Attention-based deep learning for breast lesions classification on contrast enhanced spectral mammography: a multicentre study.

MRI-based radiomics has been used to diagnose breast lesions; however, little research combining quantitative pharmacokinetic parameters of dynamic contrast-enhanced MRI (DCE-MRI) and diffusion kurtosis imaging (DKI) exists. To develop and validate a multimodal MRI-based radiomics model for the differential diagnosis of benign and malignant breast lesions and analyze the discriminative abilities of different MR sequences. Retrospective. In all, 207 female patients with 207 histopathology-confirmed breast lesions (95 benign and 112 malignant) were included in the study. Then 159 patients were assigned to the training group, and 48 patients comprised the validation group. T2 -weighted (T2 W), T1 -weighted (T1 W), diffusion-weighted MR imaging (b-values = 0, 500, 800, and 2000 seconds/mm2 ) and quantitative DCE-MRI were performed on a 3.0T MR scanner. Radiomics features were extracted from T2 WI, T1 WI, DKI, apparent diffusion coefficient (ADC) maps, and DCE pharmacokinetic parameter maps in the training set. Models based on each sequence or combinations of sequences were built using a support vector machine (SVM) classifier and used to differentiate benign and malignant breast lesions in the validation set. Optimal feature selection was performed by Spearman's rank correlation coefficients and the least absolute shrinkage and selection operator algorithm (LASSO). Receiver operating characteristic (ROC) curves were used to assess the diagnostic performance of the radiomics models in the validation set. The area under the ROC curve (AUC) of the optimal radiomics model, including T2 WI, DKI, and quantitative DCE-MRI parameter maps was 0.921, with an accuracy of 0.833. The AUCs of the models based on T1 WI, T2 WI, ADC map, DKI, and DCE pharmacokinetic parameter maps were 0.730, 0.791, 0.770, 0.788, and 0.836, respectively. The model based on radiomics features from T2 WI, DKI, and quantitative DCE pharmacokinetic parameter maps has a high discriminatory ability for benign and malignant breast lesions. 3 TECHNICAL EFFICACY STAGE: 2 J. Magn. Reson. Imaging 2020;52:596-607.

BackgroundTreatment of sepsis requires prompt initiation of appropriate empiric antibiotics. Coverage for Gram-negative bacilli (GNB) is decided based on risk factors for resistant organisms, as documented in both structured and unstructured data in the electronic health record. We hypothesized that adding large language models (LLMs) that can decode clinical narratives would improve performance of deep learning models in predicting antimicrobial resistance in GNB sepsis.MethodsThis retrospective cohort analysis included all adult patients who met CDC sepsis criteria admitted to 10 BJC hospitals from 1/2018 to 12/2023. Sepsis episodes were classified as community-onset (< 48hrs from hospitalization) or hospital-onset. GNB were stratified as ceftriaxone-susceptible (SS), ceftriaxone-resistant but cefepime-susceptible (RS), and ceftriaxone- and cefepime-resistant (RR). Culture negative or non-GNB sepsis were labeled SS. Deep learning models (XGBoost, DeepAMR) were developed using all available structured data (e.g., demographics, vitals, labs). LLMs (BioGPT, Medical-LLaMa3-8B) were developed on full or partial history and physical exam notes. Models were assessed using area under the receiver operating characteristic curve (AUROC) and area under precision-recall curve (AUPRC).ResultsA total of 29,517 patients were included, for 23,949 community-onset and 25,101 hospital-onset sepsis episodes with RS+RR GNB in 9% and 17%, respectively (Table 1). Relevant features of structured data models included history of prior infections and resistant organisms. DeepAMR was the best performing deep learning model with AUROC 0.85/AUPRC 0.40 and AUROC 0.81/AUPRC 0.50 for community- and hospital-onset sepsis, respectively (Table 2). The best performing LLM was Medical-LLaMa3-8B on full notes (Figure 1), with AUROC 0.74/AUPRC 0.21 and AUROC 0.67/AUPRC 0.30 for community- and hospital-onset sepsis, respectively. Combining deep learning and LLMs did not significantly improve the prediction of resistant infections compared to deep learning alone.ConclusionStructured medical data provided the highest prediction of resistant GNB infections in sepsis. More analysis is needed to understand why the unstructured data did not improve model predictive performance.DisclosuresMaria Cristina Vazquez Guillamet, MD, Bionano: Stocks/Bonds (Public Company)|Charisma: Stocks/Bonds (Public Company)|Ocugen: Stocks/Bonds (Public Company)

