Materials and MethodsRetrospective data from acute spontaneous ICH patients admitted to 4 stroke centers (2015-2022) and controls (2022-2023) were analyzed. Manual segmentations served as ground truth with repeated segmentations as reference standard. nnU-Net was trained (n = 775) using 5-fold cross-validation and tested on a holdout set (n = 189). Lesion detection, segmentation, and volumetric accuracy were evaluated using the Dice similarity coefficient (DSC) and Pearson correlation coefficients (r), with subanalyses for anatomical location and impact of other hemorrhage types (subarachnoid, subdural, or epidural hematoma). The model was validated on internal (n = 121) and external (n = 169) datasets. Processing time was compared to manual segmentation.ResultsTest set sensitivity was 99% for ICH and PHE and 97% for IVH. Segmentation achieved a DSC of 0.91 (ICH), 0.71 (PHE), and 0.76 (IVH), with r = 0.99 (ICH, IVH) and r = 0.92 (PHE). DSC for lobar and deep hemorrhages were 0.90 and 0.92, respectively, and 0.70 in the brainstem, with other hemorrhage types showing no significant impact on segmentation accuracy, P > .05. For internal validation, DSC was 0.88 (ICH), 0.66 (PHE), and 0.80 (IVH), with r of 0.98, 0.88, and 0.98, respectively. External validation yielded DSC values of 0.85 (ICH), 0.61 (PHE), and 0.80 (IVH), with r values of 0.97, 0.85, and 0.96. Mean processing time was 18.2 s (±5 SD), compared to 18.01 min (±20.47 SD) for manual segmentations.ConclusionnnU-Net enables reliable, time-efficient segmentation of ICH, PHE, and IVH, validated across multicenter, multivendor datasets of spontaneous ICH, showing potential to enhance clinical workflows.

Assessment of thyroid nodules is usually relied on the experience of the radiologist and is time-consuming. Classification model of thyroid nodules cannot only reduce the burden on physicians but also provide objective recommendations. However, most classification models based on deep learning simply give a prediction result of the benignity or malignancy of nodules; thus, physicians have no way of knowing how the deep learning gets the prediction result due to the black-box nature of neural networks. In this work, we integrate the explainability directly into the outputs generated by the model through combining thyroid imaging reporting and data system (TI-RADS). The inference process of the proposed method is consistent with doctor's clinical diagnosis process; therefore, doctors can better explain the diagnosis results of the model to thepatient. A multitask network based on TI-RADS (MTN-TI-RADS) for the classification of thyroid nodules is proposed. In this network, a set of TI-RADS classifications of nodules is first obtained by multitask learning, then the TI-RADS points and the corresponding risk levels are calculated, and finally, nodules are classified as benign and malignant. The classification process through the network is consistent with the diagnostic process of physician; thus, the results of classification can be easily understood by physicians. In addition, the attention modules are introduced to the spatial and channel domains to let the network focus more on criticalfeatures. To verify the classification performance of our method, we compared the results obtained through our method with the results of the radiologist's evaluation. For the 781 test nodules in the internal dataset and the 886 test nodules in the external dataset, the sensitivity and specificity of MTN-TI-RADS were 0.988, 0.912 in internal dataset, 0.949, 0.930 in external dataset, versus the senior radiologist of 0.925 ( ), 0.816 ( ), and 0.910 ( ), 0.836 ( ), respectively. And the area under the receiver operating characteristic curve of MTN-TI-RADS was 0.981 in internal dataset, 0.973 in external dataset, versus the senior radiologist of 0.905, 0.923. For the internal dataset, we also computed the accuracy of the risk level (TR1 to TR5) and the mean absolute error (MAE). The accuracy of the risk level of the proposed method is 78%, and the MAE is 1.30. The MAE of the total points (0-14 points) is 1.30. An effective and result-interpretable end-to-end thyroid nodule classification network (MTN-TI-RADS) is proposed. MTN-TI-RADS has superior ability to classify malignant and benign thyroid nodules compared to senior radiologists. Based on MTN-TI-RADS, a classification model with strong interpretation and a high degree of physician trust is constructed. The proposed classification network is consistent with the diagnosis process of physicians, thus is more reliable and interpretable, and has great potential for clinicalapplication.

