Artificial Intelligence-Based Software with CE Mark for Chest X-ray Interpretation: Opportunities and Challenges

Performance of AI to exclude normal chest radiographs to reduce radiologists' workload

INTRODUCTION

This study investigates the performance of a commercially available artificial intelligence (AI) system to identify normal chest radiographs and its potential to reduce radiologist workload.

METHODS

Retrospective analysis included consecutive chest radiographs from two medical centers between Oct 1, 2016 and Oct 14, 2016. Exclusions comprised follow-up exams within the inclusion period, bedside radiographs, incomplete images, imported radiographs, and pediatric radiographs. Three chest radiologists categorized findings into normal, clinically irrelevant, clinically relevant, urgent, and critical. A commercial AI system processed all radiographs, scoring 10 chest abnormalities on a 0-100 confidence scale. AI system performance was evaluated using the area under the ROC curve (AUC), assessing the detection of normal radiographs. Sensitivity was calculated for the default and a conservative operating point. the detection of negative predictive value (NPV) for urgent and critical findings, as well as the potential workload reduction, was calculated.

RESULTS

A total of 2603 radiographs were acquired in 2141 unique patients. Post-exclusion, 1670 radiographs were analyzed. Categories included 479 normal, 332 clinically irrelevant, 339 clinically relevant, 501 urgent, and 19 critical findings. The AI system achieved an AUC of 0.92. Sensitivity for normal radiographs was 92% at default and 53% at the conservative operating point. At the conservative operating point, NPV was 98% for urgent and critical findings, and could result in a 15% workload reduction.

CONCLUSION

A commercially available AI system effectively identifies normal chest radiographs and holds the potential to lessen radiologists' workload by omitting half of the normal exams from reporting.

CLINICAL RELEVANCE STATEMENT

The AI system is able to detect half of all normal chest radiographs at a clinically acceptable operating point, thereby potentially reducing the workload for the radiologists by 15%.

KEY POINTS

The AI system reached an AUC of 0.92 for the detection of normal chest radiographs. Fifty-three percent of normal chest radiographs were identified with a NPV of 98% for urgent findings. AI can reduce the workload of chest radiography reporting by 15%.

Quantitative CT Texture Analysis of COVID-19 Hospitalized Patients during 3-24-Month Follow-Up and Correlation with Functional Parameters

BACKGROUND

To quantitatively evaluate CT lung abnormalities in COVID-19 survivors from the acute phase to 24-month follow-up. Quantitative CT features as predictors of abnormalities' persistence were investigated.

METHODS

Patients who survived COVID-19 were retrospectively enrolled and underwent a chest CT at baseline (T0) and 3 months (T3) after discharge, with pulmonary function tests (PFTs). Patients with residual CT abnormalities repeated the CT at 12 (T12) and 24 (T24) months after discharge. A machine-learning-based software, CALIPER, calculated the CT percentage of the whole lung of normal parenchyma, ground glass (GG), reticulation (Ret), and vascular-related structures (VRSs). Differences (Δ) were calculated between time points. Receiver operating characteristic (ROC) curve analyses were performed to test the baseline parameters as predictors of functional impairment at T3 and of the persistence of CT abnormalities at T12.

RESULTS

The cohort included 128 patients at T0, 133 at T3, 61 at T12, and 34 at T24. The GG medians were 8.44%, 0.14%, 0.13% and 0.12% at T0, T3, T12 and T24. The Ret medians were 2.79% at T0 and 0.14% at the following time points. All Δ significantly differed from 0, except between T12 and T24. The GG and VRSs at T0 achieved AUCs of 0.73 as predictors of functional impairment, and area under the curves (AUCs) of 0.71 and 0.72 for the persistence of CT abnormalities at T12.

CONCLUSIONS

CALIPER accurately quantified the CT changes up to the 24-month follow-up. Resolution mostly occurred at T3, and Ret persisting at T12 was almost unchanged at T24. The baseline parameters were good predictors of functional impairment at T3 and of abnormalities' persistence at T12.

A Novel Automatic Algorithm to Support Lung Ultrasound Non-Expert Physicians in Interstitial Pneumonia Evaluation: A Single-Center Study

INTRODUCTION

Lung ultrasound (LUS) is widely used in clinical practice for identifying interstitial lung diseases (ILDs) and assessing their progression. Although high-resolution computed tomography (HRCT) remains the gold standard for evaluating the severity of ILDs, LUS can be performed as a screening method or as a follow-up tool post-HRCT. Minimum training is needed to better identify typical lesions, and the integration of innovative artificial intelligence (AI) automatic algorithms may enhance diagnostic efficiency.

