Bridging the Gap between Structured and Free-form Radiology Reporting: A Case-study on Coronary CT Angiography

Free-form radiology reports associated with coronary computed tomography angiography (CCTA) include nuanced and complicated linguistics to report cardiovascular disease. Standardization and interpretation of such reports is crucial for clinical use of CCTA. Coronary Artery Disease Reporting and Data System (CAD-RADS) has been proposed to achieve such standardization by implementing a strict template-based report writing and assignment of a score between 0 and 5 indicating the severity of coronary artery lesions. Even after its introduction, free-form unstructured report writing remains popular among radiologists. In this work, we present our attempts at bridging the gap between structured and unstructured reporting by natural language processing. We present machine learning models that while being trained only on structured reports, can predict CAD-RADS scores by analysis of free-text of unstructured radiology reports. The best model achieves 98% accuracy on structured reports and 92% 1-margin accuracy (difference of \le 1 in the predicted and the actual scores) for free-form unstructured reports. Our model also performs well under very difficult circumstances including nuanced and widely varying terminology used for reporting cardiovascular functions and diseases, scarcity of labeled data for training our model, and uneven class label distribution.

Characterizing the Progression of Pulmonary Edema Severity: Can Pairwise Comparisons in Radiology Reports Help?

Accurately assessing the severity of pulmonary edema is critical for making treatment decisions in congestive heart failure patients. However, the current scale for quantifying pulmonary edema based on chest radiographs does not have well-characterized severity levels, with substantial inter-radiologist disagreement. In this study, we investigate whether comparisons documented in radiology reports can provide accurate characterizations of pulmonary edema progression. We propose a rules-based natural language processing approach to assess the change in a patient's pulmonary edema status (e.g. better, worse, no change) by performing pairwise comparisons of consecutive radiology reports, using regular expressions and heuristics derived from clinical knowledge. Evaluated against ground-truth labels from expert radiologists, our labeler extracts comparisons describing the progression of pulmonary edema with 0.875 precision and 0.891 recall. We also demonstrate the potential utility of comparison labels in providing additional fine-grained information over noisier labels produced by models that directly estimate severity level.

An automated COVID-19 triage pipeline using artificial intelligence based on chest radiographs and clinical data

While COVID-19 diagnosis and prognosis artificial intelligence models exist, very few can be implemented for practical use given their high risk of bias. We aimed to develop a diagnosis model that addresses notable shortcomings of prior studies, integrating it into a fully automated triage pipeline that examines chest radiographs for the presence, severity, and progression of COVID-19 pneumonia. Scans were collected using the DICOM Image Analysis and Archive, a system that communicates with a hospital’s image repository. The authors collected over 6,500 non-public chest X-rays comprising diverse COVID-19 severities, along with radiology reports and RT-PCR data. The authors provisioned one internally held-out and two external test sets to assess model generalizability and compare performance to traditional radiologist interpretation. The pipeline was evaluated on a prospective cohort of 80 radiographs, reporting a 95% diagnostic accuracy. The study mitigates bias in AI model development and demonstrates the value of an end-to-end COVID-19 triage platform.

Evaluating discrepancy rates of radiology resident provisional reports for cross-sectional body imaging studies at a tertiary hospital

Introduction Urgent radiological studies obtained during on-call hours are often preliminarily read by on-call residents before consultant radiologists finalise the reports at a later time. Such provisional radiology reports provide important information to guide initial patient management. This study aims to determine discrepancy rates between provisional reports and final interpretations, and to assess the clinical significance of such discrepancies. Methods This retrospective quality assurance project reviewed a total of 1218 cross-sectional imaging studies of the body (thorax, abdomen and pelvis) done between July 2015 and May 2016 during on-call hours. The studies included 1201 Computed tomography (CT) scans and 17 Magnetic Resonance Imaging (MRI) scans. Studies with incomplete or unavailable reports were excluded. Conclusions of both the provisional and final reports of each study were reviewed for concordance, with reference to the full report if needed. Discrepancies were graded according to the ACR 2016 RADPEER scoring system. Results There were 1210 studies with complete reports. Discrepant reports were noted in 183 (15.1%) studies. Of these, 89 (7.3%) were assessed to be clinically significant and the majority of these (55) were due to interpretations which should be made most of the time. CT of the abdomen and pelvis were the most prone to discrepant reports, accounting for 148 cases (80.9%). Conclusion The majority of preliminary reports for on-call body scans were concordant with final interpretations. The discrepancy rates for provisional body scan reports provided by residents while on call were comparable to those previously reported in literature.

Surveillance of atypical femoral fractures in a nationwide fracture register

Background and purpose — To continuously assess the incidence of atypical femoral fractures (AFFs) in the population is important, to allow the evaluation of the risks and benefits associated with osteoporosis treatment. Therefore, we investigated the possibility to use the Swedish Fracture Register (SFR) as a surveillance tool for AFFs in the population and to explore means of improvement. Patients and methods — All AFF registrations in the SFR from January 1, 2015 to December 31, 2018 were enrolled in the study. For these patients, radiographs were obtained and combined with radiographs from 176 patients with normal femoral fractures, to form the study cohort. All images were reviewed and classified into AFFs or normal femur fractures by 2 experts in the field (gold-standard classification) and 1 orthopedic resident educated on the specific radiographic features of AFF (educated-user classification). Furthermore, we estimated the incidence rate of AFFs in the population captured by the register through comparison with a previous cohort and calculated the positive predictive value (PPV) and, where possible, the inter-observer agreement (Cohen’s kappa) between the different classifications. Results — Of the 178 available patients with AFF in the SFR, 104 patients were classified as AFF using the goldstandard classification, and 89 using the educated-user classification. The PPV increased from 0.58 in the SFR classification to 0.93 in the educated-user classification. The interobserver agreement between the gold-standard classification and the educated-user classification was 0.81. Interpretation — With a positive predictive value of 0.58 the Swedish Fracture Register outperforms radiology reports and reports to the Swedish Medical Products Agency on adverse drug reactions as a diagnostic tool to identify atypical femoral fractures.