RATIONALE: Cachexia is associated with increased mortality in people with chronic obstructive pulmonary disease (COPD) and is characterized by unintentional weight loss (UWL) and muscle atrophy. While associations between COPD cachexia and measures of emphysema, muscle cross-sectional area, and muscle density on computed tomography (CT) imaging of the chest have been studied, less is known about higher-order CT features, such as those generated by radiomics. The objective of this study was to assess the ability of radiomics features to predict cachexia in COPD. METHODS: We analyzed inspiratory chest CT scans from 4,606 participants enrolled in the multicenter COPDGene cohort study who completed all study visits, with cachexia status determined by participant-reported UWL during five-year and ten-year follow-up visits. The airways, lung parenchyma, and skeletal muscle regions were segmented and evaluated using in-house algorithms. We extracted 187 standardized radiomic features using PyRadiomics v3.0. Four predictive models were evaluated using Visit 1 data; two clinical models: (1) Demographics (age, gender, and pack-years of smoking), (2) demographics with spirometry (FEV1, FVC), and two radiomics models: (3) radiomic features alone (skeletal muscle, lung parenchyma and airways), (4) combined radiomic and clinical features. XGBoost, a gradient boosting algorithm, and Borderline SMOTE, a balancing algorithm, were used to predict COPD cachexia. Model performance was assessed using the area under the receiver operating characteristic curve (AUC-ROC) and precision-recall curves (AUC-PR). Paired AUC comparisons were performed using DeLong's test. The relative importance of each feature was also assessed across all models. RESULTS: Participants had a mean age of 63.0±8.7 years, 44% were female, and 25% reported UWL. The radiomics-specific models showed greater discriminative performance compared to base clinical models (AUC-ROC=0.83, 95% CI:0.79-0.87, α=0.09, p&amp;lt;0.05 and AUC-ROC=0.84, 95% CI:0.80-0.88, α=0.10, p&amp;lt;0.05 for radiomics-only and combined model, respectively). Precision-recall metrics similarly showed better performance of radiomics-specific models with AUC-PR 0.85 and 0.86 for radiomics-only and combined models, respectively, versus base-clinical models (AUC-PR 0.76 and 0.78 for demographics and demographics plus spirometry). Figure 1 demonstrates the AUC-ROC and AUC-PR for the four predictive models. Feature importance analysis indicated both muscle and lung textural characteristics were among the most discriminative radiomics features. Overall, higher-order textural radiomic features dominated over first-order features. Notably, no features from the airways showed significant importance in predicting COPD cachexia. CONCLUSION: Radiomic features from both the lung parenchyma and skeletal muscle from inspiratory CT scans improve the prediction of incident cachexia in COPD patients over clinical data alone.

The World Health Organization glioma classification highlights genetic profiles, such as isocitrate dehydrogenase (IDH) mutation. This study developed an uncertainty-aware deep learning model to predict IDH mutations in glioma patients, employing conformal prediction (CP) for uncertainty quantification (UQ). The UCSF dataset split into training (70%) and validation (30%) sets, and the UPENN dataset divided into calibration (30%) and external test (70%) sets. We developed various 3D convolutional neural network models and selected the best based on the validation set area under the precision-recall curve (AUPRC). Also, we trained a logistic regression (LR) ensemble classifier on the training set and selected the best ensemble model based on the validation set AUPRC. Finally, we employed CP, with nonconformity thresholds set at the error rate of 0.01. The conformal model was calibrated on the calibration set and was evaluated on the external test set using the AUPRC and the area under the receiver operating characteristic (AUROC). 3D convolutional models achieved a mean AUROC and AUPRC of 0.9004 and 0.8102 on the validation set, and a mean AUROC and AUPRC of 0.7340 and 0.2139 on the external test set, respectively. Using the LR ensemble model, the model achieved an AUROC and AUPRC of 0.79820 and 0.2583 on the external test set, respectively. CP with a 0.01 nonconformity threshold with 0.9917 coverage reached an AUROC and AUPRC of 0.8592 and 0.5217, respectively. Integrating CP for UQ improves the performance of deep learning models for predicting IDH mutation in glioma patients.

