Personalized Impression Generation for PET Reports Using Large Language Models

Large language models (LLMs) have shown promise in accelerating radiology reporting by summarizing clinical findings into impressions. However, automatic impression generation for whole-body PET reports presents unique challenges and has received little attention. Our study aimed to evaluate whether LLMs can create clinically useful impressions for PET reporting. To this end, we fine-tuned twelve open-source language models on a corpus of 37,370 retrospective PET reports collected from our institution. All models were trained using the teacher-forcing algorithm, with the report findings and patient information as input and the original clinical impressions as reference. An extra input token encoded the reading physician's identity, allowing models to learn physician-specific reporting styles. To compare the performances of different models, we computed various automatic evaluation metrics and benchmarked them against physician preferences, ultimately selecting PEGASUS as the top LLM. To evaluate its clinical utility, three nuclear medicine physicians assessed the PEGASUS-generated impressions and original clinical impressions across 6 quality dimensions (3-point scales) and an overall utility score (5-point scale). Each physician reviewed 12 of their own reports and 12 reports from other physicians. When physicians assessed LLM impressions generated in their own style, 89% were considered clinically acceptable, with a mean utility score of 4.08/5. On average, physicians rated these personalized impressions as comparable in overall utility to the impressions dictated by other physicians (4.03, P = 0.41). In summary, our study demonstrated that personalized impressions generated by PEGASUS were clinically useful in most cases, highlighting its potential to expedite PET reporting by automatically drafting impressions.

Automatic Personalized Impression Generation for PET Reports Using Large Language Models

Purpose

To determine if fine-tuned large language models (LLMs) can generate accurate, personalized impressions for whole-body PET reports.

Materials and Methods

Twelve language models were trained on a corpus of PET reports using the teacher-forcing algorithm, with the report findings as input and the clinical impressions as reference. An extra input token encodes the reading physician's identity, allowing models to learn physician-specific reporting styles. Our corpus comprised 37,370 retrospective PET reports collected from our institution between 2010 and 2022. To identify the best LLM, 30 evaluation metrics were benchmarked against quality scores from two nuclear medicine (NM) physicians, with the most aligned metrics selecting the model for expert evaluation. In a subset of data, model-generated impressions and original clinical impressions were assessed by three NM physicians according to 6 quality dimensions (3-point scale) and an overall utility score (5-point scale). Each physician reviewed 12 of their own reports and 12 reports from other physicians. Bootstrap resampling was used for statistical analysis.

Results

Of all evaluation metrics, domain-adapted BARTScore and PEGASUSScore showed the highest Spearman's ρ correlations (ρ=0.568 and 0.563) with physician preferences. Based on these metrics, the fine-tuned PEGASUS model was selected as the top LLM. When physicians reviewed PEGASUS-generated impressions in their own style, 89% were considered clinically acceptable, with a mean utility score of 4.08 out of 5. Physicians rated these personalized impressions as comparable in overall utility to the impressions dictated by other physicians (4.03, P=0.41).

Conclusion

Personalized impressions generated by PEGASUS were clinically useful, highlighting its potential to expedite PET reporting.

Federated learning enables big data for rare cancer boundary detection

Although machine learning (ML) has shown promise across disciplines, out-of-sample generalizability is concerning. This is currently addressed by sharing multi-site data, but such centralization is challenging/infeasible to scale due to various limitations. Federated ML (FL) provides an alternative paradigm for accurate and generalizable ML, by only sharing numerical model updates. Here we present the largest FL study to-date, involving data from 71 sites across 6 continents, to generate an automatic tumor boundary detector for the rare disease of glioblastoma, reporting the largest such dataset in the literature (n = 6, 314). We demonstrate a 33% delineation improvement for the surgically targetable tumor, and 23% for the complete tumor extent, over a publicly trained model. We anticipate our study to: 1) enable more healthcare studies informed by large diverse data, ensuring meaningful results for rare diseases and underrepresented populations, 2) facilitate further analyses for glioblastoma by releasing our consensus model, and 3) demonstrate the FL effectiveness at such scale and task-complexity as a paradigm shift for multi-site collaborations, alleviating the need for data-sharing.

Development and validation of a prediction model for estimating one-month mortality of adult COVID-19 patients presenting at emergency department with suspected pneumonia: a multicenter analysis

There are only a few models developed for risk-stratifying COVID-19 patients with suspected pneumonia in the emergency department (ED). We aimed to develop and validate a model, the COVID-19 ED pneumonia mortality index (CoV-ED-PMI), for predicting mortality in this population. We retrospectively included adult COVID-19 patients who visited EDs of five study hospitals in Texas and who were diagnosed with suspected pneumonia between March and November 2020. The primary outcome was 1-month mortality after the index ED visit. In the derivation cohort, multivariable logistic regression was used to develop the CoV-ED-PMI model. In the chronologically split validation cohort, the discriminative performance of the CoV-ED-PMI was assessed by the area under the receiver operating characteristic curve (AUC) and compared with other existing models. A total of 1678 adult ED records were included for analysis. Of them, 180 patients sustained 1-month mortality. There were 1174 and 504 patients in the derivation and validation cohorts, respectively. Age, body mass index, chronic kidney disease, congestive heart failure, hepatitis, history of transplant, neutrophil-to-lymphocyte ratio, lactate dehydrogenase, and national early warning score were included in the CoV-ED-PMI. The model was validated with good discriminative performance (AUC: 0.83, 95% confidence interval [CI]: 0.79-0.87), which was significantly better than the CURB-65 (AUC: 0.74, 95% CI: 0.69-0.79, p-value: < 0.001). The CoV-ED-PMI had a good predictive performance for 1-month mortality in COVID-19 patients with suspected pneumonia presenting at ED. This free tool is accessible online, and could be useful for clinical decision-making in the ED.

