Differential diagnosis model for tuberculous and malignant pleural effusion combining U-Net automatic segmentation and deep learning

Vollständiger Abstract

Worum geht es in dieser Arbeit?

Background We aimed to construct and validate an intelligent differential diagnosis model that integrates U-Net-based automatic segmentation with a deep learning classification model, and to evaluate its diagnostic performance and clinical value for distinguishing tuberculous pleural effusion (TPE) from malignant pleural effusion (MPE). Methods A total of 281 patients with pleural effusion confirmed by etiological or pathological evidence between January 2018 and August 2025 were included, comprising 143 patients with TPE and 138 with MPE. First, a U-Net model was employed to automatically segment pleural lesion regions on chest computed tomography (CT) images and extract regions of interest (ROIs). Subsequently, based on the segmentation results, a radiomics model, a two-dimensional deep learning (DL2D) model, and a comprehensive model integrating clinical features were constructed. Multiple machine learning algorithms, including support vector machines (SVMs), random forests (RFs), and extremely randomized trees (ERTs), were utilized for model construction and comparison. Model performance was evaluated using the area under the receiver operating characteristic curve (AUC), sensitivity, specificity, calibration curves, decision curve analysis (DCA), and the integrated discrimination improvement (IDI) and net reclassification improvement (NRI) indices. Results The U-Net segmentation model achieved Dice coefficients of 0.873 and 0.862 in the training and test sets, respectively, indicating good segmentation performance. In the test set, the comprehensive model demonstrated the best performance, with an AUC of 0.934 (95% CI 0.8733–0.9955), sensitivity of 0.875, and specificity of 0.900. Its performance was superior to that of the clinical model (AUC = 0.767), the radiomics model (AUC = 0.841), and the DL2D model (AUC = 0.776). DCA confirmed that the comprehensive model provided a higher net clinical benefit across a wide range of threshold probabilities. Furthermore, IDI and NRI analyses indicated that the comprehensive model significantly improved predictive performance relative to the individual models ( p < 0.05). Conclusion The model combining U-Net-based automatic segmentation with a deep learning classification model exhibited excellent and balanced diagnostic performance for differentiating TPE from MPE. It has the potential to provide an objective, stable, and scalable intelligent decision-support tool for clinical practice.