Diagnosing alveolar septal thickening remains challenging, and predicting alveolar diffusion is crucial for preoperative radiological assessment of pulmonary conditions in patients. Non-invasive screening of early-stage patients using imaging techniques holds significant clinical importance. In response to this challenge, we effectively predict alveolar diffusion in adenocarcinoma nodules using a radiomics and deep learning combined method, named the Spread Through Air Space (STAS) Prediction Model. Specifically, by fusing radiomic features from the lung cancer nodule lesion region with deep learning features, the mutual enhancement of feature representation leads to a remarkable area under the curve (AUC) of 0.830 in the binary classification task (STAS patients vs. non-STAS patients) for the radiomics model. Moreover, the deep learning model, utilizing ResNet-18 network to extract deep features from tumor blocks, achieves an AUC of 0.841. The combined model, incorporating both deep learning and traditional radiomic features, outperforms standalone deep learning and radiomics models by 3.50% and 4.60%, respectively. The introduction of radiomic features enhances the model’s interpretability, demonstrating promising clinical applicability.
|