Effectiveness of artificial intelligence for detecting operable lung cancer on chest radiographs

Abstract

Background: Despite the importance of early diagnosis of lung cancer and wide availability of chest radiography, the detection of operable stage lung cancer on chest radiographs (CXRs) remains challenging. This study aimed to investigate the effectiveness of artificial intelligence (AI)-based CXR analysis for detecting operable lung cancers.

Methods: Patients who underwent lung cancer surgery at two referral hospitals between March 2020 and February 2021 were retrospectively included in this study. Preoperative CXRs of the patients were analyzed using commercial AI-based lesion detection software, and the results of lesion location and types obtained using the software were reviewed by radiologists and pulmonologists, with computed tomography (CT) as a reference standard for determining nodule characteristics. Factors influencing AI detection of lung cancer on CXR were assessed using logistic regression analysis.

Results: Among the 594 patients who underwent surgery for lung cancer (median age: 65 years, 51.3% male), the sensitivity of AI for detecting lung cancer on CXR was 57.7%, and it identified 86% of CXR-visible lung cancers. Detection rates of lung cancer by AI increased according to the disease stage: 42.5% for stage IA, 86.3% for stage IB, and 90.9% for stages II-III. The detection rate increased to over 60% from stage IA2 onwards when tumor size exceeded 1 cm. Regarding lesion type on CT, 8.3%, 46.8%, and 77.3% of non-solid, part-solid, and solid nodules, respectively, were detected by AI. Multivariable analysis showed that nodule location in the upper zone [odds ratio (OR) 2.78, P<0.001], peripheral region (OR 4.59, P<0.001), and solid lesion diameter (OR 1.20, P<0.001) were significantly associated with AI detection of lung cancer.

Conclusions: AI could be an effective tool for detecting operable lung cancer on CXRs, particularly when lesions are larger and located in the upper and peripheral regions.