Performance of Multiple Pretrained BERT Models to Automate and Accelerate Data Annotation for Large Datasets.

Purpose: To develop and evaluate domain-specific and pretrained bidirectional encoder representations from transformers (BERT) models in a transfer learning task on varying training dataset sizes to annotate a larger overall dataset. Materials and Methods: The authors retrospectively reviewed 69 095 anonymized adult chest radiograph reports (reports dated April 2020-March 2021). From the overall cohort, 1004 reports were randomly selected and labeled for the presence or absence of each of the following devices: endotracheal tube (ETT), enterogastric tube (NGT, or Dobhoff tube), central venous catheter (CVC), and Swan-Ganz catheter (SGC). Pretrained transformer models (BERT, PubMedBERT, DistilBERT, RoBERTa, and DeBERTa) were trained, validated, and tested on 60%, 20%, and 20%, respectively, of these reports through fivefold cross-validation. Additional training involved varying dataset sizes with 5%, 10%, 15%, 20%, and 40% of the 1004 reports. The best-performing epochs were used to assess area under the receiver operating characteristic curve (AUC) and determine run time on the overall dataset. Results: The highest average AUCs from fivefold cross-validation were 0.996 for ETT (RoBERTa), 0.994 for NGT (RoBERTa), 0.991 for CVC (PubMedBERT), and 0.98 for SGC (PubMedBERT). DeBERTa demonstrated the highest AUC for each support device trained on 5% of the training set. PubMedBERT showed a higher AUC with a decreasing training set size compared with BERT. Training and validation time was shortest for DistilBERT at 3 minutes 39 seconds on the annotated cohort. Conclusion: Pretrained and domain-specific transformer models required small training datasets and short training times to create a highly accurate final model that expedites autonomous annotation of large datasets.Keywords: Informatics, Named Entity Recognition, Transfer Learning Supplemental material is available for this article. ©RSNA, 2022See also the commentary by Zech in this issue.

LI-RADS treatment response algorithm after first-line DEB-TACE: reproducibility and prognostic value at initial post-treatment CT/MRI.

PURPOSE: To evaluate the inter-reader reproducibility and prognostic accuracy of the Liver Imaging Reporting and Data System (LI-RADS) treatment response algorithm (LR-TR) at the time of initial post-treatment evaluation following drug-eluting beads transarterial chemoembolization (DEB-TACE) for hepatocellular carcinoma (HCC). METHODS: This retrospective study included patients with HCC who underwent first-line DEB-TACE between January 2011 and December 2015. Six readers (three fellowship-trained radiologists and three radiology trainees) independently assessed lesion-level response in up to two treated lesions per LR-TR and modified Response Evaluation Criteria in Solid Tumors (mRECIST)-target criteria, as well as patient-level response per mRECIST-overall criteria, on the initial post-treatment CT/MRI. Inter-reader agreement was calculated by Fleiss' multi-reader κ. We tested whether LR-TR, mRECIST-target, and mRECIST-overall response were associated with overall survival using Kaplan-Meier and Cox proportional hazard model analyses. RESULTS: A total of 82 patients with 113 treated target lesions were included. Inter-reader agreement was moderate for LR-TR and mRECIST-overall (κ range 0.42-0.57), and substantial for mRECIST-target (κ range 0.62-0.66), among all three reader-groups: all readers, experienced readers, and less-experienced readers. LR-TR and mRECIST-target response were not significantly associated with overall survival regardless of reader experience (P > 0.05). In contrast, mRECIST-overall response was significantly associated with overall survival when assessed by all readers (P = 0.02) and experienced readers (P = 0.03), but not by the less-experienced readers (P = 0.35). CONCLUSION: Although LR-TR algorithm has moderate inter-reader reproducibility, it alone may not predict overall survival on the initial post-treatment CT/MRI after first-line DEB-TACE for HCC.

Variability in utilization and techniques of pelvic floor imaging: findings of the SAR pelvic floor dysfunction disease-focused panel.

