A pulmonary lesion can usually be observed as a rounded, well-defined opacity on CXR. The margins may be smooth or convoluted and it is surrounded by normally aerated lung tissue. A lesion is defined by its size, location, margins, and density. Generally, if the lesion is less than 30 mm in diameter, it is considered a nodule; if more than 30 mm, it is considered a mass [2,3]. Occasionally, nodules as small as 5-6 mm in size can be visible on CXR [4]. Despite this, even larger lesions can stay undetected by a well-experienced radiologist [5]. Although most solitary pulmonary nodules are benign findings, when a lesion is discovered, the probability of its malignancy must be assessed using clinical and radiological features [6]. If a suspicion of lung cancer arises from the CXR, a contrast-enhanced chest CT is used for further assessment of the pathology nature and origin, and, in case of a malignant tumor, for staging. Recently, the role of CT has also been applied in the screening of selected patient groups: a lung cancer screening program has been running in the Czech Republic since January 2022.
For this study, a total of 9,276 CXR images were retrieved from the period between June 2020 and July 2021 from a specialized oncology center, Masaryk Memorial Cancer Institute. To determine the ground truth for the presence or absence of a pulmonary lesion on CXR, three radiologists with different levels of experience were assessed for the initial reading. At least 2 out of 3 radiologists needed to agree on the presence of a pulmonary lesion in order to establish the ground truth.
Subsequently, 300 CXR images were randomly selected for the experiment using the random.choice() function, of which 100 (55 F, 45 M) were from patients with a total of 385 confirmed pulmonary lesions (LES+ Abnormal), 100 (68 F, 32 M) were from patients with various confirmed findings other than pulmonary lesions (LES- Abnormal), and 100 (79 F, 21 M) were without any pathological findings (Normal). Five radiologists were invited to retrospectively evaluate 300 CXR images, and the performance of individual radiologists was subsequently compared with that of DLAD. The experience levels of the radiologists were as follows: RAD 1 and RAD 2 were junior radiologists with less than 2 years of experience, RAD 3 was a radiologist with more than 2 years but less than 5 years of experience, RAD 4 and RAD 5 were board-certified radiologists with more than 5 years of experience.
The performance was quantified by the means of balanced accuracy (BA), sensitivity (Se), specificity (Sp), and positive (PLR) and negative likelihood ratio (NLR). A paired design was applied to the data, i.e. each CXR was evaluated by DLAD and all assessed radiologists, and their evaluation was compared to the ground truth. To compare the Se and Sp of the DLAD with that of individual radiologists, we calculated and compared statistical parameters using confidence intervals (CI) and p-values. CIs were constructed at a two-tailed 95% confidence level.