Background and Objectives: Liver fibrosis is the key prognostic factor in patients with chronic liver diseases (CLD). Computed tomography (CT) is widely used in clinical practice, but it has limited value in assessing liver fibrosis in precirrhotic stages. Quantitative CT analysis based on radiomics can provide additional information by extracting hidden image patterns, but the optimal approach remains to be determined. The aims of this study were to evaluate automated CT-based radiomic models for predicting biopsy-proven liver fibrosis, to compare different segmentation strategies and organ inclusions approaches, and to assess its performance against vibration-controlled transient elastography (VCTE). We also examined whether these models could predict liver steatosis. Methods: In this retrospective study, 58 patients with biopsy-proven CLD and 9 controls underwent VCTE and contrast-enhanced abdominal CT within three months of biopsy. Radiomic features were extracted from portal-venous-phase images using both two-dimensional (2D) and three-dimensional (3D) segmentations of the liver, spleen, and combined liver–spleen. Multilayer perceptron neural (MLP) networks were trained to predict fibrosis staging (≥F1, ≥F2, ≥F3, and F4) and steatosis grading (≥S1, ≥S2, and S3). Model performance was assessed using area under the receiver operating characteristic curve (AUROC) and accuracy. Results: The 3D radiomic models outperformed 2D models in predicting liver fibrosis stages. In the 3D radiomic model category, the combined 3D liver–spleen model achieved very good to excellent performance (AUROCs 0.974, 0.929, 0.928, and 0.898, respectively, for ≥F1, ≥F2, ≥F3, and F4), with comparable results to VCTE (AUROCs 0.921, 0.957, 0.968, and 0.909, respectively, for ≥F1, ≥F2, ≥F3, and F4). Radiomic models showed poor predictive ability for steatosis grades (AUROCs 0.44–0.69) compared to controlled attenuation parameter (CAP) (AUROCs 0.798–0.917). Conclusions: CT-based radiomic models showed potential for predicting liver fibrosis stage. The 3D model of liver and spleen had the highest performance, comparable to VCTE. This approach could be valuable in clinical settings where elastography is unavailable or inconclusive and for opportunistic screening in patients already undergoing CT for other medical indications. In contrast, portal-venous-phase radiomics lacked predictive value for steatosis assessment. Larger, multicenter studies are required to validate these results.

Background: The accurate classification between malignant and benign breast lesions detected on mammograms is a crucial but difficult challenge for reducing false-positive recall rates and improving the efficacy of breast cancer screening. Objective: This study aims to optimize a new deep transfer learning model by implementing a novel attention mechanism in order to improve the accuracy of breast lesion classification. Methods: ResNet50 is selected as the base model to develop a new deep transfer learning model. To enhance the accuracy of breast lesion classification, we propose adding a convolutional block attention module (CBAM) to the standard ResNet50 model and optimizing a new model for this task. We assembled a large dataset with 4280 mammograms depicting suspicious soft-tissue mass-type lesions. A region of interest (ROI) is extracted from each image based on lesion center. Among them, 2480 and 1800 ROIs depict verified benign and malignant lesions, respectively. The image dataset is randomly split into two subsets with a ratio of 9:1 five times to train and test two ResNet50 models with and without using CBAM. Results: Using the area under ROC curve (AUC) as an evaluation index, the new CBAM-based ResNet50 model yields AUC = 0.866 ± 0.015, which is significantly higher than that obtained by the standard ResNet50 model (AUC = 0.772 ± 0.008) (p < 0.01). Conclusion: This study demonstrates that although deep transfer learning technology attracted broad research interest in medical-imaging informatic fields, adding a new attention mechanism to optimize deep transfer learning models for specific application tasks can play an important role in further improving model performances.