Pelvic floor disorders are common and can negatively impact quality of life. Imaging of patients with pelvic floor disorders has been extremely heterogeneous between institutions due in part to variations in clinical expectations, technical considerations, and radiologist experience. In order to assess variations in utilization and technique of pelvic floor imaging across practices, the society of abdominal radiology (SAR) disease-focused panel on pelvic floor dysfunction developed and administered an online survey to radiologists including the SAR membership. Results of the survey were compared with published recommendations for pelvic floor imaging to identify areas in need of further standardization. MRI was the most commonly reported imaging technique for pelvic floor imaging followed by fluoroscopic defecography. Ultrasound was only used by a small minority of responding radiologists. The survey responses demonstrated variability in imaging utilization, patient referral patterns, imaging protocols, patient education, and interpretation and reporting of pelvic floor imaging examinations. This survey highlighted inconsistencies in technique between institutions as well as potential gaps in knowledge that should be addressed to standardize evaluation of patients with pelvic floor dysfunction.

MR defecography technique: recommendations of the society of abdominal radiology's disease-focused panel on pelvic floor imaging.

PURPOSE: To develop recommendations for magnetic resonance (MR) defecography technique based on consensus of expert radiologists on the disease-focused panel of the Society of Abdominal Radiology (SAR). METHODS: An extensive questionnaire was sent to a group of 20 experts from the disease-focused panel of the SAR. The questionnaire encompassed details of technique and MRI protocol used for evaluating pelvic floor disorders. 75% agreement on questionnaire responses was defined as consensus. RESULTS: The expert panel reached consensus for 70% of the items and provided the basis of these recommendations for MR defecography technique. There was unanimous agreement that patients should receive coaching and explanation of commands used during MR defecography, the rectum should be distended with contrast agent, and that sagittal T2-weighted images should include the entire pelvis within the field of view. The panel also agreed unanimously that IV contrast should not be used for MR defecography. Additional areas of consensus ranged in agreement from 75 to 92%. CONCLUSION: We provide a set of consensus recommendations for MR defecography technique based on a survey of expert radiologists in the SAR pelvic floor dysfunction disease-focused panel. These recommendations can be used to develop a standardized imaging protocol.

Use of Spectral Detector Computed Tomography to Improve Liver Segmentation and Volumetry.

INTRODUCTION: Liver segmentation and volumetry have traditionally been performed using computed tomography (CT) attenuation to discriminate liver from other tissues. In this project, we evaluated if spectral detector CT (SDCT) can improve liver segmentation over conventional CT on 2 segmentation methods. MATERIALS AND METHODS: In this Health Insurance Portability and Accountability Act-compliant institutional review board-approved retrospective study, 30 contrast-enhanced SDCT scans with healthy livers were selected. The first segmentation method is based on Gaussian mixture models of the SDCT data. The second method is a convolutional neural network-based technique called U-Net. Both methods were compared against equivalent algorithms, which used conventional CT attenuation, with hand segmentation as the reference standard. Agreement to the reference standard was assessed using Dice similarity coefficient. RESULTS: Dice similarity coefficients to the reference standard are 0.93 ± 0.02 for the Gaussian mixture model method and 0.90 ± 0.04 for the CNN-based method (all 2 methods applied on SDCT). These were significantly higher compared with equivalent algorithms applied on conventional CT, with Dice coefficients of 0.90 ± 0.06 (P = 0.007) and 0.86 ± 0.06 (P < 0.001), respectively. CONCLUSION: On both liver segmentation methods tested, we demonstrated higher segmentation performance when the algorithms are applied on SDCT data compared with equivalent algorithms applied on conventional CT data.

Variations in use of "water soluble contrast challenge" for small bowel obstruction among academic radiologists: results of a national survey.

OBJECTIVE: To assess the frequency and indications for use of oral water-soluble contrast challenge as a diagnostic test for small bowel obstruction in four regions of the USA. MATERIALS AND METHODS: We distributed a 9-question web-based survey to the abdominal section heads of academic radiology departments throughout the USA (N = 97). The questions pertained to use of water-soluble contrast for management of small bowel obstruction. Descriptive statistics and Fisher's exact tests were used for data analysis. RESULTS: The overall response rate was 46%. Eighty percent of the responding hospitals had more than 500 beds in operation. Water-soluble contrast challenge was considered standard of care for management of non-operative small bowel obstruction in 60% of the responding radiology departments. The majority of the responding departments (41%) performed 2-8 contrast challenge studies per month on average. The most frequent indication for the study was distinguishing partial vs complete bowel obstruction. Eighty percent of the responding radiologists believed that the contrast challenge is useful for management of small bowel obstruction. Overall, there was no statistically significant difference in frequency and indication for use of water-soluble contrast challenge based on geographic location. CONCLUSION: The water-soluble contrast challenge was considered standard of care for non-operative management of small bowel obstruction in majority of the academic radiology departments represented in this survey. Surgeons were referring clinicians in every case. The most common clinical indication for the study was distinguishing partial versus complete small bowel obstruction.

