AI-Driven Automated Detection of Pleural Plaques on Chest CT Scans in Retired Asbestos-Exposed Workers

Our plain-English summary of AI-Driven Automated Detection of Pleural Plaques on Chest CT Scans in Retired Asbestos-Exposed Workers, by Y. Petitpas, I. Benlala, F. Baldacci, et al (Diagnostics, 2026).

The study

• Authors: Y. Petitpas, I. Benlala, F. Baldacci, G. Dournes, K. Jaunas, B. Clin, P. Brochard, A. Gislard, F. Delva, A. Lacourt, C. Paris, F. Laurent, J.C. Pairon, B.D. Senneville
• Journal: Diagnostics
• Published: 2026;16(16)
• Read it: PubMed · Publisher · Free full text

Abstract

Background/Objective: This study aimed to develop and validate an automated framework for detecting pleural plaques (PPs) at the patient level using a single chest CT scan. Methods: A database of chest CT scans with corresponding individual binary annotations for PP presence was first established through expert visual assessment, based on consensus among a board of specialized radiologists. This dataset served to train and validate a patient-level classification framework for automated PP presence detection. This framework leverages a pre-trained PP segmentation network, thus obviating the need for retraining large-scale deep learning models. This segmentation backbone is supplemented by a lightweight classification module that integrates the network's outputs to infer the presence or absence of PPs at patient level. Furthermore, a patch-wise processing strategy was employed to optimize computational efficiency and training time while retaining critical local contextual information. Results: The framework was evaluated on a cohort of 1241 retired workers with documented occupational asbestos exposure, using 10-fold cross-validation (30% training, 10% validation, 60% testing). It achieved a balanced accuracy of 97.0%, sensitivity of 96.6%, and specificity of 97.4%, demonstrating performance slightly superior to that of two expert radiologists (balanced accuracy: 90.9%/91.4%, sensitivity: 88.0%/88.8%, specificity: 93.9%/94.1% for Expert1/Expert2) and substantially superior to that of a radiologist without specialized expertise in asbestos-related manifestations (balanced accuracy: 80.0%, sensitivity: 75.1%, specificity: 84.8%). Conclusion: The proposed framework demonstrates strong potential for automated patient-level identification of pleural plaques from a single CT scan. Our approach could be particularly useful in screening settings that involve compensation claims related to occupational diseases.

In plain English

This study built and tested an artificial intelligence program that judges from one chest CT scan whether a person has pleural plaques. It used 2015 to 2018 scans of 1241 retired workers with documented occupational asbestos exposure in a French multicenter cohort. By consensus of two expert radiologists, with a third resolving disagreements, 345 people had plaques and 896 did not. The program was trained on 30% of scans and tested on 60%, over 10 random splits. Averaged over the test scans, it had a balanced accuracy of 97.0% (95% confidence interval 96.2 to 97.7), a sensitivity of 96.6% (94.7 to 98.5) and a specificity of 97.4% (96.7 to 98.0). The two expert radiologists had balanced accuracies of 90.9% (90.3 to 91.5) and 91.4% (90.8 to 92.0). A radiologist without asbestos expertise, whose readings came from original scan reports, had 80.0% (78.7 to 81.3). The authors concluded the program reached expert-level performance and could help prioritize scans for expert review, not replace experts. They stated it must be validated in independent, more varied cohorts before broader clinical use.

We wrote this plain-English summary from the full article with the help of AI, and a second AI model checked each sentence against the article. It is not the authors' wording; the abstract and the article are their own account.

What the study did and found

• Design: Development and validation of an automated (artificial intelligence) method for detecting pleural plaques on CT, compared with radiologists' readings
• Who was studied: 1241 retired workers with documented occupational asbestos exposure from a multicenter French screening cohort (ARDCO/APEXS). CT scans came from the third screening round (2015 to 2018). Of 1308 third-round participants, 67 had corrupted CT data. By the reference standard, 345 were classified as having pleural plaques and 896 as not.
• Exposure: Documented occupational asbestos exposure. The article does not give fiber types, jobs or doses.
• Main results:
  • Reference standard: consensus of the two expert radiologists, with a third radiologist resolving disagreements. It classified 345 patients as having pleural plaques and 896 as not.
  • Proposed program, mean over the test subsets: balanced accuracy 97.0% (95% CI 96.2 to 97.7), sensitivity 96.6%, specificity 97.4%.
  • Proposed program, 95% confidence intervals from Table 1: sensitivity 96.6% (94.7 to 98.5), specificity 97.4% (96.7 to 98.0).
  • Expert 1, compared with the reference standard: balanced accuracy 90.9% (95% CI 90.3 to 91.5), sensitivity 88.0%, specificity 93.9%.
  • Expert 2, compared with the reference standard: balanced accuracy 91.4% (95% CI 90.8 to 92.0), sensitivity 88.8%, specificity 94.1%.
  • Radiologist without specialized expertise in asbestos-related manifestations (readings from the initial scan reports): balanced accuracy 80.0% (95% CI 78.7 to 81.3), sensitivity 75.1%, specificity 84.8%.
  • The earlier approach of applying a threshold directly to the segmentation output produced many false positives and only moderate specificity. In Table 1 its balanced accuracy was 83.6%.
  • Changing the size of the image patches analyzed made no statistically significant difference to performance.
  • Processing time: generating the probability map took 7 minutes 52 seconds on a workstation. The final classification step then took 12 seconds on the workstation and 38 seconds on an ordinary laptop.
• The authors' conclusion: The authors concluded that the framework achieved expert-level performance for pleural plaque detection in this cohort of retired asbestos-exposed workers. They said it could support surveillance and compensation-related workflows by prioritizing cases for expert review, not by replacing expert assessment. They stated that external validation in independent and more heterogeneous cohorts is necessary before broader clinical deployment.
• Limitations the authors state:
  • The method used to combine patch-level results into a patient-level decision (max-pooling) is known to be sensitive to outliers and false positives.
  • The study population consisted of retired asbestos-exposed workers. Results may not apply to people having CT scans for other reasons, such as routine clinical practice.
  • Training curves showed overfitting: the model memorized training data but did not generalize to validation data. The authors used early stopping to reduce this.
  • External validation in independent and more heterogeneous cohorts has not yet been done and is necessary before broader clinical deployment.

Topics

Medical terms in this study, explained in our Lexicon: Pleural Plaques, CT Scan, Asbestos Exposure, Occupational Disease.

Related studies

• Do asbestos-related pleural plaques on HRCT scans cause restrictive impairment in the absence of pulmonary fibrosis? — B. Clin, C. Paris, et al. (2011)
• Deep Learning for the Automatic Quantification of Pleural Plaques in Asbestos-Exposed Subjects — I. Benlala, B.D. De Senneville, et al. (2022)
• Performance of chest radiograph and CT scan for lung cancer screening in asbestos-exposed workers — B. Clin, F. Morlais, et al. (2009)
• Inter-reader agreement in HRCT detection of pleural plaques and asbestosis in participants with previous occupational exposure to asbestos — F. Laurent, C. Paris, et al. (2014)
• 3-Dimensional Quantification of Composite Pleural Plaque Volume in Patients Exposed to Asbestos Using High-resolution Computed Tomography: A Validation Study — G. Dournes, A. Dubois, et al. (2019)
• Asbestos exposure, pleural plaques and digestive cancers — B. Clin, C. Gramond, et al. (2025)

Browse more research in Diagnostics, or return to medical research.