Interobserver Agreement and Performance of Concurrent AI Assistance for Radiographic Evaluation of Knee Osteoarthritis.

Background Due to conflicting findings in the literature, there are concerns about a lack of objectivity in grading knee osteoarthritis (KOA) on radiographs. Purpose To examine how artificial intelligence (AI) assistance affects the performance and interobserver agreement of radiologists and orthopedists of various experience levels when evaluating KOA on radiographs according to the established Kellgren-Lawrence (KL) grading system. Materials and Methods In this retrospective observer performance study, consecutive standing knee radiographs from patients with suspected KOA were collected from three participating European centers between April 2019 and May 2022. Each center recruited four readers across radiology and orthopedic surgery at in-training and board-certified experience levels. KL grading (KL-0 = no KOA, KL-4 = severe KOA) on the frontal view was assessed by readers with and without assistance from a commercial AI tool. The majority vote of three musculoskeletal radiology consultants established the reference standard. The ordinal receiver operating characteristic method was used to estimate grading performance. Light kappa was used to estimate interrater agreement, and bootstrapped t statistics were used to compare groups. Results Seventy-five studies were included from each center, totaling 225 studies (mean patient age, 55 years ± 15 [SD]; 113 female patients). The KL grades were KL-0, 24.0% (n = 54); KL-1, 28.0% (n = 63); KL-2, 21.8% (n = 49); KL-3, 18.7% (n = 42); and KL-4, 7.6% (n = 17). Eleven readers completed their readings. Three of the six junior readers showed higher KL grading performance with versus without AI assistance (area under the receiver operating characteristic curve, 0.81 ± 0.017 [SEM] vs 0.88 ± 0.011 [P < .001]; 0.76 ± 0.018 vs 0.86 ± 0.013 [P < .001]; and 0.89 ± 0.011 vs 0.91 ± 0.009 [P = .008]). Interobserver agreement for KL grading among all readers was higher with versus without AI assistance (κ = 0.77 ± 0.018 [SEM] vs 0.85 ± 0.013; P < .001). Board-certified radiologists achieved almost perfect agreement for KL grading when assisted by AI (κ = 0.90 ± 0.01), which was higher than that achieved by the reference readers independently (κ = 0.84 ± 0.017; P = .01). Conclusion AI assistance increased junior readers' radiographic KOA grading performance and increased interobserver agreement for osteoarthritis grading across all readers and experience levels. Published under a CC BY 4.0 license. Supplemental material is available for this article.

Constructing a clinical radiographic knee osteoarthritis database using artificial intelligence tools with limited human labor: A proof of principle.

OBJECTIVE: To create a scalable and feasible retrospective consecutive knee osteoarthritis (OA) radiographic database with limited human labor using commercial and custom-built artificial intelligence (AI) tools. METHODS: We applied four AI tools, two commercially available and two custom-built tools, to analyze 6 years of clinical consecutive knee radiographs from patients aged 35-79 at the University of Copenhagen Hospital, Bispebjerg-Frederiksberg Hospital, Denmark. The tools provided Kellgren-Lawrence (KL) grades, joint space widths, patella osteophyte detection, radiographic view detection, knee joint implant detection, and radiographic marker detection. RESULTS: In total, 25,778 knee radiographs from 8575 patients were included in the database after excluding inapplicable radiographs, and 92.5% of the knees had a complete OA dataset. Using the four AI tools, we saved about 800 hours of radiologist reading time and only manually reviewed 16.0% of the images in the database. CONCLUSIONS: This study shows that clinical knee OA databases can be built using AI with limited human reading time for uniform grading and measurements. The concept is scalable temporally and across geographic regions and could help diversify further OA research by efficiently including radiographic knee OA data from different populations globally. We can prevent data dredging and overfitting OA theories on existing trite cohorts by including various gene pools and continuous expansion of new clinical cohorts. Furthermore, the suggested tools and applied approaches provide an ability to retest previous hypotheses and test new hypotheses on real-life clinical data with current disease prevalence and trends.

Commercially Available Chest Radiograph AI Tools for Detecting Airspace Disease, Pneumothorax, and Pleural Effusion.