Attention based automated radiology report generation using CNN and LSTM

The automated generation of radiology reports provides X-rays and has tremendous potential to enhance the clinical diagnosis of diseases in patients. A new research direction is gaining increasing attention that involves the use of hybrid approaches based on natural language processing and computer vision techniques to create auto medical report generation systems. The auto report generator, producing radiology reports, will significantly reduce the burden on doctors and assist them in writing manual reports. Because the sensitivity of chest X-ray (CXR) findings provided by existing techniques not adequately accurate, producing comprehensive explanations for medical photographs remains a difficult task. A novel approach to address this issue was proposed, based on the continuous integration of convolutional neural networks and long short-term memory for detecting diseases, followed by the attention mechanism for sequence generation based on these diseases. Experimental results obtained by using the Indiana University CXR and MIMIC-CXR datasets showed that the proposed model attained the current state-of-the-art efficiency as opposed to other solutions of the baseline. BLEU-1, BLEU-2, BLEU-3, and BLEU-4 were used as the evaluation metrics.

CT-to-fluoroscopy registration versus scan-and-plan registration for robot-assisted insertion of lumbar pedicle screws

OBJECTIVE Pedicle screw insertion for stabilization after lumbar fusion surgery is commonly performed by spine surgeons. With the advent of navigation technology, the accuracy of pedicle screw insertion has increased. Robotic guidance has revolutionized the placement of pedicle screws with 2 distinct radiographic registration methods, the scan-and-plan method and CT-to-fluoroscopy method. In this study, the authors aimed to compare the accuracy and safety of these methods. METHODS A retrospective chart review was conducted at 2 centers to obtain operative data for consecutive patients who underwent robot-assisted lumbar pedicle screw placement. The newest robotic platform (Mazor X Robotic System) was used in all cases. One center used the scan-and-plan registration method, and the other used CT-to-fluoroscopy for registration. Screw accuracy was determined by applying the Gertzbein-Robbins scale. Fluoroscopic exposure times were collected from radiology reports. RESULTS Overall, 268 patients underwent pedicle screw insertion, 126 patients with scan-and-plan registration and 142 with CT-to-fluoroscopy registration. In the scan-and-plan cohort, 450 screws were inserted across 266 spinal levels (mean 1.7 ± 1.1 screws/level), with 446 (99.1%) screws classified as Gertzbein-Robbins grade A (within the pedicle) and 4 (0.9%) as grade B (< 2-mm deviation). In the CT-to-fluoroscopy cohort, 574 screws were inserted across 280 lumbar spinal levels (mean 2.05 ± 1.7 screws/ level), with 563 (98.1%) grade A screws and 11 (1.9%) grade B (p = 0.17). The scan-and-plan cohort had nonsignificantly less fluoroscopic exposure per screw than the CT-to-fluoroscopy cohort (12 ± 13 seconds vs 11.1 ± 7 seconds, p = 0.3). CONCLUSIONS Both scan-and-plan registration and CT-to-fluoroscopy registration methods were safe, accurate, and had similar fluoroscopy time exposure overall.

Natural Language Processing of Large-Scale Structured Radiology Reports to Identify Oncologic Patients With or Without Splenomegaly Over a 10-Year Period

PURPOSE To assess the accuracy of a natural language processing (NLP) model in extracting splenomegaly described in patients with cancer in structured computed tomography radiology reports. METHODS In this retrospective study between July 2009 and April 2019, 3,87,359 consecutive structured radiology reports for computed tomography scans of the chest, abdomen, and pelvis from 91,665 patients spanning 30 types of cancer were included. A randomized sample of 2,022 reports from patients with colorectal cancer, hepatobiliary cancer (HB), leukemia, Hodgkin lymphoma (HL), and non-HL patients was manually annotated as positive or negative for splenomegaly. NLP model training/testing was performed on 1,617/405 reports, and a new validation set of 400 reports from all cancer subtypes was used to test NLP model accuracy, precision, and recall. Overall survival was compared between the patient groups (with and without splenomegaly) using Kaplan-Meier curves. RESULTS The final cohort included 3,87,359 reports from 91,665 patients (mean age 60.8 years; 51.2% women). In the testing set, the model achieved accuracy of 92.1%, precision of 92.2%, and recall of 92.1% for splenomegaly. In the validation set, accuracy, precision, and recall were 93.8%, 92.9%, and 86.7%, respectively. In the entire cohort, splenomegaly was most frequent in patients with leukemia (32.5%), HB (17.4%), non-HL (9.1%), colorectal cancer (8.5%), and HL (5.6%). A splenomegaly label was associated with an increased risk of mortality in the entire cohort (hazard ratio 2.10; 95% CI, 1.98 to 2.22; P < .001). CONCLUSION Automated splenomegaly labeling by NLP of radiology report demonstrates good accuracy, precision, and recall. Splenomegaly is most frequently reported in patients with leukemia, followed by patients with HB.