Clinical Characteristics of Patients Returning to Emergency Department With Initial False-Negative Reverse Transcriptase Polymerase Chain Reaction (RT-PCR)-Based COVID-19 Test

BACKGROUND

The coronavirus disease 2019 (COVID-19) outbreak is an international public health emergency. Early identification of COVID-19 patients with false-negative RT-PCR tests is paramount in the ED to prevent both nosocomial and community transmission. This study aimed to compare clinical characteristics of repeat emergency department (ED) visits among coronavirus disease 2019 (COVID-19) patients with initial false-negative reverse transcriptase-polymerase chain reaction (RT-PCR)-based COVID-19 test.

METHODS

This is a retrospective, multi-center, cohort study conducted at 12 hospitals affiliated with Baylor Scott & White Health system. Patients visiting the EDs of these hospitals between June and August 2020 were screened. Patients tested negative for viral RNA by quantitative RT-PCR in the first ED visit and positive in the second ED visit were included. The primary outcome was the comparison of clinical characteristics between two consecutive ED visits including the clinical symptoms, triage vital signs, laboratory, and chest X-ray (CXR) results.

RESULTS

A total of 88 confirmed COVID-19 patients with initial false-negative RT-PCR COVID-19 test in the ED were included in the final analyses. The mean duration of symptoms in the second ED visit was significantly higher (3.6 ± 0.4 vs. 2.6 ± 0.3 days, p = 0.020). In the first ED visit, lymphocytopenia (35.2%), fever (32.6%), nausea (29.5%), and dyspnea (27.9%) are the most common signs of COVID-19 infection during the window period. There were significant increases in the rate of hypoxia (13.6% vs. 4.6%, p = 0.005), abnormal infiltrate on CXR (59.7% vs. 25.9%, p < 0.001), and aspartate aminotransferase (AST) elevation (26.1% vs. 9.1%, p < 0.001) in the second ED visit.

CONCLUSIONS

Early COVID-19 testing (less than 3 days of symptom duration) could be associated with a false-negative result. In this window period, lymphocytopenia, fever, nausea, and dyspnea are the most common early signs that can potentially be clinical hints for COVID-19 diagnosis.

Abstract TP103: Vertebral Artery Tortuosity And Morphometric Characteristics In Recurrent Cervical Artery Dissection

Background: Cervical Artery Dissection (CeAD) is an important cause of stroke in the young. Data on incidence and associations of recurrent CeAD are lacking. Increased Vertebral Artery Tortuosity Index (VTI) has been reported in patients with CeAD and associated with earlier age of arterial dissection in patients with connective tissue disease. We hypothesized that VTI may be associated with recurrent CeAD.

Methods: We reviewed data from a large single-center registry of CeAD patients enrolled between 2011-2021. CT angiography (CTA) was reviewed for neck length, vertebral artery length (VAL), linear distance from the vertebral artery origin to the vertebrobasilar junction (VOBJ), and VTI [((VAL/VOBJ)–1)*100]. Recurrence was defined as radiographic diagnosis of a new dissection during a follow-up period between initial CTA imaging and July 1, 2021. Incidence rate of recurrent dissection was calculated using Poisson regression. Differences between groups were analyzed using the Kruskal-Wallis rank sum test and Fisher’s exact test.

Results: The cohort included 155 patients: women (56%), mean (SD) age 42 (±10) years, and 116 single (carotid or vertebral) and 39 multiple artery dissections. Eight (5%) had a recurrence with a 10 year incidence rate (95% CI) of 1.23 events (0.53, 2.38) per 100 person-years. The recurrent group had a lower mean VTI than the nonrecurrent group (39.2 vs. 45.8; p = 0.024) and trended towards a lower mean age (34.6 vs. 42.3; p = 0.058). Participants with vertebral artery dissection were younger than those with carotid artery dissection with mean ages of 38.5 and 45.1 respectively (p < 0.001) and did not differ in mean VTI (46.0 vs. 44.9; p = 0.747). Morphometric characteristics of height, neck length, and BMI were not associated with recurrence.

Conclusion: In this single center cohort of patients with CeAD, greater VTI was associated with a decreased risk of recurrent dissection. This finding is hypothesis generating as adults with recurrent dissections may have distinct mechanistic or genetic factors contributing to their risk of dissection. This study is limited by relatively small sample size and the retrospective analysis. Further study of VTI and risk of incident CeAD in larger independent cohorts is warranted.

Predicting opioid-induced oversedation in hospitalised patients: a multicentre observational study

OBJECTIVES

Opioid-induced respiratory depression (OIRD) and oversedation are rare but potentially devastating adverse events in hospitalised patients. We investigated which features predict an individual patient's risk of OIRD or oversedation; and developed a risk stratification tool that can be used to aid point-of-care clinical decision-making.

DESIGN

Retrospective observational study.

SETTING

Twelve acute care hospitals in a large not-for-profit integrated delivery system.

PARTICIPANTS

All inpatients ≥18 years admitted between 1 July 2016 and 30 June 2018 who received an opioid during their stay (163 190 unique hospitalisations).

MAIN OUTCOME MEASURES

The primary outcome was occurrence of sedation or respiratory depression severe enough that emergent reversal with naloxone was required, as determined from medical record review; if naloxone reversal was unsuccessful or if there was no evidence of hypoxic encephalopathy or death due to oversedation, it was not considered an oversedation event.

RESULTS

Age, sex, body mass index, chronic obstructive pulmonary disease, concurrent sedating medication, renal insufficiency, liver insufficiency, opioid naïvety, sleep apnoea and surgery were significantly associated with risk of oversedation. The strongest predictor was concurrent administration of another sedating medication (adjusted HR, 95% CI=3.88, 2.48 to 6.06); the most common such medications were benzodiazepines (29%), antidepressants (22%) and gamma-aminobutyric acid analogue (14.7%). The c-statistic for the final model was 0.755. The 24-point Oversedation Risk Criteria (ORC) score developed from the model stratifies patients as high (>20%, ≥21 points), moderate (11%-20%, 10-20 points) and low risk (≤10%, <10 points).