To explore and compare the potential value of radiomics models based on contrast-enhanced computed tomography (CT) for noninvasive preoperative prediction of lymphovascular invasion (LVI) in laryngeal squamous cell carcinoma (LSCC). This multicenter diagnostic study retrospectively enrolled patients with LSCC from three tertiary hospitals who underwent surgical treatment. Standardized preprocessing was performed on the CT images, followed by region-of-interest segmentation and extraction of traditional radiomics features and deep learning (DL) features. Features were selected using least absolute shrinkage and selection operator (LASSO) regression. Traditional radiomics models and deep learning radiomics (DLR) models were established using logistic regression, random forest, and multilayer perceptron algorithms, respectively. A transformer-based hybrid model was developed by integrating radiomics and DL features. The predictive performance of the three types of models was evaluated and compared using the area under the curve (AUC), decision curve analysis (DCA), sample probability distribution histograms, confusion matrices, calibration curves, net reclassification index (NRI), and integrated discrimination improvement (IDI). A total of 1024 patients were allocated to the training set (center1, n =291), internal validation set ( n =126), and external test sets (Center 2, n =437; Center 3, n =170). Three radiomics models and three DLR models were constructed, and the optimal performance was observed in the DLR_ Random Forest model (AUC: 0.812-0.867). The transformer hybrid model demonstrated superior predictive performance, with AUC values of 0.881, 0.843, 0.833, and 0.836 in the training, internal validation, and external test sets, respectively. DCA indicated a higher net benefit for the Transformer model, along with an improved NRI and IDI. Radiomics models based on CT images exhibit potential for noninvasive prediction of LVI in LSCC, with the transformer hybrid model achieving the highest diagnostic performance. This approach may provide clinicians with a preoperative decision support tool to optimize treatment strategies for patients with LSCC.

BackgroundIn the pediatric intensive care unit (PICU), quantifying illness severity can be guided by risk models to enable timely identification and appropriate intervention. Logistic regression models, including the pediatric index of mortality 2 (PIM-2) and pediatric risk of mortality III (PRISM-III), produce a mortality risk score using data that are routinely available at PICU admission. Artificial neural networks (ANNs) outperform regression models in some medical fields.ObjectiveIn light of this potential, we aim to examine ANN performance, compared to that of logistic regression, for mortality risk estimation in the PICU.MethodsThe analyzed data set included patients from North American PICUs whose discharge diagnostic codes indicated evidence of infection and included the data used for the PIM-2 and PRISM-III calculations and their corresponding scores. We stratified the data set into training and test sets, with approximately equal mortality rates, in an effort to replicate real-world data. Data preprocessing included imputing missing data through simple substitution and normalizing data into binary variables using PRISM-III thresholds. A 2-layer ANN model was built to predict pediatric mortality, along with a simple logistic regression model for comparison. Both models used the same features required by PIM-2 and PRISM-III. Alternative ANN models using single-layer or unnormalized data were also evaluated. Model performance was compared using the area under the receiver operating characteristic curve (AUROC) and the area under the precision recall curve (AUPRC) and their empirical 95% CIs.ResultsData from 102,945 patients (including 4068 deaths) were included in the analysis. The highest performing ANN (AUROC 0.871, 95% CI 0.862-0.880; AUPRC 0.372, 95% CI 0.345-0.396) that used normalized data performed better than PIM-2 (AUROC 0.805, 95% CI 0.801-0.816; AUPRC 0.234, 95% CI 0.213-0.255) and PRISM-III (AUROC 0.844, 95% CI 0.841-0.855; AUPRC 0.348, 95% CI 0.322-0.367). The performance of this ANN was also significantly better than that of the logistic regression model (AUROC 0.862, 95% CI 0.852-0.872; AUPRC 0.329, 95% CI 0.304-0.351). The performance of the ANN that used unnormalized data (AUROC 0.865, 95% CI 0.856-0.874) was slightly inferior to our highest performing ANN; the single-layer ANN architecture performed poorly and was not investigated further.ConclusionsA simple ANN model performed slightly better than the benchmark PIM-2 and PRISM-III scores and a traditional logistic regression model trained on the same data set. The small performance gains achieved by this two-layer ANN model may not offer clinically significant improvement; however, further research with other or more sophisticated model designs and better imputation of missing data may be warranted.

