Artificial Intelligence (AI) based chest X-ray (CXR) screening for tuberculosis (TB) is becoming increasingly popular. Still, deploying such AI tools can be challenging due to multiple real-life barriers like software installation, workflow integration, network connectivity constraints, limited human resources available to interpret findings, etc. To understand these challenges, PATH implemented a TB REACH active case-finding program in a resource-limited setting of Nagpur in India, where an AI software device (qXR) intended for TB screening using CXR images was used. Eight private CXR laboratories that fulfilled prerequisites for AI software installation were engaged for this program. Key lessons about operational feasibility and accessibility, along with the strategies adopted to overcome these challenges, were learned during this program. This program also helped to screen 10,481 presumptive TB individuals using informal providers based on clinical history. Among them, 2,303 individuals were flagged as presumptive for TB by a radiologist or by AI based on their CXR interpretation. Approximately 15.8% increase in overall TB yield could be attributed to the presence of AI alone because these additional cases were not deemed presumptive for TB by radiologists, but AI was able to identify them. Successful implementation of AI tools like qXR in resource-limited settings in India will require solving real-life implementation challenges for seamless deployment and workflow integration.

Missed nodules in chest radiographs (CXRs) are common occurrences. We assessed the effect of artificial intelligence (AI) as a second reader on the accuracy of radiologists and non-radiology physicians in lung nodule detection and localization in CXRs. This retrospective study using the multi-reader multi-case design included 300 CXRs acquired from 40 hospitals across the US. All CXRs had a paired follow-up image (chest CT or CXR) to augment the ground truth establishment for the presence and location of nodules on CXRs by five independent thoracic radiologists. 15 readers (nine radiologists and six non-radiology physicians) read each CXR twice in a second-reader paradigm, once without AI and then immediately with AI assistance. The primary analysis assessed the difference in area-under-the-alternative-free-response-receiver-operating-characteristic-curve (AFROC) of readers with and without AI. Case-level area-under-the-receiver-operating-characteristic-curve (AUROC), sensitivity, and specificity were assessed in secondary analyses. A total of 300 CXRs (147 with nodules, 153 without nodules) from 300 patients (mean age, 64 years ± 15 [standard deviation]; 174 women) were included. The mean AFROC of readers was 0.73 without AI and 0.81 with AI (95% CI of difference, 0.05-0.10). Case-level AUROC was 0.77 without AI and 0.84 with AI (95% CI of difference, 0.04-0.09). Case-level sensitivity was 72.8% and 83.5% (95% CI of difference, 6.8-14.6) and specificity was 71.1% and 72.0% (95% CI of difference, -0.8-2.6) without and with AI, respectively. Using AI, readers detected and localized more nodules without any significant difference in false positive interpretations.

In resource-limited settings, timely treatment of acute stroke is dependent upon accurate diagnosis that draws on non-contrast computed tomography (NCCT) scans of the head. Artificial Intelligence (AI) based devices may be able to assist non-specialist physicians in NCCT interpretation, thereby enabling faster interventions for acute stroke patients in these settings. We evaluated the impact of an AI device by comparing the time to intervention (TTI) from NCCT imaging to significant intervention before (baseline) and after the deployment of AI, in patients diagnosed with stroke (ischemic or hemorrhagic) through a retrospective interrupted time series analysis at a rural hospital managed by non-specialists in India. Significant intervention was defined as thrombolysis or antiplatelet therapy in ischemic strokes, and mannitol for hemorrhagic strokes or mass effect. We also evaluated the diagnostic accuracy of the software using the teleradiologists’ reports as ground truth. Impact analysis in a total of 174 stroke patients (72 in baseline and 102 after deployment) demonstrated a significant reduction of median TTI from 80 minutes (IQR: 56·8–139·5) during baseline to 58·50 (IQR: 30·3–118.2) minutes after AI deployment (Wilcoxon rank sum test—location shift: -21·0, 95% CI: -38·0, -7·0). Diagnostic accuracy evaluation in a total of 864 NCCT scans demonstrated the sensitivity, specificity, positive predictive value (PPV) and negative predictive value (NPV) in detecting intracranial hemorrhage to be 0·89 (95% CI: 0·83–0·93), 0·99 (0·98–1·00), 0·96 (0·91–0·98) and 0·97 (0·96–0·98) respectively, and for infarct these were 0·84 (0·79–0·89), 0·81 (0·77–0·84), 0·58 (0·52–0·63), and 0·94 (0·92–0·96), respectively. AI-based NCCT interpretation supported faster abnormality detection with high accuracy, resulting in persons with acute stroke receiving significantly earlier treatment. Our results suggest that AI-based NCCT interpretation can potentially improve uptake of lifesaving interventions for acute stroke in resource-limited settings.

