Interstitial lung disease pattern recognition on full high resolution computed tomography volume: Development and evaluation of a decision-support tool for less-experimented radiologists.
Authors
Affiliations (13)
Affiliations (13)
- Department of Radiology, Hôpital Cochin, AP-HP, 75014 Paris, France.
- Department of Pulmonology, Hôpital Bichat, AP-HP, 75018 Paris, France; Université Paris Cité, 75006 Paris, France.
- Université Paris Cité, 75006 Paris, France; Department of Radiology, Hôpital Saint-Louis, AP-HP, 75010 Paris, France.
- Department of Radiology, Hôpital Pontchaillou, CHU de Rennes, 35000 Rennes, France; Université de Rennes, 35000 Rennes, France.
- Unité d'Imagerie Cardiovasculaire et Thoracique, Hôpital Pitié-Salpétrière, AP-HP, 75013 Paris, France.
- Department of Radiology, Hôpital Bichat, AP-HP - INSERM 1152, 75018 Paris, France.
- Department of Radiology, Hôpital Saint-Louis, AP-HP, 75010 Paris, France.
- Université Paris Cité, 75006 Paris, France; Department of Internal Medicine, Hôpital Cochin, AP-HP, 75014 Paris, France.
- Department of Pulmonology, Hôpital Cochin, AP-HP, 75014 Paris, France.
- Université Paris Cité, 75006 Paris, France; Department of Pulmonology, Hôpital Saint-Louis, AP-HP, 75010 Paris, France.
- Department of Pulmonology, Hôpital Saint-Louis, AP-HP, 75010 Paris, France.
- Université Paris Cité, 75006 Paris, France; Department of Radiology, Hôpital Bichat, AP-HP - INSERM 1152, 75018 Paris, France.
- Department of Radiology, Hôpital Cochin, AP-HP, 75014 Paris, France; Université Paris Cité, 75006 Paris, France. Electronic address: [email protected].
Abstract
The purpose of this study was to develop an artificial intelligence (AI) tool to assist recognition of three major interstitial lung disease (ILD) patterns on high-resolution computed tomography (HRCT) and to evaluate its added value in supporting decision-making for non-specialist radiologists. This retrospective, multicenter study included 1097 HRCT examinations. Of these, 989 (90.15%) were used for development and 108 (9.85%) for external testing. A two-stage architecture inspired by domain-specific pretraining was employed. The encoder of a three-dimensional ILD segmentation model was kept to extract 7168 disease-specific features per HRCT, which were combined with age and sex in a deep learning model to predict three radiological patterns (usual interstitial pneumonia, non-specific interstitial pneumonia and fibrotic bronchiolocentric interstitial pneumonia) as diagnosed in multidisciplinary discussions (MDD). The external test dataset was interpreted by seven thoracic radiologists to establish a second reference (majority's vote) and by eight radiology residents with and without AI assistance. Accuracy, sensitivity and specificity were calculated for each pattern. The AI system achieved 77.8% accuracy on the external test dataset using MMD as a reference standard, within the range of thoracic experts (median, 75.6%; range: 61.1-81.5). AI assistance improved residents' median accuracy (+14.8% of absolute increase) and reduced reading time by 20.7% (P < 0.001). Six out of eight residents assisted by AI (75%) performed worse than AI alone. AI can accurately classify major ILD patterns and help less-experienced readers improve their performance. However, the level of improvement was inconsistent, and non-specialists rarely equaled the performance of AI alone.