Development of a Comprehensive Model Combining Clinical, Radiomics With Deep Learning for Predicting the Micropapillary Pattern in ≤2 cm Invasive Lung Adenocarcinoma

OBJECTIVES: Micropapillary pattern (MPP) in ≤2 cm invasive lung adenocarcinoma (ILADC) is strongly associated with poor prognosis, making its preoperative identification critical for determining optimal surgical strategy. This study aimed to develop a predictive model integrating clinical features, radiomics features, and deep learning features to non-invasively identify MPP in ≤2 cm ILADC. METHODS: This retrospective study analysed 311 patients with pathologically confirmed ILADC (102 with MPP ≥ 5%, 209 without MPP) treated at Zhongda Hospital, Southeast University, from January 2018 to August 2023. Clinical features, radiomics features extracted using PyRadiomics, and deep learning features obtained from a 3D convolutional neural network (NASLung) were selected through t-tests and random forest (RF) feature-importance analysis. Three base models (Clinic, Radiomics, Deep Learning) were trained using RF or support vector machine (SVM) classifiers, and their predicted probabilities were fused to construct 3 combined models: (1) Clinic + Rad-based model (CR), (2) Clinic + Deep Learning-based model (CD), (3) Rad + Deep Learning-based model (RDL), and (4) Clinic + Rad + Deep Learning-based model (CRDL). Model performance was evaluated by receiver operating characteristic (ROC) analysis, calibration, and decision curve analysis (DCA). RESULTS: A total of 6 clinical features, 30 radiomics features, and 8 deep learning features were ultimately selected. In the testing set, the CRDL model demonstrated the best performance, achieving an AUC of 0.8817, a sensitivity of 77.4%, and a specificity of 81.0%, outperforming all other models. The calibration curve showed good agreement between predicted and observed outcomes, and the DCA further confirmed the clinical net benefit of the CRDL model. CONCLUSIONS: CRDL model effectively predicts MPP in ≤2 cm ILADC preoperatively, offering a non-invasive tool to guide surgical decision-making and optimize patient management.