Semantic CT features and differentiation model: new primary lung cancer versus metastasis after previous malignancy

New pulmonary lesions after prior cancer present a diagnostic challenge, potentially representing malignancy relapse or new primary lung cancer due to shared risk factors and/or impact of prior oncological therapies. This study evaluated radiologist-defined semantic features for differentiation of second primary lung cancer (SPLC) versus lung metastasis (LM).

651 single-timepoint, pre-treatment CT thorax scans from the multicentre retrospective AI-SONAR biomarker study (IRAS 331656 REC 23/NE/0151) were divided for review by nine thoracic oncology radiologists to evaluate eight semantic features. Logistic regression analysis was undertaken to identify significant features and a developed 'Second Malignancy Aetiology Recognition Tool' model (SMART) was compared to real-world clinical reader performance using McNemar's test.

649 scans were technically usable, 299 SPLC and 350 LM. Emphysema (p < 0.0001, OR 0.20 [95% CI 0.14-0.29]), irregular contour (p < 0.0001, OR 0.31 [95% CI 0.20-0.48]) and spiculation (p = 0.013, OR 0.51 [95% CI 0.30-0.89]) were more prevalent in SPLC (OR < 1 indicates association with SPLC). Peripheral lung distribution (p = 0.003, OR 1.80 [95% CI 1.20-2.68]) was more common in LM (OR > 1 indicates metastasis). SMART model AUC was 0.81 (95% CI 0.78-0.84), LM classification accuracy 75% vs 69% by radiology reader and McNemar p-value < 0.01 for comparative accuracy. 550/649 cases were predicted SPLC or LM by radiologists, in which the SMART model LM classification accuracy was 74% vs 77% by reader and McNemar p-value 0.20.

In new lesions after prior treated cancer, radiologist readers called SPLC more often than LM. The SMART model performed comparably with expert thoracic radiologists in the diagnosis of LM.

Question Differentiating the malignant aetiology of indeterminate lung lesions after prior cancer presents a growing diagnostic challenge. The literature and nodule guidelines are sparse for this setting. Findings A SMART model derived from semantic CT imaging features correctly classified malignant lung lesions as metastasis or lung cancer, more often than thoracic radiologists. Clinical relevance The SMART model could improve the stratification of malignant new lung lesions after prior cancer. This may lead to earlier diagnosis and optimise patient management, treatment selection and downstream outcomes.