Background Commercially available artificial intelligence (AI) tools can assist radiologists in interpreting chest radiographs, but their real-life diagnostic accuracy remains unclear. Purpose To evaluate the diagnostic accuracy of four commercially available AI tools for detection of airspace disease, pneumothorax, and pleural effusion on chest radiographs. Materials and Methods This retrospective study included consecutive adult patients who underwent chest radiography at one of four Danish hospitals in January 2020. Two thoracic radiologists (or three, in cases of disagreement) who had access to all previous and future imaging labeled chest radiographs independently for the reference standard. Area under the receiver operating characteristic curve, sensitivity, and specificity were calculated. Sensitivity and specificity were additionally stratified according to the severity of findings, number of findings on chest radiographs, and radiographic projection. The χ2 and McNemar tests were used for comparisons. Results The data set comprised 2040 patients (median age, 72 years [IQR, 58-81 years]; 1033 female), of whom 669 (32.8%) had target findings. The AI tools demonstrated areas under the receiver operating characteristic curve ranging 0.83-0.88 for airspace disease, 0.89-0.97 for pneumothorax, and 0.94-0.97 for pleural effusion. Sensitivities ranged 72%-91% for airspace disease, 63%-90% for pneumothorax, and 62%-95% for pleural effusion. Negative predictive values ranged 92%-100% for all target findings. In airspace disease, pneumothorax, and pleural effusion, specificity was high for chest radiographs with normal or single findings (range, 85%-96%, 99%-100%, and 95%-100%, respectively) and markedly lower for chest radiographs with four or more findings (range, 27%-69%, 96%-99%, 65%-92%, respectively) (P < .001). AI sensitivity was lower for vague airspace disease (range, 33%-61%) and small pneumothorax or pleural effusion (range, 9%-94%) compared with larger findings (range, 81%-100%; P value range, > .99 to < .001). Conclusion Current-generation AI tools showed moderate to high sensitivity for detecting airspace disease, pneumothorax, and pleural effusion on chest radiographs. However, they produced more false-positive findings than radiology reports, and their performance decreased for smaller-sized target findings and when multiple findings were present. © RSNA, 2023 Supplemental material is available for this article. See also the editorial by Yanagawa and Tomiyama in this issue.

Diagnostic test accuracy study of a commercially available deep learning algorithm for ischemic lesion detection on brain MRIs in suspected stroke patients from a non-comprehensive stroke center.

PURPOSE: To estimate the ability of a commercially available artificial intelligence (AI) tool to detect acute brain ischemia on Magnetic Resonance Imaging (MRI), compared to an experienced neuroradiologist. METHODS: We retrospectively included 1030 patients with brain MRI, suspected of stroke from January 6th, 2020 to 1st of April 2022, based on these criteria: Age ≥ 18 years, symptoms within four weeks before the scan. The neuroradiologist reinterpreted the MRI scans and subclassified ischemic lesions for reference. We excluded scans with interpretation difficulties due to artifacts or missing sequences. Four MRI scanner models from the same vendor were used. The first 800 patients were included consecutively, remaining enriched for less frequent lesions. The index test was a CE-approved AI tool (Apollo version 2.1.1 by Cerebriu). RESULTS: The final analysis cohort comprised 995 patients (mean age 69 years, 53 % female). A case-based analysis for detecting acute ischemic lesions showed a sensitivity of 89 % (95 % CI: 85 %-91 %) and specificity of 90 % (95 % CI: 87 %-92 %). We found no significant difference in sensitivity or specificity based on sex, age, or comorbidities. Specificity was reduced in cases with DWI artifacts. Multivariate analysis showed that increasing ischemic lesion size and fragmented lesions were independently associated with higher sensitivity, while non-acute lesion ages lowered sensitivity. CONCLUSIONS: The AI tool exhibits high sensitivity and specificity in detecting acute ischemic lesions on MRI compared to an experienced neuroradiologist. While sensitivity depends on the ischemic lesions' characteristics, specificity depends on the image quality.

Artificial Intelligence based detection of pneumoperitoneum on CT scans in patients presenting with acute abdominal pain: A clinical diagnostic test accuracy study.

PURPOSE: The primary aim was to investigate the diagnostic performance of an Artificial Intelligence (AI) algorithm for pneumoperitoneum detection in patients with acute abdominal pain who underwent an abdominal CT scan. METHOD: This retrospective diagnostic test accuracy study used a consecutive patient cohort from the Acute High-risk Abdominal patient population at Herlev and Gentofte Hospital, Denmark between January 1, 2019 and September 25, 2019. As reference standard, all studies were rated for pneumoperitoneum (subgroups: none, small, medium, and large amounts) by a gastrointestinal radiology consultant. The index test was a novel AI algorithm based on a sliding window approach with a deep recurrent neural network at its core. The primary outcome was the area under the curve (AUC) of the receiver operating characteristic (ROC). RESULTS: Of 331 included patients (median age 68 years (Range 19-100; 180 women)) 31 patients (9%) had pneumoperitoneum (large: 16, moderate: 7, small: 8). The AUC was 0.77 (95% CI 0.66-0.87). At a specificity of 99% (297/300, 95% CI: 97-100%), sensitivity was 52% (16/31, 95% CI 29-65%), and positive likelihood ratio was 52 (95% CI 16-165). When excluding cases with smaller amounts of free air (<0.25 mL) the AUC increased to 0.96 (95% CI 0.89-1.0). At 99% specificity, sensitivity was 81% (13/16) and positive likelihood ratio was 82 (95% CI 27 - 254). CONCLUSIONS: An AI algorithm identified pneumoperitoneum on CT scans in a clinical setting with low sensitivity but very high specificity, supporting its role for ruling in pneumoperitoneum.