Differential Diagnosis of Pneumoconiosis Mass Shadows and Peripheral Lung Cancer Using CT Radiomics and the AdaBoost Machine Learning Model

Yongsheng Liu¹Xiaobing Li¹Wei Wang¹Xuemei Li²Li Wang¹Qian Li¹Li Zhang¹Wutao Xie^1*Qianqian Liu¹

• ¹Chongqing Medical and Pharmaceutical College, Chongqing, China
• ²Chongqing Medical University, Chongqing, China

To develop a differential diagnostic prediction model for distinguishing large opacities in pneumoconiosis from peripheral lung cancer based on CT radiomics. [Methods] A total of 103 cases of large opacities in pneumoconiosis and 85 cases of peripheral lung cancer were retrospectively collected from routine CT scans at the First Affiliated Hospital of Chongqing Medical and Pharmaceutical College between March 2021 and June 2025. Diagnosis was confirmed by an expert panel, clinical evaluations, and pathological examinations. Patients were randomly assigned to a training set (n=132) and a test set (n=56). Lesions were delineated by at least two pneumoconiosis experts using ITK-SNAP software. Radiomic features were extracted from CT images of lung lesions in the training set, including first-order features, shape features (2D and 3D), texture features (gray-level co-occurrence matrix, gray-level run-length matrix, gray-level size-zone matrix, gray-level dependence matrix), and wavelet transform filters. Feature dimensionality reduction was applied to construct morphological biomarkers. Diagnostic prediction models were built using machine learning algorithms. Model performance was evaluated using the ROC curve and the area under the curve (AUC) in the test set. [Results] A total of 108 features were extracted from 110 large opacity regions and 85 peripheral lung cancer regions of interest (ROIs). Dimensionality reduction identified a subset of eight most significant features. LR, SVM, and AdaBoost algorithms were implemented using Python to build the models. In the training set, the accuracies of the LR, SVM, and AdaBoost models were 79.4%, 84.0%, and 80.9%, respectively; the sensitivities were 74.1%, 74.1%, and 81.0%, respectively; the specificities were 83.6%, 91.8%, and 80.8%, respectively; and the AUC values were 0.837, 0.886, and 0.900, respectively. In the test set, the accuracies of the LR, SVM, and AdaBoost models were 80.7%, 82.5%, and 86.0%, respectively; the sensitivities were 89.3%, 89.3%, and 82.1%, respectively; the specificities were 72.4%, 75.9%, and 89.7%, respectively; and the AUC values were 0.825, 0.855, and 0.900, respectively. The AUC of the AdaBoost ROC curve was significantly superior to those of the LR and SVM models. The AdaBoost model demonstrated the optimal predictive performance in both the training and test sets.