3143 - An End-to-end Intelligent Breast Cancer Radiotherapy Platform Based on Large Language Models: From Patient Consultation to Target Volume Delineation

12:30pm - 01:45pm ET

Poster Hall - Exhibit Hall A

Screen: 33

POSTER

Presenter(s)

Xiaoyue Sun, MS - Sichuan University, Chengdu, Sichuan

X. Sun¹, J. Lan^1,2, D. Gao³, X. Chen¹, X. Wu^1,2, L. LIU^1,2, L. Jia⁴, W. Zhang³, and J. Jing¹; ¹Institute of Breast Health Medicine, State Key Laboratory of Biotherapy, West China Hospital, Sichuan University and Collaborative Innovation Center, Chengdu, Sichuan, China, ²Division of Head and Neck Tumor Multimodality Treatment, Cancer Center, Institute of Breast Health Medicine, West China Hospital, Sichuan University, Chengdu, Sichuan, China, ³Shanghai United Imaging Healthcare Co., Ltd., Shanghai, China, ⁴Shenzhen United Imaging Healthcare Co., Ltd., Shenzhen, Guangdong, China

Purpose/Objective(s):

Large language models (LLM) have demonstrated significant potential in extracting clinical information from patient examination reports and simulating radiation oncologists (ROs) logic to generate radiotherapy decisions, which makes it possible to automate these repetitive tasks of ROs. In this study, we developed an end-to-end intelligent breast cancer radiotherapy platform (IBCRP) based on LLM, aiming to assist ROs in completing the workflow from outpatient information acquisition to target generation with minimal intervention.

Materials/Methods:

The IBCRP consists of three modules: clinical information module, radiotherapy decision module, and text-guided target delineation module. In the first module, patients upload ten types of patient reports, such as US, MRI, and pathology reports, by scanning a QR code via the WeChat Mini Program. Then, the system employs optimized prompt engineering to invoke Qwen-VL-Max, automatically extracting a total of 36 clinical information. ROs review and modify the extracted information through a web interface. In the second module, a local knowledge base based on clinical guidelines is established. Leveraging the patient’s clinical information, the LLM simulates the ROs’ logic and generates personalized radiotherapy plans by integrating content from the knowledge base. Finally, in the third module, key delineation-related information from the radiotherapy decision—such as lesion location and target volume scope—is fed as textual input into a text-guided target delineation model to complete segmentation.

Results:

The IBCRP has completed the integration of all modules and is currently in the preliminary testing phase. Patients now can directly photograph and upload required reports via the WeChat Mini Program, and the system completes report classification and clinical information extraction within approximately 2 minutes after upload. Among all 36 clinical information items extracted, the average accuracy exceeds 90%. ROs can centrally review the extracted clinical information through the platform’s web interface. The review interface adopts a split-screen layout, with information displayed on the left and the original report synchronized on the right. The average time for review and modification is approximately 2 minutes. The generation of radiotherapy decisions takes an average of about 1 minute, with the accuracy exceeding 85%. In the target delineation stage, the system automatically completes target delineation based on delineation information in less than 1 minute on average, and the average Dice exceeds 0.8.

Conclusion:

Leveraging LLM capability, the IBCBP enables integrated management of multiple tasks within the clinical workflow, and demonstrates clear potential for significantly reducing clinical workload for ROs.