Interpretable machine learning model based on gray-scale ultrasound image for noninvasive prediction of BPGT and MPGT

Accurate differentiation between benign parotid gland tumors (BPGT) and malignant parotid gland tumors (MPGT) is of paramount importance for clinical decision-making and treatment planning. Early and precise diagnosis can significantly influence the choice of therapeutic strategies, patient prognosis, and overall survival rates. Given the clinical significance of distinguishing between BPGT and MPGT, the present study aims to develop a nomogram model that integrates clinical information, ultrasound parameters, and radiomics features extracted from gray-scale ultrasound images to differentiate between BPGT and MPGT. Furthermore, the model’s performance will be evaluated using Shapley Additive Explanations (SHAP) analysis. This retrospective and prospective study examined grayscale ultrasound images from 425 patients diagnosed with parotid gland tumors. 242 patients from Center1 and Center2 constituted the training cohort (including 197 BPGT and 45 MPGT), while 104 patients from Center3 served as the external validation cohort (including 80 BPGT and 24 MPGT). Additionally, 79 patients were prospectively collected from three centers (including 63 BPGT and 16 MPGT) for the prospective validation cohort. Clinic model was developed using clinical data and ultrasound visual information. Radiomics features were extracted from the grayscale ultrasound images to construct radiomics models and corresponding Rad-score. Integrating the Rad score and clinical factors, through multiple logistic regression analysis, to draw a nomogram. These models’ performance was evaluated utilizing receiver operating characteristic (ROC) curve analysis, calibration curves, decision curve analysis (DCA), and the Hosmer-Lemeshow (HL) test. Additionally, SHAP analysis was employed to enhances interpretability in the Rad mode by quantifying and visualizing the impact of each feature on predictions, providing clear insights into model behavior. Four ultrasound visual features were selected using logistic regression (LR) to develop the clinical model, while nineteen features were ultimately chosen from the grayscale ultrasound images through feature screening to construct the radiomics (Rad) model. In the external and prospective validation cohort, compared with Clinic (ACC = 0.769 and 0.785) and Rad (ACC = 0.827 and 0.810), the nomogram showed the best discriminative ability (ACC = 0.817 and 0.848). The nomogram model integrating clinical features and radiomics features constructed in this study has demonstrated higher accuracy and robustness in differentiating BPGT and MPGT. SHAP analysis has enhanced the transparency and interpretability of the model. The model not only support clinical diagnosis and help reduce unnecessary biopsies and surgeries, but also lay the foundation for future multimodal precise diagnosis of parotid gland tumors.

Authors

Institutions

Publication Details

Journal
BMC Cancer
Published
2026-11-14
DOI
https://doi.org/10.1186/s12885-026-17264-9
Primary Topic
Radiomics and Machine Learning in Medical Imaging
Type
article
Field-Weighted Citation Impact
0.00
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
article

Interpretable machine learning model based on gray-scale ultrasound image for noninvasive prediction of BPGT and MPGT

Weiping Zhang, Yu-Hong Diao, Jingling Wang, Fang-Qun Chen et al.
BMC Cancer
Radiomics and Machine Learning in Medical Imaging
article

Interpretable machine learning model based on gray-scale ultrasound image for noninvasive prediction of BPGT and MPGT

Weiping Zhang, Yu-Hong Diao, Jingling Wang, Fang-Qun Chen, Li-Ping Jiang, Yi Mao, Xing-Qi Peng, Zhi-Xing Liu, Li Chen
article en

Abstract

Accurate differentiation between benign parotid gland tumors (BPGT) and malignant parotid gland tumors (MPGT) is of paramount importance for clinical decision-making and treatment planning. Early and precise diagnosis can significantly influence the choice of therapeutic strategies, patient prognosis, and overall survival rates. Given the clinical significance of distinguishing between BPGT and MPGT, the present study aims to develop a nomogram model that integrates clinical information, ultrasound parameters, and radiomics features extracted from gray-scale ultrasound images to differentiate between BPGT and MPGT. Furthermore, the model’s performance will be evaluated using Shapley Additive Explanations (SHAP) analysis. This retrospective and prospective study examined grayscale ultrasound images from 425 patients diagnosed with parotid gland tumors. 242 patients from Center1 and Center2 constituted the training cohort (including 197 BPGT and 45 MPGT), while 104 patients from Center3 served as the external validation cohort (including 80 BPGT and 24 MPGT). Additionally, 79 patients were prospectively collected from three centers (including 63 BPGT and 16 MPGT) for the prospective validation cohort. Clinic model was developed using clinical data and ultrasound visual information. Radiomics features were extracted from the grayscale ultrasound images to construct radiomics models and corresponding Rad-score. Integrating the Rad score and clinical factors, through multiple logistic regression analysis, to draw a nomogram. These models’ performance was evaluated utilizing receiver operating characteristic (ROC) curve analysis, calibration curves, decision curve analysis (DCA), and the Hosmer-Lemeshow (HL) test. Additionally, SHAP analysis was employed to enhances interpretability in the Rad mode by quantifying and visualizing the impact of each feature on predictions, providing clear insights into model behavior. Four ultrasound visual features were selected using logistic regression (LR) to develop the clinical model, while nineteen features were ultimately chosen from the grayscale ultrasound images through feature screening to construct the radiomics (Rad) model. In the external and prospective validation cohort, compared with Clinic (ACC = 0.769 and 0.785) and Rad (ACC = 0.827 and 0.810), the nomogram showed the best discriminative ability (ACC = 0.817 and 0.848). The nomogram model integrating clinical features and radiomics features constructed in this study has demonstrated higher accuracy and robustness in differentiating BPGT and MPGT. SHAP analysis has enhanced the transparency and interpretability of the model. The model not only support clinical diagnosis and help reduce unnecessary biopsies and surgeries, but also lay the foundation for future multimodal precise diagnosis of parotid gland tumors.

BMC Cancer
Second Affiliated Hospital of Nanchang University (CN), First Affiliated Hospital of Nanchang University (CN)
Radiomics and Machine Learning in Medical Imaging
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon