2023 Survey on User Experience of Artificial Intelligence Software in Radiology by the Korean Society of Radiology.

OBJECTIVE: In Korea, radiology has been positioned towards the early adoption of artificial intelligence-based software as medical devices (AI-SaMDs); however, little is known about the current usage, implementation, and future needs of AI-SaMDs. We surveyed the current trends and expectations for AI-SaMDs among members of the Korean Society of Radiology (KSR). MATERIALS AND METHODS: An anonymous and voluntary online survey was open to all KSR members between April 17 and May 15, 2023. The survey was focused on the experiences of using AI-SaMDs, patterns of usage, levels of satisfaction, and expectations regarding the use of AI-SaMDs, including the roles of the industry, government, and KSR regarding the clinical use of AI-SaMDs. RESULTS: Among the 370 respondents (response rate: 7.7% [370/4792]; 340 board-certified radiologists; 210 from academic institutions), 60.3% (223/370) had experience using AI-SaMDs. The two most common use-case of AI-SaMDs among the respondents were lesion detection (82.1%, 183/223), lesion diagnosis/classification (55.2%, 123/223), with the target imaging modalities being plain radiography (62.3%, 139/223), CT (42.6%, 95/223), mammography (29.1%, 65/223), and MRI (28.7%, 64/223). Most users were satisfied with AI-SaMDs (67.6% [115/170, for improvement of patient management] to 85.1% [189/222, for performance]). Regarding the expansion of clinical applications, most respondents expressed a preference for AI-SaMDs to assist in detection/diagnosis (77.0%, 285/370) and to perform automated measurement/quantification (63.5%, 235/370). Most respondents indicated that future development of AI-SaMDs should focus on improving practice efficiency (81.9%, 303/370) and quality (71.4%, 264/370). Overall, 91.9% of the respondents (340/370) agreed that there is a need for education or guidelines driven by the KSR regarding the use of AI-SaMDs. CONCLUSION: The penetration rate of AI-SaMDs in clinical practice and the corresponding satisfaction levels were high among members of the KSR. Most AI-SaMDs have been used for lesion detection, diagnosis, and classification. Most respondents requested KSR-driven education or guidelines on the use of AI-SaMDs.

Computed Tomography Radiomics for Preoperative Prediction of Spread Through Air Spaces in the Early Stage of Surgically Resected Lung Adenocarcinomas.

PURPOSE: To assess the added value of radiomics models from preoperative chest CT in predicting the presence of spread through air spaces (STAS) in the early stage of surgically resected lung adenocarcinomas using multiple validation datasets. MATERIALS AND METHODS: This retrospective study included 550 early-stage surgically resected lung adenocarcinomas in 521 patients, classified into training, test, internal validation, and temporal validation sets (n=211, 90, 91, and 158, respectively). Radiomics features were extracted from the segmented tumors on preoperative chest CT, and a radiomics score (Rad-score) was calculated to predict the presence of STAS. Diagnostic performance of the conventional model and the combined model, based on a combination of conventional and radiomics features, for the diagnosis of the presence of STAS were compared using the area under the curve (AUC) of the receiver operating characteristic curve. RESULTS: Rad-score was significantly higher in the STAS-positive group compared to the STAS-negative group in the training, test, internal, and temporal validation sets. The performance of the combined model was significantly higher than that of the conventional model in the training set {AUC: 0.784 [95% confidence interval (CI): 0.722-0.846] vs. AUC: 0.815 (95% CI: 0.759-0.872), p=0.042}. In the temporal validation set, the combined model showed a significantly higher AUC than that of the conventional model (p=0.001). The combined model showed a higher AUC than the conventional model in the test and internal validation sets, albeit with no statistical significance. CONCLUSION: A quantitative CT radiomics model can assist in the non-invasive prediction of the presence of STAS in the early stage of lung adenocarcinomas.

Defining Medical AI Competencies for Medical School Graduates: Outcomes of a Delphi Survey and Medical Student/Educator Questionnaire of South Korean Medical Schools.