A Technique to Identify Isoattenuating Gallstones with Dual-Layer Spectral CT: An ex Vivo Phantom Study.

Background Previously reported dual-energy CT methods for detecting noncalcified gallstones have reduced accuracy for gallstones smaller than 9 mm. Purpose To develop a dual-energy CT method for differentiating isoattenuating gallstones from bile and compare it with previously reported dual-energy CT methods by using a prospective ex vivo phantom reader study. Materials and Methods From May 2017 to May 2018, gallstones were collected from 105 patients (34 men; mean age, 51 years; age range, 18-84 years) undergoing cholecystectomy and placed inside 120-mL vials containing ox bile. The vials were placed inside a water-filled phantom and were scanned with dual-layer dual-energy CT. Thirty isoattenuating gallstones (4.3-24.7 mm in diameter) were evaluated. Conventional CT images, virtual noncontrast images, and monoenergetic images at 200 and 40 keV were created. Segmented images were created by using a two-dimensional histogram of Compton and photoelectric attenuation. Six readers evaluated the presence of isoattenuating gallstones in each image. Intra- and interreader agreement was measured by using percentage agreement, diagnostic performance was evaluated by using mean area under the receiver operating characteristic curve (AUC) estimates and pairwise comparisons, and the agreement of gallstone sizes measured at pathologic examination with those measured on segmented images was compared by using Bland-Altman analysis. Results For all gallstones, segmented images provided the highest mean intrareader (88.1%) and interreader (88.2% and 93.6%) agreements for all readers and reading sessions and the highest overall AUC (0.99; 95% confidence interval [CI]: 0.97, 1.00; adjusted P < .02 for all). For gallstones larger than 9 mm, no significant difference was found between the segmented and monoenergetic AUCs (all P > .94, adjusted P > .05 for all). For gallstones measuring 9 mm or smaller, the segmented images had the highest overall AUC (0.99; 95% CI: 0.97, 1.00; adjusted P < .01 for all). The mean difference in stone sizes was -0.6 mm, with limits of agreement from 2.6 to -3.8 mm. Conclusion Segmented images from Compton and photoelectric attenuation coefficients improve detection of isoattenuating gallstones compared with previously reported dual-energy CT methods. © RSNA, 2019 Online supplemental material is available for this article. See also the editorial by Matos in this issue.

Tumors of Renal Collecting Systems, Renal Pelvis, and Ureters: Role of MR Imaging and MR Urography Versus Computed Tomography Urography.

Hematuria evaluation remains a common problem, particularly in patients who smoke and are at risk for urothelial tumors. Lifetime surveillance of the urothelium is often required once urothelial cancer is diagnosed. Computed tomography urography (CTU) has exquisite sensitivity and specificity for identification of renal and urothelial lesions. The examination is well accepted by patients and physicians. Possible harms include radiation exposure and contrast-induced nephropathy. MR imaging is also an accurate test, but requires longer exam times, and may not demonstrate stones. We present the technical and interpretation skills required to use MR urography and CTU effectively.

Quantitative accuracy of positron emission tomography/magnetic resonance and positron emission tomography/computed tomography for cervical cancer.

With the spread of positron emission tomography/magnetic resonance (PET/MR), the question of comparability of studies becomes important. We aim to determine whether PET/MR and PET/computed tomography (PET/CT) are comparable for the case of cervical cancer. Fifteen cervical cancer patients identified by either a radiation oncologist or an oncologic surgeon had both PET/MR and PET/CT performed for initial staging within 3 weeks. We then compared the results both quantitatively (measuring standardized uptake values [SUVs] on visible lesions) as well as qualitatively (having radiologists and nuclear medicine physicians interprets the results). While interpretations between PET/MR and PET/CT varied in many cases, SUVs of primary lesions were similar to within 25% in all but one case, and correlation coefficient was 0.92. Maximum SUV ranged between 4.9 and 25.2 for PET-MR and between 5.8 and 30.4 for PET-CT for primary tumors and between 1.5 and 18.8 for PET-MR and between 1.8 and 20.8 for PET-CT for nodes. However, clinical reads often varied significantly between PET/MR and PET/CT. This suggests that SUV is similar on PET/MR and PET/CT although the differing anatomic modalities available for correlation may make the difference in terms of qualitative interpretation.

