Multi-component radiological model based on intratumoral CT threshold segmentation for predicting visceral pleural invasion in lung adenocarcinoma ≤ 30 mm

OBJECTIVES: This retrospective study aims to investigate the value of intratumoral computed tomography (CT) threshold segmentation in radiomics, deep learning (DL), and radiomics-DL combined models for predicting visceral pleural invasion (VPI) in lung adenocarcinoma (LUAD) ≤ 30 mm. MATERIALS AND METHODS: Patients with invasive LUAD who underwent surgery and had preoperative thin-slice CT scans within four weeks were enrolled from two centers (n = 816). Patients from center 1 were divided into a training set (TS, n = 591) and an internal test set (ITS, n = 98) based on surgical time. Patients from center 2 constituted the external test set (ETS, n = 127). Solid, ground-glass, and peritumoral components were extracted using intratumoral CT threshold segmentation and peritumoral expansion methods. The radiomics model was a Random Forest Classifier; the DL model was a pre-trained Vision Transformer (ViT) fine-tuned on three components; the combined model integrated features from two pipelines. Model performance was evaluated by the area under the receiver operating characteristic curve (AUC), sensitivity, and specificity. Clinical utility and model interpretability were evaluated using decision curve analysis and SHapley Additive exPlanations (SHAP), respectively. RESULTS: Radiomics, ViT, and radiomics-ViT models achieved AUCs of 0.865, 0.858, and 0.895 in TS; 0.865, 0.844, and 0.852 in ITS; and 0.844, 0.816, and 0.823 in ETS, respectively. Radiomics-ViT model achieved the highest sensitivity, with ViT features contributing the most. CONCLUSION: A hybrid multi-component feature pipeline could serve as a reliable and highly sensitive tool for VPI prediction in LUAD ≤ 30 mm. CRITICAL RELEVANCE STATEMENT: The multi-component radiomics-ViT model achieved high sensitivity for VPI prediction in LUAD ≤ 30 mm, which is a promising tool for preoperative treatment design and prognostic assessment. KEY POINTS: CT attenuation-defined components remain underexplored in artificial intelligence (AI) models for VPI prediction. Solid, ground-glass, and peritumoral components enable AI models for VPI prediction. Multi-component inputs effectively capture tumor heterogeneity for VPI prediction in LUAD ≤ 30 mm.