Interpretable Multimodal Artificial Intelligence Predicts High-risk Pathology in Lung Cancer Associated With Cystic Airspaces

Lung cancer associated with cystic airspaces (LCCA) is an uncommon radiological presentation of lung cancer characterized by cystic or air-containing spaces with adjacent or surrounding abnormal lung tissue. Although many LCCA lesions are detected at an early clinical stage, their pathological behavior can vary considerably. Some tumors may contain high-risk pathological features associated with greater invasive potential and a less favorable prognosis. Identifying these features before surgery may therefore help improve individualized surgical planning and clinical decision-making.

This observational study aims to develop and validate an artificial intelligence (AI)-based multimodal model for the preoperative prediction of pathological high-risk features in patients with LCCA. The model will integrate information derived from preoperative chest computed tomography (CT) images with routinely available clinical and radiological variables.

Eligible patients with surgically resected LCCA will be included in the study. Preoperative CT images will be analyzed using AI-based imaging methods to extract tumor-related imaging information. Clinical and radiological characteristics available before surgery will also be collected. These data will be integrated to construct a multimodal prediction model.

The reference standard will be the final postoperative pathological examination of the resected tumor. Pathological high-risk features will include predefined histological and invasive characteristics associated with more aggressive tumor behavior. The ability of the multimodal model to predict these pathological features before surgery will be evaluated using measures including the area under the receiver operating characteristic curve, sensitivity, specificity, accuracy, positive predictive value, and negative predictive value.

The performance of the multimodal model will also be compared with models based on CT imaging information alone and clinical or tabular information alone to determine whether combining multiple sources of preoperative information provides additional predictive value. Model interpretation methods will be used to identify imaging regions and clinical variables that contribute to individual predictions, thereby improving the transparency and clinical interpretability of the AI model.