Classification of Invasiveness and Prediction of Ki-67 Expression Level in Pulmonary Ground-Glass Nodules Based on CT Images Using Artificial Intelligence

Abstract

This work aimed to accurately determine the invasiveness level of pulmonary ground-glass nodules (GGNs) preoperatively and to adjust personalized treatment plans accordingly. This paper presents a multi-task artificial intelligence framework to simultaneously perform three-class invasiveness classification of GGNs and prediction of Ki-67 expression levels, and enhances the clinical transparency and potential utility of the model through multi-dimensional explainability analysis. A total of 900 patients with histologically confirmed lung adenocarcinoma presenting as GGNs were retrospectively enrolled from three hospitals: 60% were assigned to the training set, 15% to the internal validation set, 15% to the external validation set, and 10% to the external test set. A multi-task multi-instance learning framework integrating 3D Res2Net deep features, radiomic features, and clinical variables (MT-MIL-RF) was proposed. A gated attention mechanism and a cross-task attention module were introduced to learn the potential biological association between invasiveness classification and Ki-67 prediction. An independent interpretability structure has been constructed, and both Shapley Additive exPlanations (SHAP) variable attribution and Gradient-weighted Class Activation Mapping (Grad-CAM) spatial localisation have been employed. The principal contribution of this work is a validated AI framework that, for the first time, provides a simultaneous, non-invasive assessment of both the invasiveness and proliferative activity (Ki-67) of GGNs. By demonstrating that a multi-task learning strategy with cross-task attention significantly outperforms conventional single-task models, we offer a clinically promising and transparent tool to improve personalized surgical planning and risk stratification for lung adenocarcinoma.