Abstract
Introduction: Chest radiographs remain the most frequently advised and often the first-line imaging investigation for diagnosis of suspected respiratory disorders. The incidence of a pulmonary nodule in the general population on chest radiographs is a low 0.2%. As a result, most radiologists and clinicians alike are not even looking for it or miss it. However, a nodule can have a variety of etiological causes and in cases the cause is malignant, then there is a high likelihood that the patient will present later with a more widely spread neoplasia, thus impacting both patient morbidity and mortality negatively.
We developed a deep learning algorithm to diagnose pulmonary nodules and furthermore compared the actual impact it had on the performance of consultant clinicians for diagnosing the disease.
Methods: The model is an ensemble of two FPN with Xception encoder, trained on 5325 radiographs and externally validated on 310 radiographs. Our study group encompassed- 4 consultant clinicians from separate fields, 3 radiology consultants & 5 residents. They participated in an exercise to detect pulmonary nodules unaided and aided by the predictions of the model.
Results: The standalone AI has a specificity of 88% (83-92%) and sensitivity (CI) of 78% (69-85%). With the help of AI, the entire cohort’s Cohen Kappa score went up from the mean±SD of 58±17 to 66±8. The AUC improved from 0.79 to 0.83, with the greatest improvement seen in clinicians, increasing from 0.74 to 0.85.
Conclusion: The performance of clinicians to detect nodules improved significantly when they were provided the algorithmic aid, thus proving its clinical utility.
Footnotes
Cite this article as Eur Respir J 2022; 60: Suppl. 66, 4308.
This article was presented at the 2022 ERS International Congress, in session “-”.
This is an ERS International Congress abstract. No full-text version is available. Further material to accompany this abstract may be available at www.ers-education.org (ERS member access only).
- Copyright ©the authors 2022