Develop a radiomics nomogram that integrates deep learning, radiomics, and clinical variables to predict epidermal growth factor receptor (EGFR) mutation status in patients with stage I non-small cell lung cancer (NSCLC). We retrospectively included 438 patients who underwent curative surgery and completed driver-gene mutation tests for stage I NSCLC from four academic medical centers. Predictive models were established by extracting and analyzing radiomic features in intratumoral, peritumoral, and habitat regions of CT images to identify EGFR mutation status in stage I NSCLC. Additionally, three deep learning models based on the intratumoral region were constructed. A nomogram was developed by integrating representative radiomic signatures, deep learning, and clinical features. Model performance was assessed by calculating the area under the receiver operating characteristic (ROC) curve. The established habitat radiomics features demonstrated encouraging performance in discriminating between EGFR mutant and wild-type, with predictive ability superior to other single models (AUC 0.886, 0.812, and 0.790 for the training, validation, and external test sets, respectively). The radiomics-based nomogram exhibited excellent performance, achieving the highest AUC values of 0.917, 0.837, and 0.809 in the training, validation, and external test sets, respectively. Decision curve analysis (DCA) indicated that the nomogram provided a higher net benefit than other radiomics models, offering valuable information for treatment.

Identifying final pathology of FNA-indeterminant nodules before surgical resection could decrease the number of unnecessary surgeries and total cost to patients. This project explores how radiomics (RM) and deep learning (DTLM) models may be combined to improve the potential for clinical interpretability of machine learning models in the task of classifying indeterminant thyroid nodules on ultrasound. Two radiomic and deep learning combination models were created: a simple classifier combination model (SCM) and an interpretability-driven combination model (ICM). SCM provided a nodule malignancy score. ICM merged radiomic and deep learning features through correlation and provided echogenicity-related, composition-related, and shape/margin-related malignancy scores which were averaged to yield an overall nodule malignancy score. Models were trained and tested on a de-identified dataset of 476 grayscale ultrasound images collected under IRB approval containing 222 images from 69 indeterminant nodules with a final pathology of malignant and 254 images from 82 indeterminant nodules with a final pathology of benign. Models were tested using 5- fold cross-validation by nodule over 100 iterations. Receiver-operating characteristic (ROC) analysis was conducted with area under the ROC curve (AUC) serving as the statistic of merit for model performance. Models yielded mean AUC [95%CI] of 0.75 [.67,.83], 0.70 [.62,.78], 0.77 [0.70,0.84], 0.76 [.69,.84] for RM, DTLM, SCM, and ICM respectively. This work failed to demonstrate a statistically significant difference in model performances. However, the ICM presents a novel method for combining radiomics and deep learning features focused on improving interpretability for clinical implementation in the task of indeterminant thyroid nodule classification.

To develop and test an optimized radiomics model based on multi-planar automated breast volume scan (ABVS) images to identify malignant and benign breast lesions. Patients (n=200) with breast lesions who underwent ABVS examinations were included. For each patient, 208 radiomics features were extracted from the ABVS images, including axial plane and coronal plane. Recursive feature elimination, random forest, and chi-square test were used to select features. A support vector machine, logistic regression, and extreme gradient boosting were utilized as classifiers to differentiate malignant and benign breast lesions. The area under the curve, sensitivity, specificity, accuracy, and precision was used to evaluate the performance of the radiomics models. Generalization of the radiomics models was verified through 5-fold cross-validation. For a single plane or a combination of planes, a combination of recursive feature elimination, and support vector machine yielded the best performance when identifying breast lesions. The machine learning models based on a combination of planes performed better than those based on a single plane. Regarding the axial plane and coronal plane, the machine learning model using a combination of recursive feature elimination and support vector machine yielded the optimal identification performance: average area under the curve (0.857 ± 0.058, 95% confidence interval, 0.763-0.957); the average values of sensitivity, specificity, accuracy, and precision were 87.9, 68.2, 80.7, and 82.9%, respectively. The optimized radiomics model based on ABVS images can provide valuable information for identifying benign and malignant breast lesions preoperatively and guide the accurate clinical treatment. Further external validation is required.

