Ultrasound-based deep learning radiomics for the differential diagnosis of benign and malignant subpleural pulmonary lesions

TL;DR

This study creates a deep learning model using ultrasound images to accurately distinguish between benign and malignant lung lesions, improving diagnosis and reducing unnecessary procedures.

Contribution

A novel ultrasound-based clinical deep learning radiomics model for diagnosing subpleural pulmonary lesions is developed and validated.

Findings

The CDLR model achieved high AUC values of 0.987 (training) and 0.924 (validation) for differentiating benign and malignant lesions.

The model outperformed standalone clinical, radiomics, and deep learning models in validation, with high sensitivity, specificity, and accuracy.

Grad-CAM and SHAP analysis enhanced model interpretability by highlighting key image regions and feature contributions.

Abstract

This study aims to develop an ultrasound-driven clinical deep learning radiomics (CDLR) model for the differential diagnosis of benign and malignant subpleural pulmonary lesions (SPLs), with the goal of guiding personalized treatment and minimizing unnecessary interventions.

A retrospective analysis was conducted on 609 SPL patients from July 2020 to February 2024 at Guangxi Medical University. The dataset was divided into training (487 cases) and validation (122 cases) cohorts. Prior to ultrasound-guided lung mass biopsy, 1561 radiomics (Rad) features were extracted from every ultrasound image, alongside 128 deep transfer learning (DTL) features after dimensionality reduction and compression based on ResNet-50. Feature selection was performed, followed by the development of a deep learning radiomics (DLR) model using a Support Vector Machine (SVM), which was then used to derive the model’s feature. Clinical data were analyzed through univariate and multivariate logistic regression, generating the clinical features. The DLR and clinical features were integrated using SVM to create the CDLR model for differentiating benign and malignant SPLs. Model performance was evaluated using the area under the receiver operating characteristic curve (AUC), and its clinical utility was assessed via decision curve analysis (DCA). The Shapley Additive Explanation (SHAP) method and Gradient weighted Class Activation Mapping (Grad-CAM) visualization were employed to enhance model interpretability.

The CDLR model demonstrated high accuracy in distinguishing benign and malignant SPLs. The AUC values for the training and validation set were 0.987 and 0.924, respectively. Notably, the CDLR model outperformed clinical, standalone Rad, DTL, and DLR models in the validation cohort. The model also achieved the highest sensitivity (0.871), specificity (0.897), and accuracy (0.877). Grad-CAM visualization highlighted key regions of interest within ultrasound images, and SHAP analysis identified the contributions of clinical, deep learning, and radiomics features.

The ultrasound-based CDLR model provides a robust tool for differentiating benign and malignant SPLs, offering superior diagnostic performance compared to existing ultrasound diagnostic criteria. This model is valuable for early lung cancer screening and can reduce unnecessary biopsies or surgeries for pulmonary masses.

Figures