Phantom Validation of Spectral Detector Computed Tomography-Derived Virtual Monoenergetic, Virtual Noncontrast, and Iodine Quantification Images.

PURPOSE: Spectral detector computed tomography (SDCT) is a new CT technology that uses a dual-layer detector to perform energy separation. We aim to assess 3 clinical concepts using a phantom model: noise profile across the virtual monoenergetic (VME) spectrum, accuracy of iodine quantification, and virtual noncontrast (VNC) reconstructions' ability to remove iodine contribution to attenuation. METHODS: Six vials containing varying concentrations of iodinated contrast (0-6 mg/mL) diluted in water were placed in a water bath and scanned on an SDCT scanner. Virtual monoenergetic (40-200 keV at 10-keV increments), iodine-no-water, and VNC reconstructions were created. Attenuation (in Hounsfield units [HU]), VME noise at each energy level, CT-derived iodine concentration, and VNC attenuation were recorded. RESULTS: Virtual monoenergetic noise was improved at all energies compared with conventional images (conventional, 9.8-11.2; VME, 7.5-9.5). Noise profile showed a slightly higher image noise at 40 keV, but was otherwise relatively flat across the energy spectrum. On iodine-no-water reconstructions, measured varied from actual iodine concentration by ±0.1 mg/mL (SD, 0.16-0.36). Virtual noncontrast attenuation was within 5 HU of water attenuation at all iodine concentrations. CONCLUSION: Reconstructions of SDCT show lower VME image noise, accurate iodine quantification, and VNC attenuation values within 5 HU of expected in a phantom model.

Dual-layer spectral detector CT: non-inferiority assessment compared to dual-source dual-energy CT in discriminating uric acid from non-uric acid renal stones ex vivo.

PURPOSE: To assess the non-inferiority of dual-layer spectral detector CT (SDCT) compared to dual-source dual-energy CT (dsDECT) in discriminating uric acid (UA) from non-UA stones. METHODS: Fifty-seven extracted urinary calculi were placed in a cylindrical phantom in a water bath and scanned on a SDCT scanner (IQon, Philips Healthcare) and second- and third-generation dsDECT scanners (Somatom Flash and Force, Siemens Healthcare) under matched scan parameters. For SDCT data, conventional images and virtual monoenergetic reconstructions were created. A customized 3D growing region segmentation tool was used to segment each stone on a pixel-by-pixel basis for statistical analysis. Median virtual monoenergetic ratios (VMRs) of 40/200, 62/92, and 62/100 for each stone were recorded. For dsDECT data, dual-energy ratio (DER) for each stone was recorded from vendor-specific postprocessing software (Syngo Via) using the Kidney Stones Application. The clinical reference standard of X-ray diffraction analysis was used to assess non-inferiority. Area under the receiver-operating characteristic curve (AUC) was used to assess diagnostic performance of detecting UA stones. RESULTS: Six pure UA, 47 pure calcium-based, 1 pure cystine, and 3 mixed struvite stones were scanned. All pure UA stones were correctly separated from non-UA stones using SDCT and dsDECT (AUC = 1). For UA stones, median VMR was 0.95-0.99 and DER 1.00-1.02. For non-UA stones, median VMR was 1.4-4.1 and DER 1.39-1.69. CONCLUSION: SDCT spectral reconstructions demonstrate similar performance to those of dsDECT in discriminating UA from non-UA stones in a phantom model.

Associations between Tumor Vascularity, Vascular Endothelial Growth Factor Expression and PET/MRI Radiomic Signatures in Primary Clear-Cell-Renal-Cell-Carcinoma: Proof-of-Concept Study.