PURPOSE: Given the increasing significance and potential impact of artificial intelligence (AI) technology on health care delivery, there is an increasing demand to integrate AI into medical school curricula. This study aimed to define medical AI competencies and identify the essential competencies for medical graduates in South Korea. METHOD: An initial Delphi survey conducted in 2022 involving 4 groups of medical AI experts (n = 28) yielded 42 competency items. Subsequently, an online questionnaire survey was carried out with 1,955 participants (1,174 students and 781 professors) from medical schools across South Korea, utilizing the list of 42 competencies developed from the first Delphi round. A subsequent Delphi survey was conducted with 33 medical educators from 21 medical schools to differentiate the essential AI competencies from the optional ones. RESULTS: The study identified 6 domains encompassing 36 AI competencies essential for medical graduates: (1) understanding digital health and changes driven by AI; (2) fundamental knowledge and skills in medical AI; (3) ethics and legal aspects in the use of medical AI; (4) medical AI application in clinical practice; (5) processing, analyzing, and evaluating medical data; and (6) research and development of medical AI, as well as subcompetencies within each domain. While numerous competencies within the first 4 domains were deemed essential, a higher percentage of experts indicated responses in the last 2 domains, data science and medical AI research and development, were optional. CONCLUSIONS: This medical AI framework of 6 competencies and their subcompetencies for medical graduates exhibits promising potential for guiding the integration of AI into medical curricula. Further studies conducted in diverse contexts and countries are necessary to validate and confirm the applicability of these findings. Additional research is imperative for developing specific and feasible educational models to integrate these proposed competencies into pre-existing curricula.

Development of a deep learning based image processing tool for enhanced organoid analysis.

Contrary to 2D cells, 3D organoid structures are composed of diverse cell types and exhibit morphologies of various sizes. Although researchers frequently monitor morphological changes, analyzing every structure with the naked eye is difficult. Given that deep learning (DL) has been used for 2D cell image segmentation, a trained DL model may assist researchers in organoid image recognition and analysis. In this study, we developed OrgaExtractor, an easy-to-use DL model based on multi-scale U-Net, to perform accurate segmentation of organoids of various sizes. OrgaExtractor achieved an average dice similarity coefficient of 0.853 from a post-processed output, which was finalized with noise removal. Correlation between CellTiter-Glo assay results and daily measured organoid images shows that OrgaExtractor can reflect the actual organoid culture conditions. The OrgaExtractor data can be used to determine the best time point for organoid subculture on the bench and to maintain organoids in the long term.

Identifying major predictors for parenting stress in a caregiver of autism spectrum disorder using machine learning models.

Introduction: Previous studies have investigated predictive factors for parenting stress in caregivers of autism spectrum disorder (ASD) patients using traditional statistical approaches, but their study settings and results were inconsistent. Herein, this study aimed to identify major predictors for parenting stress in this population by developing explainable machine learning models. Methods: Study participants were collected from the Department of Child and Adolescent Psychiatry, Severance Hospital, Yonsei University College of Medicine, Seoul, the Republic of Korea between March 2016 and October 2020. A total of 36 model features were used, which include subscales of the Minnesota Multiphasic Personality Inventory-2 (MMPI-2) for caregivers' psychopathology, Social Responsiveness Scale-2 for core symptoms, and Child Behavior Checklist (CBCL) for behavioral problems. Machine learning classifiers [eXtreme Gradient Boosting (XGBoost), random forest (RF), logistic regression, and support vector machine (SVM) classifier] were generated to predict severe total parenting stress and its subscales (parental distress, parent-child dysfunctional interaction, and difficult child). Model performance was assessed by area under the receiver operating curve (AUC), sensitivity, specificity, accuracy, positive predictive value, and negative predictive value. We utilized the SHapley Additive exPlanations tree explainer to investigate major predictors. Results: A total of 496 participants were included [mean age of ASD patients 6.39 (SD 2.24); 413 men (83.3%)]. The best-performing models achieved an AUC of 0.831 (RF model; 95% CI 0.740-0.910) for parental distress, 0.814 (SVM model; 95% CI 0.720-0.896) for parent-child dysfunctional interaction, 0.813 (RF model; 95% CI 0.724-0.891) for difficult child, and 0.862 (RF model; 95% CI 0.783-0.930) for total parenting stress on the test set. For the total parenting stress, ASD patients' aggressive behavior and anxious/depressed, and caregivers' depression, social introversion, and psychasthenia were the top 5 leading predictors. Conclusion: By using explainable machine learning models (XGBoost and RF), we investigated major predictors for each subscale of the parenting stress index in caregivers of ASD patients. Identified predictors for parenting stress in this population might help alert clinicians whether a caregiver is at a high risk of experiencing severe parenting stress and if so, providing timely interventions, which could eventually improve the treatment outcome for ASD patients.