CONCLUSIONS

The ORC risk score identifies patients at high risk for OIRD or oversedation from routinely collected data, enabling targeted monitoring for early detection and intervention. It can also be applied to preventive strategies-for example, clinical decision support offered when concurrent prescriptions for opioids and other sedating medications are entered that shows how the chosen combination impacts the patient's risk.

Screening Digital Breast Tomosynthesis: Radiation Dose Among Patients With Breast Implants

OBJECTIVE

To compare the mean glandular dose (MGD), cancer detection rate (CDR), and recall rate (RR) among screening examinations of patients with breast implants utilizing various digital breast tomosynthesis (DBT)-based imaging protocols.

METHODS

This IRB-approved retrospective study included 1998 women with breast implants who presented for screening mammography between December 10, 2013, and May 29, 2020. Images were obtained using various protocol combinations of DBT and 2D digital mammography. Data collected included MGD, implant type and position, breast density, BI-RADS final assessment category, CDR, and RR. Statistical analysis utilized type II analysis of variance and the chi-square test.

RESULTS

The highest MGD was observed in the DBT only protocol, while the 2D only protocol had the lowest (10.29 mGy vs 5.88 mGy, respectively). Statistically significant difference in MGD was observed across protocols (P < 0.0001). The highest per-view MGD was among DBT full-field (FF) views in both craniocaudal and mediolateral oblique projections (P < 0.0001). No significant difference was observed in RR among protocols (P = 0.17). The combined 2D (FF only) + DBT implant-displaced (ID) views protocol detected the highest number of cancers (CDR, 7.2 per 1000), but this was not significantly different across protocols (P = 0.48).

CONCLUSION

The combination of 2D FF views and DBT ID views should be considered for women with breast implants in a DBT-based screening practice when aiming to minimize radiation exposure without compromising the sensitivity of cancer detection. Avoidance of DBT FF in this patient population is recommended to minimize radiation dose.

Mortality Variations of COVID-19 from Different Hospital Settings During Different Pandemic Phases: A Multicenter Retrospective Study

Introduction

Diverse coronavirus disease 2019 (COVID-19) mortalities have been reported but focused on identifying susceptible patients at risk of more severe disease or death. This study aims to investigate the mortality variations of COVID-19 from different hospital settings during different pandemic phases.

Methods

We retrospectively included adult (≥18 years) patients who visited emergency departments (ED) of five hospitals in the state of Texas and who were diagnosed with COVID-19 between March–November 2020. The included hospitals were dichotomized into urban and suburban based on their geographic location. The primary outcome was mortality that occurred either during hospital admission or within 30 days after the index ED visit. We used multivariable logistic regression to investigate the associations between independent variables and outcome. Generalized additive models were employed to explore the mortality variation during different pandemic phases.

Results

A total of 1,788 adult patients who tested positive for COVID-19 were included in the study. The median patient age was 54.6 years, and 897 (50%) patients were male. Urban hospitals saw approximately 59.5% of the total patients. A total of 197 patients died after the index ED visit. The analysis indicated visits to the urban hospitals (odds ratio [OR] 2.14, 95% confidence interval [CI], 1.41, 3.23), from March to April (OR 2.04, 95% CI, 1.08, 3.86), and from August to November (OR 2.15, 95% CI, 1.37, 3.38) were positively associated with mortality.

Conclusion

Visits to the urban hospitals were associated with a higher risk of mortality in patients with COVID-19 when compared to visits to the suburban hospitals. The mortality risk rebounded and showed significant difference between urban and suburban hospitals since August 2020. Optimal allocation of medical resources may be necessary to bridge this gap in the foreseeable future.

Federated Learning used for predicting outcomes in SARS-COV-2 patients

'Federated Learning' (FL) is a method to train Artificial Intelligence (AI) models with data from multiple sources while maintaining anonymity of the data thus removing many barriers to data sharing. During the SARS-COV-2 pandemic, 20 institutes collaborated on a healthcare FL study to predict future oxygen requirements of infected patients using inputs of vital signs, laboratory data, and chest x-rays, constituting the "EXAM" (EMR CXR AI Model) model. EXAM achieved an average Area Under the Curve (AUC) of over 0.92, an average improvement of 16%, and a 38% increase in generalisability over local models. The FL paradigm was successfully applied to facilitate a rapid data science collaboration without data exchange, resulting in a model that generalised across heterogeneous, unharmonized datasets. This provided the broader healthcare community with a validated model to respond to COVID-19 challenges, as well as set the stage for broader use of FL in healthcare.

Dinutuximab in adult-onset chemotherapy refractory high-risk neuroblastoma

INTRODUCTION

Neuroblastoma is the most common extracranial solid tumor in pediatrics but is considerably uncommon in adults, with approximately 1 case per 10 million diagnosed per year and is associated with poor prognosis. There are no standard treatment protocols for adult-onset neuroblastomas and there are only a few published case reports on neuroblastoma in adults.

CASE REPORT

We report our treatment experience in a 41-year-old female diagnosed with high-risk, poorly differentiated neuroblastoma.

MANAGEMENT AND OUTCOME

Our patient received two cycles of dinutuximab adapted from the Children's Oncology Group ANBL1221 protocol. The patient experienced pain, neuropathy, pruritus, and infusion reactions which were managed with supportive care. Due to the lack of tumor regression, dinutuximab was omitted from future treatments. Currently, the patient is asymptomatic from her disease and remains off of all therapy and pain medication.

DISCUSSION

While dinutuximab has produced promising outcomes in pediatric patients, it is not without potentially severe adverse effects. Serious reactions of capillary leak syndrome, infusion reactions, pain, and neuropathy have been reported. Clinicians must be cognizant of the treatment-related toxicities associated with dinutuximab therapy, ranging from pain, neuropathy, pruritus, and infusion reactions as explored in this patient case.