Studies have shown that tumor angiogenesis is an essential process for tumor growth, proliferation and metastasis. Also, tumor angiogenesis is an important prognostic factor of clear cell renal cell carcinoma (ccRCC), as well as a factor in guiding treatment with antiangiogenic agents. Here, we attempted to find the associations between tumor angiogenesis and radiomic imaging features from PET/MRI. Specifically, sparse canonical correlation analysis was conducted on 3 feature datasets (i.e., radiomic imaging features, tumor microvascular density (MVD), and vascular endothelial growth factor (VEGF) expression) from 9 patients with primary ccRCC. In order to overcome the potential bias of intratumoral heterogeneity of angiogenesis, this study investigated the relationship between regional expressions of angiogenesis and VEGF, and localized radiomic features from different parts within the tumors. Our study highlighted the significant strong correlations between radiomic features and MVD, and also demonstrated that the spatiotemporal features extracted from DCE-MRI provided stronger radiomic correlation to MVD than the textural features extracted from Dixon sequences and FDG PET. Furthermore, PET/MRI, which takes advantage of the combined functional and structural information, had higher radiomics correlation to MVD than solely utilizing PET or MRI alone.

Detection of Upper Tract Urothelial Malignancies by Computed Tomography Urography in Patients Referred for Hematuria at a Large Tertiary Referral Center.

OBJECTIVE: To evaluate the age-stratified prevalence of upper tract urothelial malignancies diagnosed on computed tomography urography in a large cohort of patients referred for initial evaluation of hematuria. MATERIALS AND METHODS: A total of 1123 consecutive adults without a history of urothelial cancer underwent initial computed tomography urography for gross hematuria (n = 652), microscopic hematuria (n = 457), or unspecified hematuria (n = 14) at a single institution from October 2006 to October 2012. Imaging findings suggestive of urothelial lesions were correlated with clinical information, including cystoscopy, cytology, and surgical pathology reports. Patients subsequently diagnosed with urothelial cancer following a normal radiographic evaluation were identified and analyzed. Age, gender, smoking history, and location and type of malignancy were analyzed. RESULTS: Upper tract urothelial cancer was detected in 4 (0.36%) patients, with a mean age of 66.5 years. All 4 patients presented with gross hematuria and were current or former smokers. None of the 535 patients under age 55 who underwent computed tomography urography were diagnosed with upper tract disease regardless of age, smoking history, or degree of hematuria. Likewise, no upper tract cancers were detected in patients referred for microscopic hematuria, regardless of age. CONCLUSION: Detection of upper tract urothelial cancer by computed tomography urography is exceedingly rare in patients presenting at a tertiary referral center with hematuria, particularly in the lower risk strata (younger age, microscopic hematuria). Further investigation into risk-stratified approaches to imaging for hematuria workup is warranted to minimize unnecessary costs and radiation exposure.

Alternate Metabolic Programs Define Regional Variation of Relevant Biological Features in Renal Cell Carcinoma Progression.

PURPOSE: Clear cell renal cell carcinoma (ccRCC) has recently been redefined as a highly heterogeneous disease. In addition to genetic heterogeneity, the tumor displays risk variability for developing metastatic disease, therefore underscoring the urgent need for tissue-based prognostic strategies applicable to the clinical setting. We have recently employed the novel PET/magnetic resonance (MR) image modality to enrich our understanding of how tumor heterogeneity can relate to gene expression and tumor biology to assist in defining individualized treatment plans. EXPERIMENTAL DESIGN: ccRCC patients underwent PET/MR imaging, and these images subsequently used to identify areas of varied intensity for sampling. Samples from 8 patients were subjected to histologic, immunohistochemical, and microarray analysis. RESULTS: Tumor subsamples displayed a range of heterogeneity for common features of hypoxia-inducible factor expression and microvessel density, as well as for features closely linked to metabolic processes, such as GLUT1 and FBP1. In addition, gene signatures linked with disease risk (ccA and ccB) also demonstrated variable heterogeneity, with most tumors displaying a dominant panel of features across the sampled regions. Intriguingly, the ccA- and ccB-classified samples corresponded with metabolic features and functional imaging levels. These correlations further linked a variety of metabolic pathways (i.e., the pentose phosphate and mTOR pathways) with the more aggressive, and glucose avid ccB subtype. CONCLUSIONS: Higher tumor dependency on exogenous glucose accompanies the development of features associated with the poor risk ccB subgroup. Linking these panels of features may provide the opportunity to create functional maps to enable enhanced visualization of the heterogeneous biologic processes of an individual's disease. Clin Cancer Res; 22(12); 2950-9. ©2016 AACR.