A machine learning model for predicting hepatocellular carcinoma risk in patients with chronic hepatitis B.

BACKGROUND: Machine learning (ML) algorithms can be used to overcome the prognostic performance limitations of conventional hepatocellular carcinoma (HCC) risk models. We established and validated an ML-based HCC predictive model optimized for patients with chronic hepatitis B (CHB) infections receiving antiviral therapy (AVT). METHODS: Treatment-naïve CHB patients who were started entecavir (ETV) or tenofovir disoproxil fumarate (TDF) were enrolled. We used a training cohort (n = 960) to develop a novel ML model that predicted HCC development within 5 years and validated the model using an independent external cohort (n = 1937). ML algorithms consider all potential interactions and do not use predefined hypotheses. RESULTS: The mean age of the patients in the training cohort was 48 years, and most patients (68.9%) were men. During the median 59.3 (interquartile range 45.8-72.3) months of follow-up, 69 (7.2%) patients developed HCC. Our ML-based HCC risk prediction model had an area under the receiver-operating characteristic curve (AUC) of 0.900, which was better than the AUCs of CAMD (0.778) and REAL B (0.772) (both p < .05). The better performance of our model was maintained (AUC = 0.872 vs. 0.788 for CAMD and 0.801 for REAL B) in the validation cohort. Using cut-off probabilities of 0.3 and 0.5, the cumulative incidence of HCC development differed significantly among the three risk groups (p < .001). CONCLUSIONS: Our new ML model performed better than models in terms of predicting the risk of HCC development in CHB patients receiving AVT.

Detection of Pacemaker and Identification of MRI-conditional Pacemaker Based on Deep-learning Convolutional Neural Networks to Improve Patient Safety.

With the increased availability of magnetic resonance imaging (MRI) and a progressive rise in the frequency of cardiac device implantation, there is an increased chance that patients with implanted cardiac devices require MRI examination during their lifetime. Though MRI is generally contraindicated in patients who have undergone pacemaker implantation with electronic circuits, the recent introduction of MR Conditional pacemaker allows physicians to take advantage of MRI to assess these patients during diagnosis and treatment. When MRI examinations of patients with pacemaker are requested, physicians must confirm whether the device is a conventional pacemaker or an MR Conditional pacemaker by reviewing chest radiographs or the electronic medical records (EMRs). The purpose of this study was to evaluate the utility of a deep convolutional neural network (DCNN) trained to detect pacemakers on chest radiographs and to determine the device's subclassification. The DCNN perfectly detected pacemakers on chest radiographs and the accuracy of the subclassification of pacemakers using the internal and external test datasets were 100.0% (n = 106/106) and 90.1% (n = 279/308). The DCNN can be applied to the radiologic workflow for double-checking purposes, thereby improving patient safety during MRI and preventing busy physicians from making errors.

Influence of computed tomography slice thickness on deep learning-based, automatic coronary artery calcium scoring software performance.

Background: The influence of computed tomography (CT) slice thickness on the accuracy of deep learning (DL)-based, automatic coronary artery calcium (CAC) scoring software has not been explored yet. Methods: This retrospective study included 844 subjects (477 men, mean age of 58.9±10.7 years) who underwent electrocardiogram (ECG)-gated CAC scoring CT scans with 1.5 and 3 mm slice thickness values between September 2013 and October 2020. Automatic CAC scoring was performed using DL-based software (3D patch-based U-Net architectures). Manual CAC scoring was set as the reference standard. The reliability of automatic CAC scoring was evaluated using intraclass correlation coefficients (ICCs) for both the 1.5 and 3 mm datasets. The agreement of CAC severity categories [Agatston score (AS) 0, 1-100, 101-400, >400] between automatic CAC scoring and the reference standard was analyzed using weighted kappa (κ) statistics for both 1.5 and 3 mm datasets. Results: The CAC scoring agreement between the automatic CAC scoring and reference standard was excellent (ICC 0.982 for 1.5 mm, 0.969 for 3 mm, respectively). The categorical agreement of CAC severity between two methods was excellent for both 1.5 and 3 mm scans, with better agreement for 3 mm scans (weighted κ: 0.851 and 0.961, 95% confidence intervals: 0.823-0.879 and 0.945-0.974, respectively). Conclusions: Automatic CAC scoring shows excellent agreement with the reference standard for both 1.5 and 3 mm scans but results in lower agreement in the CAC severity category for 1.5 mm scans.

