Scholarlink

Purpose <p>Pneumonia is a common, clinically important cause of acute respiratory illness. Chest CT aids detection and extent assessment, but interpretation can be variable and time sensitive. We evaluated InferVision for pneumonia detection and decision support on chest CT against radiologist-consensus reference standard.</p> Methods <p>We performed a retrospective diagnostic-accuracy evaluation of a fixed InferVision chest CT pneumonia-assessment system in 115 consecutive clinical chest CT examinations. The system generated a case-level pneumonia flag, a continuous probability score, and quantitative lesion-burden outputs. Two expert radiologists independently assessed each examination for pneumonia presence/absence, with disagreements resolved by consensus as the imaging-based reference standard. InferVision outputs were compared with this reference to assess diagnostic performance, agreement, discrimination, calibration, threshold trade-offs, decision-curve net benefit, and lesion-volume agreement in true-positive cases.</p> Results <p>Pneumonia was present in 55/115 (47.8%) examinations by consensus. At the prespecified operating point, InferVision produced 46 true positives, 56 true negatives, 4 false positives, and 9 false negatives, yielding sensitivity 83.6%, specificity 93.3%, PPV 92.0%, NPV 86.2%, accuracy 88.7%, and F1-score 0.876; AI–consensus agreement was κ = 0.773, and inter-radiologist agreement was κ = 0.821. Discrimination was good (AUC 0.886; average precision 0.887) with partial calibration (Brier 0.143). Decision-curve analysis indicated clinical utility across thresholds (pt = 0.30: net benefit 0.359 vs. treat-all 0.255; pt = 0.50: 0.287 vs. − 0.043). In true positives, AI lesion volumes correlated strongly with reference (<i>r</i> = 0.981) with small positive bias (+ 2.88&#xa0;cm³).</p> Conclusion <p>InferVision demonstrated high specificity and precision with good discrimination for pneumonia detection on CT and provided potentially useful decision-support outputs and standardized extent quantification, supporting its potential role as an adjunct to radiologist interpretation.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Clinical evaluation of an AI-based pneumonia detection system on chest CT: a qualitative and quantitative analysis

  • 期刊论文

  • 发表日期：

    2026年4月9日

查看全文
  • Hamza Sekkat,
  • Youssef Madkouri,
  • Abdellah Khallouqi,
  • Mohammed Talbi,
  • Oussama El Mouden,
  • Abdellah Halimi,
  • Omar El rhazouani,
  • Youssef El Merabet,
  • Farida Bentayeb

摘要

Purpose

Pneumonia is a common, clinically important cause of acute respiratory illness. Chest CT aids detection and extent assessment, but interpretation can be variable and time sensitive. We evaluated InferVision for pneumonia detection and decision support on chest CT against radiologist-consensus reference standard.

Methods

We performed a retrospective diagnostic-accuracy evaluation of a fixed InferVision chest CT pneumonia-assessment system in 115 consecutive clinical chest CT examinations. The system generated a case-level pneumonia flag, a continuous probability score, and quantitative lesion-burden outputs. Two expert radiologists independently assessed each examination for pneumonia presence/absence, with disagreements resolved by consensus as the imaging-based reference standard. InferVision outputs were compared with this reference to assess diagnostic performance, agreement, discrimination, calibration, threshold trade-offs, decision-curve net benefit, and lesion-volume agreement in true-positive cases.

Results

Pneumonia was present in 55/115 (47.8%) examinations by consensus. At the prespecified operating point, InferVision produced 46 true positives, 56 true negatives, 4 false positives, and 9 false negatives, yielding sensitivity 83.6%, specificity 93.3%, PPV 92.0%, NPV 86.2%, accuracy 88.7%, and F1-score 0.876; AI–consensus agreement was κ = 0.773, and inter-radiologist agreement was κ = 0.821. Discrimination was good (AUC 0.886; average precision 0.887) with partial calibration (Brier 0.143). Decision-curve analysis indicated clinical utility across thresholds (pt = 0.30: net benefit 0.359 vs. treat-all 0.255; pt = 0.50: 0.287 vs. − 0.043). In true positives, AI lesion volumes correlated strongly with reference (r = 0.981) with small positive bias (+ 2.88 cm³).

Conclusion

InferVision demonstrated high specificity and precision with good discrimination for pneumonia detection on CT and provided potentially useful decision-support outputs and standardized extent quantification, supporting its potential role as an adjunct to radiologist interpretation.