Study protocol for a multicentre implementation trial of monotherapy anticoagulation to expedite home treatment of patients diagnosed with venous thromboembolism in the emergency department

INTRODUCTION

In the USA, many emergency departments (EDs) have established protocols to treat patients with newly diagnosed deep vein thrombosis (DVT) as outpatients. Similar treatment of patients with pulmonary embolism (PE) has been proposed, but no large-scale study has been published to evaluate a comprehensive, integrated protocol that employs monotherapy anticoagulation to treat patients diagnosed with DVT and PE in the ED.

METHODS AND ANALYSIS

This protocol describes the implementation of the Monotherapy Anticoagulation To expedite Home treatment of Venous ThromboEmbolism (MATH-VTE) study at 33 hospitals in the USA. The study was designed and executed to meet the requirements for the Standards for Reporting Implementation Studies guideline. The study was funded by investigator-initiated awards from industry, with Indiana University as the sponsor. The study principal investigator and study associates travelled to each site to provide on-site training. The protocol identically screens patients with both DVT or PE to determine low risk of death using either the modified Hestia criteria or physician judgement plus a negative result from the simplified PE severity index. Patients must be discharged from the ED within 24 hours of triage and treated with either apixaban or rivaroxaban. Overall effectiveness is based upon the primary efficacy and safety outcomes of recurrent VTE and bleeding requiring hospitalisation respectively. Target enrolment of 1300 patients was estimated with efficacy success defined as the upper limit of the 95% CI for the 30-day frequency of VTE recurrence below 2.0%. Thirty-three hospitals in 17 states were initiated in 2016-2017.

ETHICS AND DISSEMINATION

All sites had Institutional Review Board approval. We anticipate completion of enrolment in June 2020; study data will be available after peer-reviewed publication. MATH-VTE will provide information from a large multicentre sample of US patients about the efficacy and safety of home treatment of VTE with monotherapy anticoagulation.

Abstract 55: Comparison of Sequential Multi-Detector CT and Cone-Beam CT Perfusion Maps in 54 Subjects With an Acute Ischemic Stroke

Introduction: Time from diagnostic imaging to groin puncture highly correlates with outcome and often accounts for delays between hospital arrival and EVT. Our study comparing image quality and information content of MDCTP and CBCTP provides feasibility data for selected AIS patients to go straight to the angio-suite for comprehensive imaging and treatment.

Methods: AIS patients eligible for EVT underwent MDCTP, then a CBCTP study on arrival in angio-suite. Of 939 admitted June 2017-April 2019, 226 (24%) received EVT. Of these 54 (35%) were enrolled to receive additional CBCTP imaging. Inability to obtain consent and co-morbidities were major causes for non-enrollment. Times from the start of MDCTP to angio-suite and from angio-suite arrival to first arterial image were recorded. Acquired CBCTP data were reconstructed and processed with an in-house toolbox. MDCTP and CBCTP data were matched for slice thickness and angulation and were processed using RAPID CTP (iSchemaView, Inc.). The rCBF, rCBV, MTT, tMAX maps were randomized to generate 3 unique evaluation sets. 3 neuroradiologists scored diagnostic image quality, artifacts, mismatch pattern detection and EVT indication using 5-point Likert scales. Stroke laterality was compared with the clinical standard for diagnostic accuracy.

Results: Accuracies for stroke diagnosis are 97% [95%, 97%] with MDCTP and 92% [90%, 95%] with CBCTP. Cohen’s Kappa between observers is 0.90 for MDCTP-based diagnosis and 0.89 for CBCTP-based diagnosis. Scores of CBCTP to make the stroke diagnosis, detect mismatch pattern, and make treatment decision were non-inferior to corresponding scores for MDCTP (alpha=0.05) within 10% of the whole score range. Subjective scores of MDCTP for image quality and artifacts were slightly superior to those of CBCTP (1.8 vs. 2.3, p<0.01).

Conclusions: In this study, a direct to angio-suite workflow provided non-inferior perfusion imaging for AIS patient triage while saving nearly one hour per patient.

Abstract TP76: Quantitative Comparison of Multidetector CT and Cone-Beam CT Perfusion Maps in Large Vessel Occlusion Stroke Patients Undergoing Mechanical Thrombectomy

Cerebral perfusion evaluation using CT or MR perfusion is the gold standard modality to select large vessel occlusion (LVO) stroke patients presenting >6 hours from symptom onset. The availability of cone beam C-arm CT perfusion (CBCTP) in angiography suites could reduce time to endovascular revascularization. We aimed to evaluate the reliability of using CBCTP when compared to multidetector CT perfusion (MDCTP). In this prospective, single-arm, interventional study, 14 LVO anterior circulation thrombectomy patients underwent both a 128 slice MDCTP in the ED and a CBCTP <30 minutes apart prior to groin puncture. CBCTP was acquired using a prototype acquisition mode enabling 10 consecutive C-Arm rotations with nearly continuous data acquisition. A total of 60 cc of contrast layered with 60 cc of saline were injected covering arterial inflow, parenchymal phase and venous outflow. Image data was reconstructed into CBF, CBV, MTT and TTP maps. Three types of measurements were used to compare modalities. In measurement 1, 6 circular regions of interest (ROI) (400mm 2 ) were placed in the anterior arterial territory. In measurement 2, circular ROIs were placed in the ASPECTS regions (cortical 300mm 2 , subcortical 200mm 2 ). In measurement 3, a ROI was drawn around the entire affected area. All ROIs were placed in the basal ganglia and supraganglionic level of both brain sides. Rates (unaffected/affected area) between MDCTP and CBCTP were compared for all sequences. The intraclass correlation coefficient (ICC) was calculated using a single rater, consistency, two-way random-effects model. Measurement 1 found a moderate degree of agreement between MDCTP and CBCTP in CBF, CBV, MTT and TTP rates with ICCs of 0.58 (CI 0.42 - 0.69), 0.65 (CI 0.53 - 0.74), 0.77 (CI 0.68 - 0.83) and 0.52 (CI 0.35 - 0.65). In measurement 2, moderate agreement was found in CBF, CBV and MTT rates; with ICCs of 0.51 (CI 0.32 - 0.65), 0.57 (CI 0.4 - 0.69) and 0.62 (CI 0.47 - 0.73). The results of measurement 3 found an excellent (ICC=0.95, CI 0.88 - 0.98), good (ICC=0.83, CI 0.62 - 0.9) and moderate (ICC=0.7, CI 0.34 - 0.87), degree of agreement in the CBV, MTT and CBF rates, respectively. These results demonstrate promising accuracy of CBCTP in the evaluating ischemic tissue in patient presenting with LVO acute stroke.