In this multicentric study, individual patient data from 3 different centers were analyzed. Consecutive patients receiving standardized multiparametric breast magnetic resonance imaging for standard nonscreening indications were included. At each center, 2 experienced radiologists with more than 5 years of experience retrospectively interpreted the examinations in consensus and applied the KS to every histologically verified lesion. The corresponding mean ADC of each lesion was measured using a Wielema type 4 region of interest. According to established methods, the KS and ADC were combined, yielding the KS+ score. Diagnostic accuracy was evaluated by the area under the receiver operating characteristics curve (AUROC) and compared between the KS, ADC, and KS+ (DeLong test). Likewise, the potential to help avoid unnecessary biopsies was compared between the KS, ADC, and KS+ based on established high sensitivity thresholds (McNemar test). A total of 450 lesions in 414 patients (mean age, 51.5 years; interquartile range, 42-60.8 years) were included, with 219 lesions being malignant (48.7%; 95% confidence interval [CI], 44%-53.4%). The performance of the KS (AUROC, 0.915; CI, 0.886-0.939) was significantly better than that of the ADC (AUROC, 0.848; CI, 0.811-0.880; P < 0.001). The largest difference between these parameters was observed when assessing subcentimeter lesions (AUROC, 0.909 for KS; CI, 0.849-0.950 vs 0.811 for ADC; CI, 0.737-0.871; P = 0.02).The use of the KS+ (AUROC, 0.918; CI, 0.889-0.942) improved the performance slightly, but without any significant difference relative to a single KS or ADC reading (P = 0.64).When applying high sensitivity thresholds for avoiding unnecessary biopsies, the KS and ADC achieved equal sensitivity (97.7% for both; cutoff values, >4 for KS and ≤1.4 × 10-3 mm2/s for ADC). However, the rate of potentially avoidable biopsies was higher when using the KS (specificity: 65.4% for KS vs 32.9% for ADC; P < 0.0001). The KS was superior to the KS+ in avoiding unnecessary biopsies. Both the KS and ADC may be used to distinguish benign from malignant breast lesions. However, KS proved superior in this task including, most of all, when assessing small lesions less than 1 cm. Using the KS may avoid twice as many unnecessary biopsies, and the combination of both the KS and ADS does not improve diagnostic performance.

BackgroundFetal alcohol syndrome (FAS) is a lifelong developmental disability that occurs among individuals with prenatal alcohol exposure (PAE). With improved prediction models, FAS can be diagnosed or treated early, if not completely prevented.ObjectiveIn this study, we sought to compare different machine learning algorithms and their FAS predictive performance among women who consumed alcohol during pregnancy. We also aimed to identify which variables (eg, timing of exposure to alcohol during pregnancy and type of alcohol consumed) were most influential in generating an accurate model.MethodsData from the collaborative initiative on fetal alcohol spectrum disorders from 2007 to 2017 were used to gather information about 595 women who consumed alcohol during pregnancy at 5 hospital sites around the United States. To obtain information about PAE, questionnaires or in-person interviews, as well as reviews of medical, legal, or social service records were used to gather information about alcohol consumption. Four different machine learning algorithms (logistic regression, XGBoost, light gradient-boosting machine, and CatBoost) were trained to predict the prevalence of FAS at birth, and model performance was measured by analyzing the area under the receiver operating characteristics curve (AUROC). Of the total cases, 80% were randomly selected for training, while 20% remained as test data sets for predicting FAS. Feature importance was also analyzed using Shapley values for the best-performing algorithm.ResultsOverall, there were 20 cases of FAS within a total population of 595 individuals with PAE. Most of the drinking occurred in the first trimester only (n=491) or throughout all 3 trimesters (n=95); however, there were also reports of drinking in the first and second trimesters only (n=8), and 1 case of drinking in the third trimester only (n=1). The CatBoost method delivered the best performance in terms of AUROC (0.92) and area under the precision-recall curve (AUPRC 0.51), followed by the logistic regression method (AUROC 0.90; AUPRC 0.59), the light gradient-boosting machine (AUROC 0.89; AUPRC 0.52), and XGBoost (AUROC 0.86; AURPC 0.45). Shapley values in the CatBoost model revealed that 12 variables were considered important in FAS prediction, with drinking throughout all 3 trimesters of pregnancy, maternal age, race, and type of alcoholic beverage consumed (eg, beer, wine, or liquor) scoring highly in overall feature importance. For most predictive measures, the best performance was obtained by the CatBoost algorithm, with an AUROC of 0.92, precision of 0.50, specificity of 0.29, F1 score of 0.29, and accuracy of 0.96.ConclusionsMachine learning algorithms were able to identify FAS risk with a prediction performance higher than that of previous models among pregnant drinkers. For small training sets, which are common with FAS, boosting mechanisms like CatBoost may help alleviate certain problems associated with data imbalances and difficulties in optimization or generalization.

Assessing robustness to adversarial attacks in attention-based networks: Case of EEG-based motor imagery classification

Multicenter deep learning for multi-abnormality screening on hip radiographs: development, external validation, and assisted reader study.