Non-Invasive Prediction of Lung Cancer Histological Differentiation via Radiomics and Multi-Binary Classification Models.
Authors
Affiliations (3)
Affiliations (3)
- School of Health Science and Engineering, University of Shanghai for Science and Technology, Shanghai, China.
- Department of Medical Affairs, People's Hospital of Yangzhong City, Yangzhong, Jiangsu, China.
- Department of Pathology, People's Hospital of Yangzhong City, Yangzhong, Jiangsu, China.
Abstract
The histological differentiation of Non-Small Cell Lung Cancer (NSCLC) is a critical prognostic factor that influences therapeutic strategies and patient outcomes. However, conventional assessment methods relying on postoperative pathology or biopsy are invasive and limited by sampling bias. Therefore, it is of great clinical significance to develop a non-invasive, imaging-based approach for accurate preoperative differentiation evaluation. This retrospective study included 184 NSCLC patients with preoperative chest CT scans and confirmed pathological differentiation grades from 2022 to 2024. Radiomics features were extracted using PyRadiomics, followed by feature selection via the LASSO algorithm. A novel three-task binary classification strategy was proposed to replace conventional trinary classification, including low vs. non-low, moderate vs. non-moderate, and high vs. non-high differentiation. Four machine learning models-GBDT, RF, XGBoost, and LightGBM-were constructed and evaluated using ROC analysis, confusion matrices, and SHAP-based interpretability analysis. The GBDT model achieved the highest AUC (0.849) in the low differentiation classification task, while the RF model outperformed others in predicting high differentiation (AUC = 0.7188). The moderate differentiation task showed relatively poor performance across all models (AUC < 0.55). SHAP analysis revealed that features such as original_firstorder_Kurtosis, glrlm_RunEntropy, and wavelet-HLL_firstorder_Median played key roles in differentiating tumor grades, highlighting their biological relevance and potential utility in clinical interpretation. The proposed multi-binary strategy improved classification granularity and interpretability. Ensemble learning models demonstrated robust performance across tasks, especially for extreme differentiation levels. This study, which combines radiomics with a multi-task machine learning framework, demonstrates prediction and can improve the accuracy and interpretability of preoperative lung cancer differentiation. The proposed model provides a non-invasive, quantitative tool with the potential to support individualized clinical decision-making. Further multicenter validation and multimodal data integration are warranted to enhance its clinical applicability.