Chest Radiography of Tuberculosis: Determination of Activity Using Deep Learning Algorithm.

BACKGROUND: Inactive or old, healed tuberculosis (TB) on chest radiograph (CR) is often found in high TB incidence countries, and to avoid unnecessary evaluation and medication, differentiation from active TB is important. This study develops a deep learning (DL) model to estimate activity in a single chest radiographic analysis. METHODS: A total of 3,824 active TB CRs from 511 individuals and 2,277 inactive TB CRs from 558 individuals were retrospectively collected. A pretrained convolutional neural network was fine-tuned to classify active and inactive TB. The model was pretrained with 8,964 pneumonia and 8,525 normal cases from the National Institute of Health (NIH) dataset. During the pretraining phase, the DL model learns the following tasks: pneumonia vs. normal, pneumonia vs. active TB, and active TB vs. normal. The performance of the DL model was validated using three external datasets. Receiver operating characteristic analyses were performed to evaluate the diagnostic performance to determine active TB by DL model and radiologists. Sensitivities and specificities for determining active TB were evaluated for both the DL model and radiologists. RESULTS: The performance of the DL model showed area under the curve (AUC) values of 0.980 in internal validation, and 0.815 and 0.887 in external validation. The AUC values for the DL model, thoracic radiologist, and general radiologist, evaluated using one of the external validation datasets, were 0.815, 0.871, and 0.811, respectively. CONCLUSION: This DL-based algorithm showed potential as an effective diagnostic tool to identify TB activity, and could be useful for the follow-up of patients with inactive TB in high TB burden countries.

T1 Map-Based Radiomics for Prediction of Left Ventricular Reverse Remodeling in Patients With Nonischemic Dilated Cardiomyopathy.

OBJECTIVE: This study aimed to develop and validate models using radiomics features on a native T1 map from cardiac magnetic resonance (CMR) to predict left ventricular reverse remodeling (LVRR) in patients with nonischemic dilated cardiomyopathy (NIDCM). MATERIALS AND METHODS: Data from 274 patients with NIDCM who underwent CMR imaging with T1 mapping at Severance Hospital between April 2012 and December 2018 were retrospectively reviewed. Radiomic features were extracted from the native T1 maps. LVRR was determined using echocardiography performed ≥ 180 days after the CMR. The radiomics score was generated using the least absolute shrinkage and selection operator logistic regression models. Clinical, clinical + late gadolinium enhancement (LGE), clinical + radiomics, and clinical + LGE + radiomics models were built using a logistic regression method to predict LVRR. For internal validation of the result, bootstrap validation with 1000 resampling iterations was performed, and the optimism-corrected area under the receiver operating characteristic curve (AUC) with 95% confidence interval (CI) was computed. Model performance was compared using AUC with the DeLong test and bootstrap. RESULTS: Among 274 patients, 123 (44.9%) were classified as LVRR-positive and 151 (55.1%) as LVRR-negative. The optimism-corrected AUC of the radiomics model in internal validation with bootstrapping was 0.753 (95% CI, 0.698-0.813). The clinical + radiomics model revealed a higher optimism-corrected AUC than that of the clinical + LGE model (0.794 vs. 0.716; difference, 0.078 [99% CI, 0.003-0.151]). The clinical + LGE + radiomics model significantly improved the prediction of LVRR compared with the clinical + LGE model (optimism-corrected AUC of 0.811 vs. 0.716; difference, 0.095 [99% CI, 0.022-0.139]). CONCLUSION: The radiomic characteristics extracted from a non-enhanced T1 map may improve the prediction of LVRR and offer added value over traditional LGE in patients with NIDCM. Additional external validation research is required.

Deep-Learning-Based Detection of Vertebral Fracture and Osteoporosis Using Lateral Spine X-Ray Radiography.

