Transformer network–based disease subtyping from multidimensional lesion-layer features-ZENTIME PUBLISHING CORPORATION LIMITED

Home Progress in Medical Devices All issues Issue 3

Transformer network–based disease subtyping from multidimensional lesion-layer features

Linrong Yuan^*, Yutong Xie^*, Danhong Li^* , Jianghui Li, Miao Yu, Siqi Wang, Yu Wang, He Ren

Faculty of Medical Instrumentation, Shanghai University of Medicine and Health Sciences, Shanghai 201318, China.

^*The authors contribute equally.

Address correspondence to: He Ren, Faculty of Medical Instrumentation, Shanghai University of Medicine and Health Sciences; No. 279, Zhouzhu Highway, Pudong New Area, Shanghai 201318, China. Tel: +86-18817581363. E-mail: renh@sumhs.edu.cn.

DOI: https://doi.org/10.61189/941872mmikqi

Received July 26, 2025; Accepted September 10, 2025; Published September 30, 2025

Highlights

● A total of 289 patient CT datasets were analyzed, and 15 optimal radiomic features were identified using ANOVA, correlation analysis, and random forest ranking, ensuring high discriminative power and clinical interpretability.

● The proposed model demonstrated excellent performance (Accuracy: 0.98, Area Under the Curve: 0.99) in training set, demonstrating robust learning capacity and the ability to distinguish lesion subtypes from multidimensional radiomic features.

● By leveraging serialized radiomic trends rather than isolated feature analysis, this study provides a new paradigm for early screening and personalized diagnosis of lung adenocarcinoma.

Abstract

Objective: To develop and validate a Transformer-based radiomics model for classifying lung adenocarcinoma subtypes from computed tomography imaging data. Methods: We retrospectively collected 289 computed tomography images of lung adenocarcinoma, including adenocarcinoma in situ, minimally invasive adenocarcinoma, and invasive adenocarcinoma. Correlation-based feature analysis was employed and identified 15 optimal radiomic features. A Transformer-based classification model incorporating multi-head attention and position-wise feed-forward Networks was subsequently constructed. Results: The proposed model achieved a training accuracy of 0.98, test accuracy of 0.914, training recall of 0.942, test recall of 0.874, training F1-score of 0.940, test F1-score of 0.871, training area under the curve of 0.99, and test area under the curve of 0.88. Conclusion: This Transformer-based radiomics model effectively classifies lung adenocarcinoma subtypes, aiding early screening, diagnosis, and personalized treatment strategies to improve patient prognosis.

Keywords: Computed tomography, lung adenocarcinoma subtype analysis, radiomic characteristics, deep learning, transformer network model

Download

Cite

Views

Downloads

Lastest Issue

Design and simulation of lower limb exoskeleton based on online gait generation algorithm Optimal design and experimental study of Mg alloy electrodes for tissue welding Measurement methods for positioning accuracy of multileaf collimators in radiation therapy: A mini review Research advances of beamforming algorithms in medical ultrasound systems Research progress of ankle-foot rehabilitation robots Gait prediction for lower limb exoskeleton robots based on real-time adaptive Kalman filtering Research progress on intestinal anastomosis technology and related devices Research progress on fixation methods for mandibular defect reconstruction Application of deep learning in the diagnosis of gastrointestinal diseases Research progress on simulation of radiofrequency ablation for liver cancer treatment