Computed Tomography Imaging in Aortic Dissection

Emergency physicians often rely on a “triple-rule-out” computed tomography (CT) where image acquisition is timed to obtain image quality equivalent to dedicated coronary CT angiography, pulmonary CT angiography, and thoracic aorta CT angiography. This case highlights the importance of obtaining CT angiography dedicated to the aorta in the setting of high clinical suspicion for aortic disease if initial CT pulmonary angiogram is negative for aortic disease.

Analysis of central venous catheter utilization at a quaternary care hospital

Central line-associated bloodstream infections (CLABSIs) are one of the most dangerous and costly types of hospital-acquired infections. Incidence of CLABSI can be significantly reduced through proper aseptic techniques, surveillance, and active management strategies, including elimination of idle central line days. This quality improvement project examined two central venous catheter (CVC) cohorts. The institutional electronic health record (EHR) was utilized to generate a daily report indicating CVC utilization by patient care unit. The EHR was further scrutinized for documentation of appropriate indications for CVC use employing an appropriateness tool developed by the institutional vascular access team. Cohort 1 included 12 National Healthcare Safety Network-reportable units audited on a daily basis over a 4-week time period; cohort 2 included selected National Healthcare Safety Network-nonreportable units audited on a daily basis over a 2-week time period. Central venous catheters that did not meet defined indications as outlined by the institutional vascular access team's data collection checklist were escalated the same day to the unit clinical nurse manager for review and possible removal. The percentage of clinically nonindicated CVCs in cohort 1 fell by 65% over the 4-week period of daily audit and real-time feedback, with similar results noted for cohort 2. In conclusion, real-time audit and feedback regarding appropriate clinical indications for CVC use can result in decreased idle or nonindicated central line days, potentially contributing to decreased CLABSI rates.

Credentialing for robotic lobectomy: what is the learning curve? A retrospective analysis of 272 consecutive cases by a single surgeon

Credentialing processes for surgeons seeking robotic thoracic surgical privileges are not evidence-based, and the learning curve has not been reported. The goal of this study is to review our experience with robotic lobectomies and provide evidence for the development of a more uniform credentialing process. We performed a retrospective review of the first 272 consecutive robotic lobectomies performed between 2011 and 2017 by a single surgeon with prior video-assisted thoracoscopic (VATS) experience. Primary outcomes were operative duration, blood loss, chest tube duration, length of hospital stay, intraoperative complication, and conversion to thoracotomy. The patients were subdivided by surgical date into two cohorts of 120 consecutive patients to compare differences in outcomes, thereby illustrating the learning curve. Between 2011 and 2017, 272 patients (median age 67.5 years) underwent a robotic lobectomy by a single surgeon. The majority of patients (157/272) had early stage (T1N0) adenocarcinoma. For the entire cohort, median operative time was 160 min (83-317 min). The median blood loss was 75 mL (10-4000 mL). Median chest tube duration was 2 days (1-23 days) and median hospital stay was 3 days (1-25 days). Intraoperative complications occurred in seven patients. Only six patients required conversion to thoracotomy. Using multivariable logistic regression, it was found that the age, gender, and stage do not factor into conversion to thoracotomy, but BMI was found to be a significant covariate (p 0.043). As the surgeon performs more surgeries, there is a significantly shorter operative time (p < 0.001), decreased blood loss (p < 0.001), and shorter hospital stay (p < 0.014). When the first 120 and last 120 surgeries were compared, there was significantly less blood loss (234.6 vs 78.69 cc, p < 0.001), shorter operative time (181.9 vs 147.4 min, p < 0.001), shorter tube duration (3.49 vs 3.11 days, p 0.007), and shorter length of stay (4.03 vs 3.48 days, p < 0.001), respectively. More intraoperative complications were observed during the first 120 surgeries (6/120) compared to the last 120 surgeries (0/120; Fischer exact p = 0.029). Regression model plots did not show any apparent and significant change points, but rather a steady improvement. The more cases the surgeon does, the better is the outcome in terms of operative duration, blood loss, post-operative length of stay and intraoperative complications. The learning curve for robotic surgery for a surgeon with prior VATS experience is that of a continuous improvement with experience instead of a particular change point. Since most thoracic surgeons who perform robotic-assisted surgery have already gotten past their VATS learning curves, they no longer have a definable learning curve for robotic surgery. Hence, if a surgeon is already proficient and credentialed to perform VATS lung resections, he or she is no longer faced with a significant learning curve for robotic lung resections, and should be credentialed to do so once he or she has undergone the appropriate training with the equipment and technology.