Osteoporosis and vertebral fractures (VFs) remain underdiagnosed. The addition of deep learning methods to lateral spine radiography (a simple, widely available, low-cost test) can potentially solve this problem. In this study, we develop deep learning scores to detect osteoporosis and VF based on lateral spine radiography and investigate whether their use can improve referral of high-risk individuals to bone-density testing. The derivation cohort consisted of patients aged 50 years or older who underwent lateral spine radiography in Severance Hospital, Korea, from January 2007 to December 2018, providing a total of 26,299 lateral spine plain X-rays for 9276 patients (VF prevalence, 18.6%; osteoporosis prevalence, 40.3%). Two individual deep convolutional neural network scores to detect prevalent VF (VERTE-X pVF score) and osteoporosis (VERTE-X osteo score) were tested on an internal test set (20% hold-out set) and external test set (another hospital cohort [Yongin], 395 patients). VERTE-X pVF, osteo scores, and clinical models to detect prevalent VF or osteoporosis were compared in terms of the areas under the receiver-operating-characteristics curves (AUROCs). Net reclassification improvement (NRI) was calculated when using deep-learning scores to supplement clinical indications for classification of high-risk individuals to dual-energy X-ray absorptiometry (DXA) testing. VERTE-X pVF and osteo scores outperformed clinical models in both the internal (AUROC: VF, 0.93 versus 0.78; osteoporosis, 0.85 versus 0.79) and external (VF, 0.92 versus 0.79; osteoporosis, 0.83 versus 0.65; p < 0.01 for all) test sets. VERTE-X pVF and osteo scores improved the reclassification of individuals with osteoporosis to the DXA testing group when applied together with the clinical indications for DXA testing in both the internal (NRI 0.10) and external (NRI 0.14, p < 0.001 for all) test sets. The proposed method could detect prevalent VFs and osteoporosis, and it improved referral of individuals at high risk of fracture to DXA testing more than clinical indications alone. © 2023 The Authors. Journal of Bone and Mineral Research published by Wiley Periodicals LLC on behalf of American Society for Bone and Mineral Research (ASBMR).

Development and Validation of a Deep Learning-Based Synthetic Bone-Suppressed Model for Pulmonary Nodule Detection in Chest Radiographs.

Importance: Dual-energy chest radiography exhibits better sensitivity than single-energy chest radiography, partly due to its ability to remove overlying anatomical structures. Objectives: To develop and validate a deep learning-based synthetic bone-suppressed (DLBS) nodule-detection algorithm for pulmonary nodule detection on chest radiographs. Design, Setting, and Participants: This decision analytical modeling study used data from 3 centers between November 2015 and July 2019 from 1449 patients. The DLBS nodule-detection algorithm was trained using single-center data (institute 1) of 998 chest radiographs. The DLBS algorithm was validated using 2 external data sets (institute 2, 246 patients; and institute 3, 205 patients). Statistical analysis was performed from March to December 2021. Exposures: DLBS nodule-detection algorithm. Main Outcomes and Measures: The nodule-detection performance of DLBS model was compared with the convolution neural network nodule-detection algorithm (original model). Reader performance testing was conducted by 3 thoracic radiologists assisted by the DLBS algorithm or not. Sensitivity and false-positive markings per image (FPPI) were compared. Results: Training data consisted of 998 patients (539 men [54.0%]; mean [SD] age, 54.2 [9.82] years), and 2 external validation data sets consisted of 246 patients (133 men [54.1%]; mean [SD] age, 55.3 [8.7] years) and 205 patients (105 men [51.2%]; mean [SD] age, 51.8 [9.1] years). Using the external validation data set of institute 2, the bone-suppressed model showed higher sensitivity compared with that of the original model for nodule detection (91.5% [109 of 119] vs 79.8% [95 of 119]; P < .001). The overall mean of FPPI with the bone-suppressed model was reduced compared with the original model (0.07 [17 of 246] vs 0.09 [23 of 246]; P < .001). For the observer performance testing with the data of institute 3, the mean sensitivity of 3 radiologists was 77.5% (95% [CI], 69.9%-85.2%), whereas that of radiologists assisted by DLBS modeling was 92.1% (95% CI, 86.3%-97.3%; P < .001). The 3 radiologists had a reduced number of FPPI when assisted by the DLBS model (0.071 [95% CI, 0.041-0.111] vs 0.151 [95% CI, 0.111-0.210]; P < .001). Conclusions and Relevance: This decision analytical modeling study found that the DLBS model was more sensitive to detecting pulmonary nodules on chest radiographs compared with the original model. These findings suggest that the DLBS model could be beneficial to radiologists in the detection of lung nodules in chest radiographs without need of the specialized equipment or increase of radiation dose.

Fully automated radiomics-based machine learning models for multiclass classification of single brain tumors: Glioblastoma, lymphoma, and metastasis.

