Related Experiment Video
Updated: Sep 9, 2025
Guidelines and Experience Using Imaging Biomarker Explorer IBEX for Radiomics
Published on: January 8, 2018
An Interpretable Radiomics Model Integrating Ultrasound and Clinical Features for Multiclass Classification of
Yin Zheng1, Chen Chen2, Hongwei You3
1The Second School of Clinical Medicine, Zhejiang Chinese Medical University, Hangzhou, Zhejiang, China; Department of Diagnostic Ultrasound Imaging & Interventional Therapy, Zhejiang Cancer Hospital, Hangzhou, Zhejiang, China.
Objective:
Accurate identification of common lesions (benign, metastatic carcinoma, lymphoma) of axillary lymph nodes (ALNs) is of vital importance for clinical decision-making, directly affecting tumor staging, treatment strategy formulation and prognosis of patients. However, current diagnostic methods by radiologists have limited accuracy in differentiating between benign lesions, metastatic carcinoma, and lymphoma.This study aimed to develop and validate the Axillary Lymph Node Interpretable Prediction (ALNIP) model for classifying axillary lymph nodes into benign, metastatic carcinoma, or lymphoma based on ultrasound imaging.
Methods:
For this multicenter retrospective study, 1480 ALNs were collected from institution 1 constituting the training set (n = 806) and test set (n = 344), and institutions 2 and 3 serving as external validation sets 1 (n = 214) and 2 (n = 116). Radiomics features were extracted from ultrasound and selected using ElasticNet regression to construct the radiomics model. Based on these features, ALNIP model was developed by incorporating radiomics features with clinical variables and subsequently validated. The results were compared with the performance of radiologists with different levels of seniority. Histograms of radiomics feature distributions and the SHapley Additive exPlanations (SHAP) method were used to perform interpretable analyses.
Results:
Nine radiomics features were selected, and Logistic-Radiomics model performed best achieving a MicroAUC of 0.835 (p < 0.05). ALNIP model, built upon this Logistic-Radiomics model, achieved MicroAUCs of 0.924, 0.905, and 0.853 on the test set, the external validation set 1, and the external validation set 2, respectively. ALNIP model outperformed significantly radiologists of all experience levels (p < 0.05).
Conclusions:
The ALNIP model offers a reliable and interpretable approach to predicting ALN status preoperatively, particularly for ALNs that are difficult to assess on ultrasound, with significant potential for noninvasive diagnosis.