Design, Construction, and Initial Results of a Prototype Multi-Contrast X-Ray Breast Imaging System

By integrating a grating-based interferometer with a clinical full field digital mammography (FFDM) system, a prototype multi-contrast (absorption, phase, and dark field) x-ray breast imaging system was developed in this work. Unlike previous benchtop-based multi-contrast x-ray imaging systems that usually have relatively long source-to-detector distance and vibration isolators or dampers for the interferometer, the FFDM hardware platform is subject to mechanical vibration and the constraint of compact system geometry. Current grating fabrication technology also imposes additional constraints on the design of the grating interferometer. Based on these technical constraints and the x-ray beam properties of the FFDM system, three gratings were designed and integrated with the FFDM system. When installing the gratings, no additional vibration damping device was used in order to test the robustness of multi-contrast imaging system against mechanical vibration. The measured visibility of the diffraction fringes was 23±3%, and two images acquired 60 minutes apart demonstrated good system reproducibility with no visible signal drift. Preliminary results generated from the prototype system demonstrate the multi-contrast imaging capability of the system. The three contrast mechanisms provide mutually complementary information of the phantom object. This prototype system provides a much needed platform for evaluating the true clinical utility of the multi-contrast x-ray imaging method for the diagnosis of breast cancer.

A Brain Electrical Activity Electroencephalographic-Based Biomarker of Functional Impairment in Traumatic Brain Injury: A Multi-Site Validation Trial

The potential clinical utility of a novel quantitative electroencephalographic (EEG)-based Brain Function Index (BFI) as a measure of the presence and severity of functional brain injury was studied as part of an independent prospective validation trial. The BFI was derived using quantitative EEG (QEEG) features associated with functional brain impairment reflecting current consensus on the physiology of concussive injury. Seven hundred and twenty adult patients (18-85 years of age) evaluated within 72 h of sustaining a closed head injury were enrolled at 11 U.S. emergency departments (EDs). Glasgow Coma Scale (GCS) score was 15 in 97%. Standard clinical evaluations were conducted and 5 to 10 min of EEG acquired from frontal locations. Clinical utility of the BFI was assessed for raw scores and percentile values. A multinomial logistic regression analysis demonstrated that the odds ratios (computed against controls) of the mild and moderate functionally impaired groups were significantly different from the odds ratio of the computed tomography (CT) postive (CT+, structural injury visible on CT) group (p = 0.0009 and p = 0.0026, respectively). However, no significant differences were observed between the odds ratios of the mild and moderately functionally impaired groups. Analysis of variance (ANOVA) demonstrated significant differences in BFI among normal (16.8%), mild TBI (mTBI)/concussed with mild or moderate functional impairment, (61.3%), and CT+ (21.9%) patients (p < 0.0001). Regression slopes of the odds ratios for likelihood of group membership suggest a relationship between the BFI and severity of impairment. Findings support the BFI as a quantitative marker of brain function impairment, which scaled with severity of functional impairment in mTBI patients. When integrated into the clinical assessment, the BFI has the potential to aid in early diagnosis and thereby potential to impact the sequelae of TBI by providing an objective marker that is available at the point of care, hand-held, non-invasive, and rapid to obtain.

Modified ideal observer model (MIOM) for high-contrast and high-spatial resolution CT imaging tasks

PURPOSE

Although a variety of mathematical observer models have been developed to predict human observer performance for low contrast lesion detection tasks, their predictive power for high contrast and high spatial resolution discrimination imaging tasks, including those in CT bone imaging, could be limited. The purpose of this work was to develop a modified observer model that has improved correlation with human observer performance for these tasks.

METHODS

The proposed observer model, referred to as the modified ideal observer model (MIOM), uses a weight function to penalize components in the task function that have less contribution to the actual human observer performance for high contrast and high spatial resolution discrimination tasks. To validate MIOM, both human observer and observer model studies were performed, each using exactly the same CT imaging task [discrimination of a connected component in a high contrast (1000 HU) high spatial resolution bone fracture model (0.3 mm)] and experimental CT image data. For the human observer studies, three physicist observers rated the connectivity of the fracture model using a five-point Likert scale; for the observer model studies, a total of five observer models, including both conventional models and the proposed MIOM, were used to calculate the discrimination capability of the CT images in resolving the connected component. Images used in the studies encompassed nine different reconstruction kernels. Correlation between human and observer model performance for these kernels were quantified using the Spearman rank correlation coefficient (ρ). After the validation study, an example application of MIOM was presented, in which the observer model was used to select the optimal reconstruction kernel for a High-Resolution (Hi-Res, GE Healthcare) CT scan technique.

RESULTS

The performance of the proposed MIOM correlated well with that of the human observers with a Spearman rank correlation coefficient ρ of 0.88 (P = 0.003). In comparison, the value of ρ was 0.05 (P = 0.904) for the ideal observer, 0.05 (P = 0.904) for the non-prewhitening observer, -0.18 (P = 0.634) for the non-prewhitening observer with eye filter and internal noise, and 0.30 (P = 0.427) for the prewhitening observer with eye filter and internal noise. Using the validated MIOM, the optimal reconstruction kernel for the Hi-Res mode to perform high spatial resolution and high contrast discrimination imaging tasks was determined to be the HD Ultra kernel at the center of the scan field of view (SFOV), or the Lung kernel at the peripheral region of the SFOV. This result was consistent with visual observations of nasal CT images of an in vivo canine subject.

CONCLUSION

Compared with other observer models, the proposed modified ideal observer model provides significantly improved correlation with human observers for high contrast and high spatial resolution CT imaging tasks.

Emergency Department Triage of Traumatic Head Injury Using a Brain Electrical Activity Biomarker: A Multisite Prospective Observational Validation Trial

OBJECTIVES

A brain electrical activity biomarker for identifying traumatic brain injury (TBI) in emergency department (ED) patients presenting with high Glasgow Coma Scale (GCS) after sustaining a head injury has shown promise for objective, rapid triage. The main objective of this study was to prospectively evaluate the efficacy of an automated classification algorithm to determine the likelihood of being computed tomography (CT) positive, in high-functioning TBI patients in the acute state.