AIM

This study aims to assess the effectiveness of a novel AI algorithm in automatic ILD recognition and scoring in comparison to an expert LUS sonographer. The "SensUS Lung" device, equipped with an automatic algorithm, was employed for the automatic recognition of the typical ILD patterns and to calculate an index grading of the interstitial involvement.

METHODS

We selected 33 Caucasian patients in follow-up for ILDs exhibiting typical HRCT patterns (honeycombing, ground glass, fibrosis). An expert physician evaluated all patients with LUS on twelve segments (six per side). Next, blinded to the previous evaluation, an untrained operator, a non-expert in LUS, performed the exam with the SensUS device equipped with the automatic algorithm ("SensUS Lung") using the same protocol. Pulmonary functional tests (PFT) and DLCO were conducted for all patients, categorizing them as having reduced or preserved DLCO. The SensUS device indicated different grades of interstitial involvement named Lung Staging that were scored from 0 (absent) to 4 (peak), which was compared to the Lung Ultrasound Score (LUS score) by dividing it by the number of segments evaluated. Statistical analyses were done with Wilcoxon tests for paired values or Mann-Whitney for unpaired samples, and correlations were performed using Spearman analysis; p < 0.05 was considered significant.

RESULTS

Lung Staging was non-inferior to LUS score in identifying the risk of ILDs (median SensUS 1 [0-2] vs. LUS 0.67 [0.25-1.54]; p = 0.84). Furthermore, the grade of interstitial pulmonary involvement detected with the SensUS device is directly related to the LUS score (r = 0.607, p = 0.002). Lung Staging values were inversely correlated with forced expiratory volume at first second (FEV1%, r = -0.40, p = 0.027), forced vital capacity (FVC%, r = -0.39, p = 0.03) and forced expiratory flow (FEF) at 25th percentile (FEF25%, r = -0.39, p = 0.02) while results directly correlated with FEF25-75% (r = 0.45, p = 0.04) and FEF75% (r = 0.43, p = 0.01). Finally, in patients with reduced DLCO, the Lung Staging was significantly higher, overlapping the LUS (reduced median 1 [1-2] vs. preserved 0 [0-1], p = 0.001), and overlapping the LUS (reduced median 18 [4-20] vs. preserved 5.5 [2-9], p = 0.035).

CONCLUSIONS

Our data suggest that the considered AI automatic algorithm may assist non-expert physicians in LUS, resulting in non-inferior-to-expert LUS despite a tendency to overestimate ILD lesions. Therefore, the AI algorithm has the potential to support physicians, particularly non-expert LUS sonographers, in daily clinical practice to monitor patients with ILDs. The adopted device is user-friendly, offering a fully automatic real-time analysis. However, it needs proper training in basic skills.

Computer-Aided Detection for Chest Radiography to Improve the Quality of Tuberculosis Diagnosis in Vietnam's District Health Facilities: An Implementation Study

In Vietnam, chest radiography (CXR) is used to refer people for GeneXpert (Xpert) testing to diagnose tuberculosis (TB), demonstrating high yield for TB but a wide range of CXR abnormality rates. In a multi-center implementation study, computer-aided detection (CAD) was integrated into facility-based TB case finding to standardize CXR interpretation. CAD integration was guided by a programmatic framework developed for routine implementation. From April through December 2022, 24,945 CXRs from TB-vulnerable populations presenting to district health facilities were evaluated. Physicians interpreted all CXRs in parallel with CAD (qXR 3.0) software, for which the selected TB threshold score was ≥0.60. At three months, there was 47.3% concordance between physician and CAD TB-presumptive CXR results, 7.8% of individuals who received CXRs were referred for Xpert testing, and 858 people diagnosed with Xpert-confirmed TB per 100,000 CXRs. This increased at nine months to 76.1% concordant physician and CAD TB-presumptive CXRs, 9.6% referred for Xpert testing, and 2112 people with Xpert-confirmed TB per 100,000 CXRs. Our programmatic CAD-CXR framework effectively supported physicians in district facilities to improve the quality of referral for diagnostic testing and increase TB detection yield. Concordance between physician and CAD CXR results improved with training and was important to optimize Xpert testing.