New lung lesions after previous cancer can create a difficult diagnostic question because they may reflect either a second primary lung cancer or lung metastasis. A multicentre retrospective AI-SONAR biomarker analysis published in European Radiology assessed whether visible CT features, defined by radiologists, could support this distinction. The work focused on single-timepoint, pre-treatment CT thorax scans from patients with new lung lesions occurring within 10 years of prior radically treated cancer. Nine thoracic oncology radiologists reviewed predefined imaging features, and a semantic-feature model called the Second Malignancy Aetiology Recognition Tool, or SMART, was developed to classify lesion origin. The model was compared with clinical radiologist assessment in a real-world diagnostic setting.
Semantic Features Shape Diagnostic Assessment
The dataset included 649 technically usable CT thorax scans from the malignant arm of AI-SONAR. Cases were divided between second primary lung cancer and lung metastasis, with diagnosis confirmed by biopsy or multidisciplinary team consensus, including follow-up imaging when relevant. The scans came from two host institutions and included external imports, giving the dataset a varied clinical profile.
Radiologists assessed eight visible CT features: emphysema, lesion contour, spiculation, lobulation, density, cavitation, feeding vessel and lesion distribution. The region of interest was the lung lesion that had determined the clinical diagnosis through biopsy sampling or multidisciplinary review. For most feature assessment, radiologists focused on the lesion itself. For clinical classification, they reviewed the wider thoracic field.
Several CT patterns helped distinguish between the two malignant categories. Emphysema, irregular contour and spiculation appeared more often in second primary lung cancer. Peripheral distribution appeared more often in lung metastasis. Although all assessed features showed some discriminatory value in initial testing, four features remained important when assessed together: emphysema, contour, spiculation and distribution.
SMART Combines CT Findings into a Model
The SMART model was built from the four semantic CT features that remained significant in combined assessment. Its purpose was to provide a structured, scan-based method for classifying new malignant lung lesions after previous cancer. The model did not rely on broader clinical variables, preserving a focus on CT appearances alone.
Across the malignant cohort, SMART achieved an area under the curve of 0.81 and an overall accuracy of 75%. When compared with radiologist classification of lung metastasis, the model showed higher sensitivity and higher overall accuracy. Radiologists showed higher specificity, meaning they were more likely to correctly identify cases that were not metastasis. However, radiologist assessment more often classified metastatic lesions as not metastasis.
Must Read: Hybrid CT Predicts GGN Invasiveness in Lung Cancer
A more restricted comparison included cases in which radiologists labelled lesions as either second primary lung cancer or lung metastasis. In that subgroup, radiologist accuracy was slightly higher than SMART accuracy, while SMART retained higher sensitivity for lung metastasis. Radiologists performed particularly well in recognising second primary lung cancer. These results show that the model and expert clinical assessment offered different strengths, with SMART contributing a more sensitive approach to metastatic classification.
Reader Variation Highlights the Need for Structure
Radiologist agreement varied across the CT features. Cavitation showed the strongest agreement, while contour, spiculation, distribution and overall clinical classification showed moderate agreement. Emphysema, lobulation, density and feeding vessel showed lower agreement. This variation matters because semantic features depend on visual interpretation, and not every feature is equally reproducible across readers.
Clinical classification also showed a tendency to favour second primary lung cancer over lung metastasis. This matters in practice because the management of a new lung cancer differs considerably from metastatic disease after prior cancer. Classification may influence treatment selection, further imaging, biopsy decisions and use of healthcare resources.
The SMART model offers a standardised way to combine multiple CT features, but limitations remain. It was developed and evaluated within the same AI-SONAR dataset, so independent external validation is still needed. Most lesions were solid, limiting evaluation of ground glass and part-solid patterns. Clinical factors were not included, although age differed between the two lesion groups. Further work may assess whether adding clinical variables, sequential imaging or radiomic features can improve classification.
Semantic CT features can support differentiation between second primary lung cancer and lung metastasis in patients with new lung lesions after prior radically treated cancer. Emphysema, irregular contour, spiculation and distribution form the basis of the SMART model, which performed comparably with expert thoracic radiologist assessment and showed higher sensitivity for metastasis. Radiologist variation and underclassification of metastatic disease underline the need for structured decision support, although roader validation is needed before wider clinical use.
Source: European Radiology
Image Credit: iStock
References:
Kalsi HS, Linton-Reid K, Kim C et al. (2026) Semantic CT features and differentiation model: new primary lung cancer versus metastasis after previous malignancy. Eur Radiol: In Press.