BACKGROUND AND PURPOSE: To investigate the diagnostic performance of fully automated radiomics-based models for multiclass classification of a single enhancing brain tumor among glioblastoma, central nervous system lymphoma, and metastasis. MATERIALS AND METHODS: The training and test sets were comprised of 538 cases (300 glioblastomas, 73 lymphomas, and 165 metastases) and 169 cases (101 glioblastomas, 29 lymphomas, and 39 metastases), respectively. After fully automated segmentation, radiomic features were extracted. Three conventional machine learning classifiers, including least absolute shrinkage and selection operator (LASSO), adaptive boosting (Adaboost), and support vector machine with the linear kernel (SVC), combined with one of four feature selection methods, including forward sequential feature selection, F score, mutual information, and LASSO, were trained. Additionally, one ensemble classifier based on the three classifiers was used. The diagnostic performance of the optimized models was tested in the test set using the accuracy, F1-macro score, and the area under the receiver operating characteristic curve (AUCROC). RESULTS: The best performance was achieved when the LASSO was used as a feature selection method. In the test set, the best performance was achieved by the ensemble classifier, showing an accuracy of 76.3% (95% CI, 70.0-82.7), a F1-macro score of 0.704, and an AUCROC of 0.878. CONCLUSION: Our fully automated radiomics-based models for multiclass classification might be useful for differential diagnosis of a single enhancing brain tumor with a good diagnostic performance and generalizability.

Clinical application of deep learning-based synthetic CT from real MRI to improve dose planning accuracy in Gamma Knife radiosurgery: a proof of concept study.

Dose planning for Gamma Knife radiosurgery (GKRS) uses the magnetic resonance (MR)-based tissue maximum ratio (TMR) algorithm, which calculates radiation dose without considering heterogeneous radiation attenuation in the tissue. In order to plan the dose considering the radiation attenuation, the Convolution algorithm should be used, and additional radiation exposure for computed tomography (CT) and registration errors between MR and CT are entailed. This study investigated the clinical feasibility of synthetic CT (sCT) from GKRS planning MR using deep learning. The model was trained using frame-based contrast-enhanced T1-weighted MR images and corresponding CT slices from 54 training subjects acquired for GKRS planning. The model was applied prospectively to 60 lesions in 43 patients including benign tumor such as meningioma and pituitary adenoma, metastatic brain tumors, and vascular disease of various location for evaluating the model and its application. We evaluated the sCT and compared between treatment plans made with MR only (TMR 10 plan), MR and real CT (rCT; Convolution with rCT [Conv-rCT] plan), and MR and synthetic CT (Convolution with sCT [Conv-sCT] plan). The mean absolute error (MAE) of 43 sCT was 107.35 ± 16.47 Hounsfield units. The TMR 10 treatment plan differed significantly from plans made by Conv-sCT and Conv-rCT. However, the Conv-sCT and Conv-rCT plans were similar. This study showed the practical applicability of deep learning based on sCT in GKRS. Our results support the possibility of formulating GKRS treatment plans while considering radiation attenuation in the tissue using GKRS planning MR and no radiation exposure.

Platelet-to-White Blood Cell Ratio Is Associated with Adverse Outcomes in Cirrhotic Patients with Acute Deterioration.

Background: The platelet-to-white blood cell ratio (PWR) is a hematologic marker of the systemic inflammatory response. Recently, the PWR was revealed to have a role as an independent prognostic factor for mortality in patients with hepatitis B virus (HBV)-related acute-on-chronic failure (ACLF) and HBV-related liver cirrhosis (LC) with acute decompensation (AD). However, the prognostic role of the PWR still needs to be investigated in LC patients with AD. In this study, we analyzed whether the PWR could stratify the risk of adverse outcomes (death or liver transplantation (LT)) in these patients. Methods: A prospective cohort of 1670 patients with AD of liver cirrhosis ((age: 55.2 ± 7.8, male = 1226 (73.4%)) was enrolled and evaluated for 28-day and overall adverse outcomes. Results: During a median follow-up of 8.0 months (range, 1.9−15.5 months), 424 (25.4%) patients had adverse outcomes (death = 377, LT = 47). The most common etiology of LC was alcohol use (69.7%). The adverse outcome rate was higher for patients with a PWR ≤ 12.1 than for those with a PWR > 12.1. A lower PWR level was a prognostic factor for 28-day adverse outcomes (PWR: hazard ratio 1.707, p = 0.034) when adjusted for the etiology of cirrhosis, infection, ACLF, and the MELD score. In the subgroup analysis, the PWR level stratified the risk of 28-day adverse outcomes regardless of the presence of ACLF or the main form of AD but not for those with bacterial infection. Conclusions: A lower PWR level was associated with 28-day adverse outcomes, indicating that the PWR level can be a useful and simple tool for stratifying the risk of 28-day adverse outcomes in LC patients with AD.