METHODS

Adult patients admitted to the ED for evaluation within 72 hours of sustaining a closed head injury with GCS 12 to 15 were candidates for study. A total of 720 patients (18-85 years) meeting inclusion/exclusion criteria were enrolled in this observational, prospective validation trial, at 11 U.S. EDs. GCS was 15 in 97%, with the first and third quartiles being 15 (interquartile range = 0) in the study population at the time of the evaluation. Standard clinical evaluations were conducted and 5 to 10 minutes of electroencephalogram (EEG) was acquired from frontal and frontal-temporal scalp locations. Using an a priori derived EEG-based classification algorithm developed on an independent population and applied to this validation population prospectively, the likelihood of each subject being CT+ was determined, and performance metrics were computed relative to adjudicated CT findings.

RESULTS

Sensitivity of the binary classifier (likely CT+ or CT-) was 92.3% (95% confidence interval [CI] = 87.8%-95.5%) for detection of any intracranial injury visible on CT (CT+), with specificity of 51.6% (95% CI = 48.1%-55.1%) and negative predictive value (NPV) of 96.0% (95% CI = 93.2%-97.9%). Using ternary classification (likely CT+, equivocal, likely CT-) demonstrated enhanced sensitivity to traumatic hematomas (≥1 mL of blood), 98.6% (95% CI = 92.6%-100.0%), and NPV of 98.2% (95% CI = 95.5%-99.5%).

CONCLUSION

Using an EEG-based biomarker high accuracy of predicting the likelihood of being CT+ was obtained, with high NPV and sensitivity to any traumatic bleeding and to hematomas. Specificity was significantly higher than standard CT decision rules. The short time to acquire results and the ease of use in the ED environment suggests that EEG-based classifier algorithms have potential to impact triage and clinical management of head-injured patients.

Reduction of Beam Hardening Artifacts in Cone-Beam CT Imaging via SMART-RECON Algorithm

When an automatic exposure control is introduced in C-arm cone beam CT data acquisition, the spectral inconsistencies between acquired projection data are exacerbated. As a result, conventional water/bone correction schemes are not as effective as in conventional diagnostic x-ray CT acquisitions with a fixed tube potential. In this paper, a new method was proposed to reconstruct several images with different degrees of spectral consistency and thus different levels of beam hardening artifacts. The new method relies neither on prior knowledge of the x-ray beam spectrum nor on prior compositional information of the imaging object. Numerical simulations were used to validate the algorithm.

Gastrointestinal Endoscopic Ultrasound-Guided Fine-Needle Aspiration Biopsy Specimens: Adequate Diagnostic Yield and Accuracy Can Be Achieved without On-Site Evaluation

BACKGROUND

Endoscopic ultrasound-guided fine-needle aspiration biopsy (EUS-FNA) is the preferred method for biopsying the gastrointestinal tract, and rapid on-site cytological evaluation is considered standard practice. Our institution does not perform on-site evaluation; this study analyzes our overall diagnostic yield, accuracy, and incidence of nondiagnostic cases to determine the validity of this strategy.

DESIGN

Data encompassing clinical information, procedural records, and cytological assessment were analyzed for gastrointestinal EUS-FNA procedures (n = 85) performed at Vancouver General Hospital from January 2012 to January 2013. We compared our results with those of studies that had on-site evaluation and studies that did not have on-site evaluation.

RESULTS

Eighty-five biopsies were performed in 78 patients, from sites that included the pancreas, the stomach, the duodenum, lymph nodes, and retroperitoneal masses. Malignancies were diagnosed in 45 (53%) biopsies, while 24 (29%) encompassed benign entities. Suspicious and atypical results were recorded in 8 (9%) and 6 (7%) cases, respectively. Only 2 (2%) cases received a cytological diagnosis of 'nondiagnostic'. Our overall accuracy was 72%, our diagnostic yield was 98%, and our nondiagnostic rate was 2%. Our results did not significantly differ from those of studies that did have on-site evaluation.

CONCLUSION

Our study highlights that adequate diagnostic accuracy can be achieved without on-site evaluation.

Grating-based phase contrast tomosynthesis imaging: proof-of-concept experimental studies

PURPOSE

This paper concerns the feasibility of x-ray differential phase contrast (DPC) tomosynthesis imaging using a grating-based DPC benchtop experimental system, which is equipped with a commercial digital flat-panel detector and a medical-grade rotating-anode x-ray tube. An extensive system characterization was performed to quantify its imaging performance.

METHODS

The major components of the benchtop system include a diagnostic x-ray tube with a 1.0 mm nominal focal spot size, a flat-panel detector with 96 μm pixel pitch, a sample stage that rotates within a limited angular span of ± 30°, and a Talbot-Lau interferometer with three x-ray gratings. A total of 21 projection views acquired with 3° increments were used to reconstruct three sets of tomosynthetic image volumes, including the conventional absorption contrast tomosynthesis image volume (AC-tomo) reconstructed using the filtered-backprojection (FBP) algorithm with the ramp kernel, the phase contrast tomosynthesis image volume (PC-tomo) reconstructed using FBP with a Hilbert kernel, and the differential phase contrast tomosynthesis image volume (DPC-tomo) reconstructed using the shift-and-add algorithm. Three inhouse physical phantoms containing tissue-surrogate materials were used to characterize the signal linearity, the signal difference-to-noise ratio (SDNR), the three-dimensional noise power spectrum (3D NPS), and the through-plane artifact spread function (ASF).

RESULTS

While DPC-tomo highlights edges and interfaces in the image object, PC-tomo removes the differential nature of the DPC projection data and its pixel values are linearly related to the decrement of the real part of the x-ray refractive index. The SDNR values of polyoxymethylene in water and polystyrene in oil are 1.5 and 1.0, respectively, in AC-tomo, and the values were improved to 3.0 and 2.0, respectively, in PC-tomo. PC-tomo and AC-tomo demonstrate equivalent ASF, but their noise characteristics quantified by the 3D NPS were found to be different due to the difference in the tomosynthesis image reconstruction algorithms.