Purpose To develop and validate a deep learning model for automatic segmentation of primary central nervous system lymphoma (PCNSL) at postcontrast T1-weighted MRI. Materials and Methods Data were retrospectively collected from patients with pathologically proven immunocompetent PCNSL between September 2010 and February 2022. Postcontrast T1-weighted MRI scans were used to train and validate a deep learning model based on the nnU-Net framework. Manual segmentation by neuroradiologists served as the reference standard. The model was trained using an internal dataset from a single center and tested on both internal and external test sets from seven additional centers. Performance was assessed using Dice score, mean average surface distance, and F1 score. Statistical comparisons were performed using Mann-Whitney U test and bootstrap resampling for CIs. Results The study included 135 patients (68 female, 66 male, and one of unspecified sex; internal dataset: mean age ±SD, 67.0 years ± 12.0; external dataset: mean age, 75.5 years ± 13.6). The model achieved a mean Dice score of 0.84 (95% CI: 0.79, 0.88) on the internal test set (n = 44) and 0.88 (95% CI: 0.84, 0.91) on the external test set (n = 48), with no evidence of a difference between test sets (P = .59). Performance varied by lesion type; accuracy was highest in homogeneous discrete lesions, and performance was slightly decreased when numerous poorly defined infracentimetric lesions occurred. Strong volumetric correlation was observed between automatic and manual segmentations (internal: r = 0.99, P < .001; external: r = 0.98, P < .001). Conclusion A deep learning model achieved accurate and robust automatic segmentation of PCNSL across multiple clinical centers with different MRI acquisition parameters. Keywords: Brain Lymphoma, Brain Tumor, Automatic Segmentation, Artificial Intelligence, Deep Learning Supplemental material is available for this article. © RSNA, 2025.

BackgroundArtificial intelligence (AI) has shown promise for automating spinal alignment assessment in adolescent idiopathic scoliosis (AIS). However, AI models typically exhibit reduced accuracy and robustness when deployed across multiple medical centers due to variability in imaging protocols and data characteristics, potentially compromising clinical diagnosis and treatment decisions.ObjectiveThis study aimed to develop a real-time, plug-and-play data transformation method to enhance the robustness of deep learning models against data heterogeneity in radiographs, thereby improving their performance in assessing AIS across multiple medical centers.MethodsIn this retrospective multicenter study, 3899 full-spine radiographs from 7 hospitals (2 from Hong Kong and 5 from Mainland China), collected between January 2012 and August 2024, were included. Data from 2 hospitals in Hong Kong (n=3034) were used for model training and internal validation, while radiographs from the 5 mainland hospitals (n=865) formed 5 independent external validation datasets. A novel pixel intensity–based data transformation method was developed to standardize image contrast and brightness across datasets and integrated into the model training process to enhance our previously developed AI model, SpineHRNet+. The enhanced model's accuracy and robustness for cobb angle (CA) prediction and severity classification were evaluated using both internal and external datasets. Data heterogeneity across centers was quantified by brightness and contrast differences. CA prediction accuracy was evaluated using residual analysis, linear regression (coefficient of determination [R²]), and Bland-Altman analyses. Model performance for disease severity classification was assessed using sensitivity, specificity, precision, negative predictive value, accuracy, and confusion matrix analysis. The transformation method aligns pixel intensity distributions across datasets using statistical profiling and optimization, ensuring consistent image characteristics while preserving anatomical integrity.ResultsThe developed data transformation method significantly reduced contrast variability between datasets, improving consistency in image characteristics and enabling more reliable AI analysis. The enhanced SpineHRNet+ achieved consistent and accurate CA predictions across external validation datasets, with mean prediction errors within 4° (SD 3.12°), and maintained an R² greater than 0.90 for all centers. The sensitivity and negative predictive value for disease severity grading improved to 90.18% and 93.16%, respectively. Bland-Altman analyses demonstrated robust agreement, with 95% limits of agreement within 7.51° across all datasets.ConclusionsThe proposed data transformation approach effectively addressed data heterogeneity, significantly improving the accuracy and robustness of SpineHRNet+ in multicenter AIS assessments. The real-time processing capability and preservation of anatomical integrity underscore the method’s clinical practicality, enabling scalable and reliable AI applications in diverse health care environments.