Development and Validation of a Visually Explainable Deep Learning Model for Classification of C-shaped Canals of the Mandibular Second Molars in Periapical and Panoramic Dental Radiographs.

INTRODUCTION: The purpose of this study was to develop and validate a visually explainable deep learning model for the classification of C-shaped canals of the mandibular second molars in dental radiographs. METHODS: The periapical and panoramic images of 1000 mandibular second molars were collected from 372 patients. The diagnostic performance of the deep learning system using periapical and panoramic radiographs was investigated in respect to its ability to determine whether the second mandibular molar showed a C-shaped canal configuration. The assessment of the canal configuration of cone-beam computed tomographic volumes from 372 patients (740 mandibular second molars) was used as a gold standard. RESULTS: The deep convolutional neural network algorithm model showed high accuracy in predicting the C-shaped canal variation among mandibular second molars in both periapical and panoramic images. The model demonstrated best results when using image patches including only the root portion of the tooth and when using both periapical and panoramic images for training (area under the curve [AUC] = 0.99). The model's diagnostic performance using only the root portion of the tooth (AUC: periapical = 0.98 and panoramic = 0.95) was similar to a specialist (AUC: periapical = 0.95 and panoramic = 0.96) and better than a novice general clinician (AUC: periapical = 0.89 and panoramic = 0.91). Both the specialist and general clinician showed better diagnostic performance when reading panoramic radiographs compared with periapical images. CONCLUSIONS: With further optimization of the test data using a larger data set and improvements made in the model, a deep learning system may be expected to effectively diagnose C-shaped canals and aid clinicians in practice and education.

A Deep Learning Model with High Standalone Performance for Diagnosis of Unruptured Intracranial Aneurysm.

PURPOSE: This study aimed to investigate whether a deep learning model for automated detection of unruptured intracranial aneurysms on time-of-flight (TOF) magnetic resonance angiography (MRA) can achieve a target diagnostic performance comparable to that of human radiologists for approval from the Korean Ministry of Food and Drug Safety as an artificial intelligence-applied software. MATERIALS AND METHODS: In this single-center, retrospective, confirmatory clinical trial, the diagnostic performance of the model was evaluated in a predetermined test set. After sample size estimation, the test set consisted of 135 aneurysm-containing examinations with 168 intracranial aneurysms and 197 aneurysm-free examinations. The target sensitivity and specificity were set as 87% and 92%, respectively. The patient-wise sensitivity and specificity of the model were analyzed. Moreover, the lesion-wise sensitivity and false-positive detection rate per case were also investigated. RESULTS: The sensitivity and specificity of the model were 91.11% [95% confidence interval (CI): 84.99, 95.32] and 93.91% (95% CI: 89.60, 96.81), respectively, which met the target performance values. The lesion-wise sensitivity was 92.26%. The overall false-positive detection rate per case was 0.123. Of the 168 aneurysms, 13 aneurysms from 12 examinations were missed by the model. CONCLUSION: The present deep learning model for automated detection of unruptured intracranial aneurysms on TOF MRA achieved the target diagnostic performance comparable to that of human radiologists. With high standalone performance, this model may be useful for accurate and efficient diagnosis of intracranial aneurysm.

Radiomics machine learning study with a small sample size: Single random training-test set split may lead to unreliable results.