CONCLUSIONS

It is feasible to simultaneously generate x-ray differential phase contrast, phase contrast, and absorption contrast tomosynthesis images using a grating-based data acquisition setup. The method shows promise in improving the visibility of several low-density materials and therefore merits further investigation.

Correlation between human observer performance and model observer performance in differential phase contrast CT

PURPOSE

With the recently expanding interest and developments in x-ray differential phase contrast CT (DPC-CT), the evaluation of its task-specific detection performance and comparison with the corresponding absorption CT under a given radiation dose constraint become increasingly important. Mathematical model observers are often used to quantify the performance of imaging systems, but their correlations with actual human observers need to be confirmed for each new imaging method. This work is an investigation of the effects of stochastic DPC-CT noise on the correlation of detection performance between model and human observers with signal-known-exactly (SKE) detection tasks.

METHODS

The detectabilities of different objects (five disks with different diameters and two breast lesion masses) embedded in an experimental DPC-CT noise background were assessed using both model and human observers. The detectability of the disk and lesion signals was then measured using five types of model observers including the prewhitening ideal observer, the nonprewhitening (NPW) observer, the nonprewhitening observer with eye filter and internal noise (NPWEi), the prewhitening observer with eye filter and internal noise (PWEi), and the channelized Hotelling observer (CHO). The same objects were also evaluated by four human observers using the two-alternative forced choice method. The results from the model observer experiment were quantitatively compared to the human observer results to assess the correlation between the two techniques.

RESULTS

The contrast-to-detail (CD) curve generated by the human observers for the disk-detection experiments shows that the required contrast to detect a disk is inversely proportional to the square root of the disk size. Based on the CD curves, the ideal and NPW observers tend to systematically overestimate the performance of the human observers. The NPWEi and PWEi observers did not predict human performance well either, as the slopes of their CD curves tended to be steeper. The CHO generated the best quantitative agreement with human observers with its CD curve overlapping with that of human observer. Statistical equivalence between CHO and humans can be claimed within 11% of the human observer results, including both the disk and lesion detection experiments.

CONCLUSIONS

The model observer method can be used to accurately represent human observer performance with the stochastic DPC-CT noise for SKE tasks with sizes ranging from 8 to 128 pixels. The incorporation of the anatomical noise remains to be studied.

Statistical model based iterative reconstruction (MBIR) in clinical CT systems. Part II. Experimental assessment of spatial resolution performance

PURPOSE

Statistical model based iterative reconstruction (MBIR) methods have been introduced to clinical CT systems and are being used in some clinical diagnostic applications. The purpose of this paper is to experimentally assess the unique spatial resolution characteristics of this nonlinear reconstruction method and identify its potential impact on the detectabilities and the associated radiation dose levels for specific imaging tasks.

METHODS

The thoracic section of a pediatric phantom was repeatedly scanned 50 or 100 times using a 64-slice clinical CT scanner at four different dose levels [CTDIvol =4, 8, 12, 16 (mGy)]. Both filtered backprojection (FBP) and MBIR (Veo(®), GE Healthcare, Waukesha, WI) were used for image reconstruction and results were compared with one another. Eight test objects in the phantom with contrast levels ranging from 13 to 1710 HU were used to assess spatial resolution. The axial spatial resolution was quantified with the point spread function (PSF), while the z resolution was quantified with the slice sensitivity profile. Both were measured locally on the test objects and in the image domain. The dependence of spatial resolution on contrast and dose levels was studied. The study also features a systematic investigation of the potential trade-off between spatial resolution and locally defined noise and their joint impact on the overall image quality, which was quantified by the image domain-based channelized Hotelling observer (CHO) detectability index d'.

RESULTS

(1) The axial spatial resolution of MBIR depends on both radiation dose level and image contrast level, whereas it is supposedly independent of these two factors in FBP. The axial spatial resolution of MBIR always improved with an increasing radiation dose level and/or contrast level. (2) The axial spatial resolution of MBIR became equivalent to that of FBP at some transitional contrast level, above which MBIR demonstrated superior spatial resolution than FBP (and vice versa); the value of this transitional contrast highly depended on the dose level. (3) The PSFs of MBIR could be approximated as Gaussian functions with reasonably good accuracy. (4) Thez resolution of MBIR showed similar contrast and dose dependence. (5) Noise standard deviation assessed on the edges of objects demonstrated a trade-off with spatial resolution in MBIR. (5) When both spatial resolution and image noise were considered using the CHO analysis, MBIR led to significant improvement in the overall CT image quality for both high and low contrast detection tasks at both standard and low dose levels.

CONCLUSIONS

Due to the intrinsic nonlinearity of the MBIR method, many well-known CT spatial resolution and noise properties have been modified. In particular, dose dependence and contrast dependence have been introduced to the spatial resolution of CT images by MBIR. The method has also introduced some novel noise-resolution trade-off not seen in traditional CT images. While the benefits of MBIR regarding the overall image quality, as demonstrated in this work, are significant, the optimal use of this method in clinical practice demands a thorough understanding of its unique physical characteristics.

Grating based x-ray differential phase contrast imaging without mechanical phase stepping

Grating-based x-ray differential phase contrast imaging (DPCI) often uses a phase stepping procedure to acquire data that enables the extraction of phase information. This method prolongs the time needed for data acquisition by several times compared with conventional x-ray absorption image acquisitions. A novel analyzer grating design was developed in this work to eliminate the additional data acquisition time needed to perform phase stepping in DPCI. The new analyzer grating was fabricated such that the linear grating structures are shifted from one detector row to the next; the amount of the lateral shift was equal to a fraction of the x-ray diffraction fringe pattern. The x-ray data from several neighboring detector rows were then combined to extract differential phase information. Initial experimental results have demonstrated that the new analyzer grating enables accurate DPCI signal acquisition from a single x-ray exposure like conventional x-ray absorption imaging.