Transformer‐Based Segmentation and Feature Fusion for Multi‐Class Breast Ultrasound Classification
International Journal of Imaging Systems and Technology · 2026-05 · 36 (3)
DOI: 10.1002/ima.70312 ↗Kurz gefasst
Fachartikel in der Zeitschrift International Journal of Imaging Systems and Technology, online erschienen am 29.04.2026. Verfasst von Liqun Shi und 2 weiteren Autor:innen. Die Arbeit nennt 49 Literaturangaben.
Abstract (Originalsprache)
This study aimed to develop and evaluate a comprehensive AI‐driven pipeline for multi‐class classification of breast lesions in ultrasound images, integrating advanced segmentation, radiomic and deep feature fusion, and state‐of‐the‐art machine learning classifiers. This multi‐center study analyzed 3574 breast ultrasound images from eight institutions. After applying quality control and exclusion criteria, 2876 cases from seven centers were used for training and internal testing, and 698 cases from the eighth center were used for external validation. Five segmentation models (U‐Net, nnU‐Net, TransUNet, Swin‐UNet, and MedT, the Medical Transformer) were evaluated using Dice Similarity Coefficient (Dice), Intersection over Union (IoU), Hausdorff Distance (HD), and Average Symmetric Surface Distance (ASSD). Tumor regions were automatically segmented, followed by extraction of radiomic and deep features. Features were filtered using the intraclass correlation coefficient (ICC), reduced via standard deviation and Pearson correlation, and selected using LASSO, RFE, and MI. Three classifiers (Extreme Gradient Boosting or XGBoost, Tabular Network or TabNet, and Tabular Transformer or TabTransformer) were applied to radiomic, deep, and fused features. Classification was assessed using accuracy, AUC, sensitivity, confusion matrices, q ‐values, and t ‐distributed Stochastic Neighbor Embedding ( t ‐SNE) plots. MedT consistently achieved the highest segmentation accuracy (Dice = 95.32%) and generated the most stable features. The best classification setup, which used MedT for segmentation, RFE‐selected fused features, and TabTransformer for classification, reached accuracies of 95.2%, 94.0%, and 93.1% on the training, test, and external datasets, respectively. The corresponding AUCs were 97.1%, 95.8%, and 94.7%. Confusion matrices showed high precision across all categories. Q ‐value analysis confirmed the statistical reliability of fused features, and t ‐SNE plots illustrated strong post‐classification separability. This study presents a robust and generalizable diagnostic pipeline for classifying breast ultrasound lesions. The combination of MedT‐based segmentation, feature fusion, and transformer‐based classification offers a reliable solution for automated diagnosis.
Zitieren
APA
Shi, L., Shi, M., & Zhu, H. (2026). Transformer‐Based Segmentation and Feature Fusion for Multi‐Class Breast Ultrasound Classification. International Journal of Imaging Systems and Technology, 36(3). https://doi.org/10.1002/ima.70312
Vancouver
Shi L, Shi M, Zhu H. Transformer‐Based Segmentation and Feature Fusion for Multi‐Class Breast Ultrasound Classification. International Journal of Imaging Systems and Technology. 2026;36(3). doi:10.1002/ima.70312
BibTeX
@article{Shi2026,
author = {Shi, Liqun and Shi, Mingli and Zhu, Huihui},
title = {Transformer‐Based Segmentation and Feature Fusion for Multi‐Class Breast Ultrasound Classification},
journal = {International Journal of Imaging Systems and Technology},
year = {2026},
volume = {36},
number = {3},
doi = {10.1002/ima.70312}
}
EUVIMED