This study aims to determine how randomly splitting a dataset into training and test sets affects the estimated performance of a machine learning model and its gap from the test performance under different conditions, using real-world brain tumor radiomics data. We conducted two classification tasks of different difficulty levels with magnetic resonance imaging (MRI) radiomics features: (1) "Simple" task, glioblastomas [n = 109] vs. brain metastasis [n = 58] and (2) "difficult" task, low- [n = 163] vs. high-grade [n = 95] meningiomas. Additionally, two undersampled datasets were created by randomly sampling 50% from these datasets. We performed random training-test set splitting for each dataset repeatedly to create 1,000 different training-test set pairs. For each dataset pair, the least absolute shrinkage and selection operator model was trained and evaluated using various validation methods in the training set, and tested in the test set, using the area under the curve (AUC) as an evaluation metric. The AUCs in training and testing varied among different training-test set pairs, especially with the undersampled datasets and the difficult task. The mean (±standard deviation) AUC difference between training and testing was 0.039 (±0.032) for the simple task without undersampling and 0.092 (±0.071) for the difficult task with undersampling. In a training-test set pair with the difficult task without undersampling, for example, the AUC was high in training but much lower in testing (0.882 and 0.667, respectively); in another dataset pair with the same task, however, the AUC was low in training but much higher in testing (0.709 and 0.911, respectively). When the AUC discrepancy between training and test, or generalization gap, was large, none of the validation methods helped sufficiently reduce the generalization gap. Our results suggest that machine learning after a single random training-test set split may lead to unreliable results in radiomics studies especially with small sample sizes.

Bone Radiomics Score Derived From DXA Hip Images Enhances Hip Fracture Prediction in Older Women.

Dual-energy X-ray absorptiometry (DXA)-based bone mineral density testing is standard to diagnose osteoporosis to detect individuals at high risk of fracture. A radiomics approach to extract quantifiable texture features from DXA hip images may improve hip fracture prediction without additional costs. Here, we investigated whether bone radiomics scores from DXA hip images could improve hip fracture prediction in a community-based cohort of older women. The derivation set (143 women who sustained hip fracture [mean age 73 years, time to fracture median 2.1 years] versus 290 age-matched women [mean age 73 years] who did not sustain hip fracture during follow-up [median 5.5 years]) were split into the train set (75%) and the test set (25% hold-out set). Among various models using 14 selected features out of 300 texture features mined from DXA hip images in the train set, random forest model was selected as the best model to build a bone radiomics score (range 0 to 100) based on the performance in the test set. In a community-based cohort (2029 women, mean age 71 years) as the clinical validation set, the bone radiomics score was calculated using a model fitted in the train set. A total of 34 participants (1.7%) sustained hip fracture during median follow-up of 5.4 years (mean bone radiomics score 40 ± 16 versus 28 ± 12 in non-fractured, p < 0.001). A one-point bone radiomics score increment was associated with a 4% elevated risk of incident hip fracture (adjusted hazard ratio [aHR] = 1.04, p = 0.001) after adjustment for age, body mass index (BMI), previous history of fracture, and femoral neck T-score, with improved model fit when added to covariates (likelihood ratio chi-square 10.74, p = 0.001). The association between bone radiomics score with incident hip fracture remained robust (aHR = 1.06, p < 0.001) after adjustment for FRAX hip fracture probability. Bone radiomics scores estimated from texture features of DXA hip images have the potential to improve hip fracture prediction. © 2021 American Society for Bone and Mineral Research (ASBMR).

Radiomics With Ensemble Machine Learning Predicts Dopamine Agonist Response in Patients With Prolactinoma.

CONTEXT: Early identification of the response of prolactinoma patients to dopamine agonists (DA) is crucial in treatment planning. OBJECTIVE: To develop a radiomics model using an ensemble machine learning classifier with conventional magnetic resonance images (MRIs) to predict the DA response in prolactinoma patients. DESIGN: Retrospective study. SETTING: Severance Hospital, Seoul, Korea. PATIENTS: A total of 177 prolactinoma patients who underwent baseline MRI (109 DA responders and 68 DA nonresponders) were allocated to the training (n = 141) and test (n = 36) sets. Radiomic features (n = 107) were extracted from coronal T2-weighed MRIs. After feature selection, single models (random forest, light gradient boosting machine, extra-trees, quadratic discrimination analysis, and linear discrimination analysis) with oversampling methods were trained to predict the DA response. A soft voting ensemble classifier was used to achieve the final performance. The performance of the classifier was validated in the test set. RESULTS: The ensemble classifier showed an area under the curve (AUC) of 0.81 [95% confidence interval (CI), 0.74-0.87] in the training set. In the test set, the ensemble classifier showed an AUC, accuracy, sensitivity, and specificity of 0.81 (95% CI, 0.67-0.96), 77.8%, 78.6%, and 77.3%, respectively. The ensemble classifier achieved the highest performance among all the individual models in the test set. CONCLUSIONS: Radiomic features may be useful biomarkers to predict the DA response in prolactinoma patients.