Deep learning model predicts growth of lung ground-glass nodules

Pulmonary ground-glass nodules, the hazy spots frequently picked up in low-dose CT screening, often grow so slowly that patients face years of follow-up scans before doctors can tell whether a nodule is benign or malignant. A new retrospective study published in BMC Cancer describes a multimodal deep learning framework designed to shorten that waiting game by predicting, early on, whether a nodule is likely to grow.

The research team analyzed data from 421 patients with pulmonary nodules seen between December 2021 and October 2024, of whom 147 had nodules that grew and 274 had stable nodules. Their approach combined three kinds of information: spatial features extracted from CT images by a three-dimensional convolutional neural network based on the ResNet3D architecture, clinical variables modeled with a random forest algorithm, and the patient’s epidermal growth factor receptor (EGFR) mutation status. The separate data streams were merged using a probabilistic averaging fusion strategy.

In parallel, the researchers extracted 107 radiomic features from the images using the PyRadiomics toolkit and narrowed these down to 15 representative features through recursive feature elimination. Exploratory analyses using the Wilcoxon rank-sum test and logistic regression identified three radiomic metrics—TotalEnergy, ClusterShade, and LeastAxisLength—as independently correlated with nodule growth. These features were also strongly associated with nodule type, EGFR mutation status, and typical imaging manifestations.

On an independent test set, the fused model achieved an area under the receiver operating characteristic curve (AUC) of 0.9123. A DeLong test showed that the fusion model was significantly superior to the single ResNet3D model. However, no statistically significant AUC difference was observed between the fusion model and the random forest model that incorporated EGFR status.

The role of EGFR mutations proved more nuanced than the molecular data’s prominence in the model might suggest. Although EGFR mutations were more common in growing nodules, the predictive improvement contributed by EGFR information was only marginal and did not reach statistical significance. In other words, the molecular marker added little beyond what imaging and clinical variables already provided.

A subgroup analysis focused on progressive pure ground-glass nodules revealed an extremely low sensitivity, with only four progressive pure ground-glass nodule positive cases in the test set. The authors note that this scarcity renders the subgroup analysis exploratory, and the framework’s performance in that specific subgroup cannot be considered established.

The study’s conclusion is measured: the multimodal framework effectively predicts growth trends of pulmonary ground-glass nodules by integrating deep radiomics, clinical features, and molecular data, and the modeling strategy provides tentative evidence supporting clinical risk evaluation. The authors suggest it may assist clinicians in streamlining follow-up protocols appropriately, though the retrospective design and the exploratory nature of some analyses mean the findings warrant further validation before broader clinical use.

Subject of Research: A multimodal deep learning framework for predicting the growth trends of pulmonary ground-glass nodules

Article Title: Multimodal deep learning framework integrating deep radiomics, clinical features, and EGFR Status to predict growth trends of pulmonary ground-glass nodules

Article References: Huang, S., Qiao, K., Li, Z., Lin, C., Hai, J., Wu, Y., Yan, B., Zhang, Q., Chen, J., Yang, X., Wang, Z., & Wang, X. (2026). Multimodal deep learning framework integrating deep radiomics, clinical features, and EGFR Status to predict growth trends of pulmonary ground-glass nodules. BMC Cancer. https://doi.org/10.1186/s12885-026-17257-8

Image Credits: AI Generated

DOI: 10.1186/s12885-026-17257-8

Keywords: ground-glass nodules, deep learning, radiomics, EGFR, lung cancer screening, computed tomography, ResNet3D, random forest, PyRadiomics, BMC Cancer, nodule growth, predictive modeling