Objectives: Artificial intelligence (AI) algorithms have the potential to assist radiologists in the reporting of head computed tomography (CT) scans. We investigated the performance of an AI-based software device used in a large teleradiology practice for intracranial haemorrhage (ICH) detection. Methods: A randomly selected subset of all non-contrast CT head (NCCTH) scans from patients aged ≥18 years referred for urgent teleradiology reporting from 44 different hospitals within the United Kingdom over a 4-month period was considered for this evaluation. Thirty auditing radiologists evaluated the NCCTH scans and the AI output retrospectively. Agreement between AI and auditing radiologists is reported along with failure analysis. Results: A total of 1315 NCCTH scans from as many distinct patients (median age, 73 years [IQR 53-84]; 696 [52.9%] females) were evaluated. One hundred twelve (8.5%) scans had ICH. Overall agreement, positive percent agreement, negative percent agreement, and Gwet’s AC1 of AI with radiologists were found to be 93.5% (95% CI, 92.1-94.8), 85.7% (77.8-91.6), 94.3% (92.8-95.5) and 0.92 (0.90-0.94), respectively, in detecting ICH. 9 out of 16 false negative outcomes were due to missed subarachnoid haemorrhages and these were predominantly subtle haemorrhages. The most common reason for false positive results was due to motion artefacts. Conclusions: AI demonstrated very good agreement with the radiologists in the detection of ICH. Advances in knowledge: Real-world evaluation of an AI-based CT head interpretation device is reported. Knowledge of scenarios where false negative and false positive results are possible will help reporting radiologists.

Introduction Diagnosing and treating lung cancer in early stages is essential for survival outcomes. The chest X-ray (CXR) remains the primary screening tool to identify lung cancers in the UK; however, there is a shortfall of radiologists, while demand continues to increase. Image analysis by machine-learning software has the potential to support radiology workflows with a focus on immediate triage of suspicious X-rays. The RADICAL study will evaluate Qure.ai’s ‘qXR’ software in reducing reporting time for suspicious X-rays in NHS Greater Glasgow & Clyde. Methods and analysis This is a stepped-wedge cluster-randomised study consisting of a retrospective technical evaluation and prospective clinical effectiveness study alongside the assessment of acceptability via qualitative work and evaluation of cost-effectiveness via a cost utility analysis. The primary objective is to assess the clinical effectiveness of qXR to prioritise patients suspected with lung cancer on CXR for follow-up CT. Secondary objectives will look at the utility, safety, technical performance, health economics and acceptability of the intervention. The study period is 24 months, consisting of an initial 12 month data collection period and a 12 month follow-up period. All the standard care CXRs from outpatient and primary care requests will be securely transmitted to Qure.ai software ‘qXR’ for interpretation. Images with features of cancer will be flagged as ‘Urgent Suspicion of Cancer’ and be prioritised for radiologist review within the existing reporting workflow. Ethics and dissemination The study will follow the principles of Good Clinical Practice. The protocol was granted REC approval in August 2023 from North West—Greater Manchester West Research Ethics Committee (REC 23/NW/0211). This study was registered on clinicaltrials.gov (NCT06044454). An interim report will be produced for use by the Scottish Government. The results from this study will be presented at artificial intelligence, radiology and respiratory meetings and published in peer-reviewed journals. Trial registration number NCT06044454.

Background: Artificial intelligence (AI) based computer-aided detection devices are recommended for screening and triaging of pulmonary tuberculosis (TB) using digital chest x-ray (CXR) images (soft copies). Most AI algorithms are trained using input data from digital CXR Digital Imaging and Communications in Medicine (DICOM) files. There can be scenarios when only digital CXR films (hard copies) are available for interpretation. A smartphone-captured photo of the digital CXR film may be used for AI to process in such a scenario. There is a gap in the literature investigating if there is a significant difference in the performance of AI algorithms when digital CXR DICOM files are used as input for AI to process as opposed to photos of the digital CXR films being used as input. Objective: The primary objective was to compare the agreement of AI in detecting radiological signs of TB when using DICOM files (denoted as CXRd) as input versus when using smartphone-captured photos of digital CXR films (denoted as CXRp) with human readers. Methods: Pairs of CXRd and CXRp images were obtained retrospectively from patients screened for TB. AI results were obtained using both the CXRd and CXRp files. The majority consensus on the presence or absence of TB in CXR pairs was obtained from a panel of 3 independent radiologists. The positive and negative percent agreement of AI in detecting radiological signs of TB in CXRd and CXRp were estimated by comparing with the majority consensus. The distribution of AI probability scores was also compared. Results: A total of 1278 CXR pairs were analyzed. The positive percent agreement of AI was found to be 92.22% (95% CI 89.94-94.12) and 90.75% (95% CI 88.32-92.82), respectively, for CXRd and CXRp images (P=.09). The negative percent agreement of AI was 82.08% (95% CI 78.76-85.07) and 79.23% (95% CI 75.75-82.42), respectively, for CXRd and CXRp images (P=.06). The median of the AI probability score was 0.72 (IQR 0.11-0.97) in CXRd and 0.72 (IQR 0.14-0.96) in CXRp images (P=.75). Conclusions: We did not observe any statistically significant differences in the output of AI in digital CXRs and photos of digital CXR films.