To optimize surgical procedures and prevent retained surgical instruments, precise identification of required instruments during surgical treatment is essential. However, establishing ground truth data can be a labor-intensive barrier for researchers. Therefore, we developed and evaluated a novel system for detecting laparoscopic surgical instruments during laparoscopic cholecystectomy through virtual image creation. Virtual images were created by synthesizing laparoscopic instrument photos with surgical video backgrounds. The 311 instrument images and 1610 background images from 52 patients were augmented through random brightness, contrast, crop, rotation, scaling, flipping, and perspective transformations, resulting in 6023 composite images. These data were split into training, tuning, and internal test sets. Based on synthetic data, we developed a system comprising two-step processes. The first model is a unified instrument localization model that detects surgical instruments, and the second model is an instrument-type classification model that categorizes the detected surgical instruments. External and public datasets were used to evaluate generalizability. The unified instrument localization model achieved average precision (AP) values with intersection over union (IoU) of 0.5 of 0.981, 0.882, and 0.689 for internal, external, and public datasets, respectively. The instrument-type classification model demonstrated area under the curve (AUC) values of 0.959 for seven instrument types in the external dataset and 0.749 for four instrument types in the public dataset. The final two-step instrument detection model demonstrated an AUC of 0.848 for the external dataset and 0.688 for the public dataset, which showed significantly superior performance compared to conventional multi-class instrument models. This validated deep learning model using synthetically generated data provides a reliable framework for surgical instrument detection. Our approach demonstrates strong performance and generalizability, suggesting its potential for improving operative workflow efficiency and surgical education across various minimally invasive procedures.

3D multi-scale, multi-task, and multi-label deep learning for prediction of lymph node metastasis in T1 lung adenocarcinoma patients' CT images

IntroductionUse of stereotactic ablative radiotherapy (SABR) for central lung tumors can result in up to a 35% incidence of late pulmonary toxicity. We evaluated an automated scoring method to quantify post-SABR bronchial changes by using artificial intelligence (AI)-based airway segmentation. Materials and methodsCentral lung SABR patients treated at Amsterdam UMC (AUMC, internal reference dataset) and Peter MacCallum Cancer Centre (PMCC, external validation dataset) were identified. Patients were eligible if they had pre- and post-SABR CT scans with ≤ 1 mm resolution. The first step of the automated scoring method involved AI-based airway auto-segmentation using MEDPSeg, an end-to-end deep learning-based model. The Vascular Modeling Toolkit in 3D Slicer was then used to extract a centerline curve through the auto-segmented airway lumen, and cross-sectional measurements were computed along each bronchus for all CT scans. For AUMC patients, airway stenosis/occlusion was evaluated by both visual assessment and automated scoring. Only the automated method was applied to the PMCC dataset. ResultsStudy patients comprised 26 from AUMC, and 33 from PMCC. Visual scoring identified stenosis/occlusion in 8 AUMC patients (31 %), most frequently in the segmental bronchi. After airway auto-segmentation, minor manual edits were needed in 9 % of patients. Segmentation for a single scan averaged 83sec (range 73–136). Automated scoring nearly doubled detected airway stenosis/occlusion (n = 15, 58 %), and allowed for earlier detection in 5/8 patients who had also visually scored changes. Estimated rates were 48 % and 66 % at 1- and 2-years, respectively, for the internal dataset. The automated detection rate was 52 % in the external dataset, with 1- and 2-year risks of 56 % and 61 %, respectively. ConclusionAn AI-based automated scoring method allows for detection of more bronchial stenosis/occlusion after lung SABR, and at an earlier time-point. This tool can facilitate studies to determine early airway changes and establish more reliable airway tolerance doses.