195 - A Deep Learning-Driven Framework for Large Language Model-Assisted Automatic Target Volume Localization and Delineation for Enhancing Spinal Metastases Stereotactic Body Radiotherapy Workflow

Z. Yang¹, M. Kazemimoghadam², L. Wang³, G. A. Szalkowski⁴, C. F. Chuang¹, L. Liu⁵, S. G. Soltys¹, E. Pollom¹, E. Rahimy¹, H. Jiang⁶, D. Park⁷, A. Persad⁷, Y. Hori⁷, J. Fu¹, I. O. Romero⁸, L. Zalavari⁹, M. Chen², W. Lu², and X. Gu¹⁰; ¹Department of Radiation Oncology, Stanford University School of Medicine, Stanford, CA, ²Medical Artificial Intelligence and Automation (MAIA) Lab, Department of Radiation Oncology, UT Southwestern Medical Center, Dallas, TX,, Dallas, TX, ³Department of Radiation Oncology, Stanford University School of Medicine, Palo Alto, CA, ⁴Georgia Institute of Technology, Atlanta, GA, United States, ⁵Department of Radiation Oncology, Stanford University, Stanford, CA, ⁶NeuralRad LLC, Madison, WI, ⁷Department of Neurosurgery, Stanford University School of Medicine, Stanford, CA, ⁸Stanford University School of Medicine, Palo Alto, CA, ⁹Stanford Health Care, Stanford, CA, United States, ¹⁰Stanford University Department of Radiation Oncology, Palo Alto, CA

Purpose/Objective(s): Precise target volume delineation is essential in achieving high-quality treatment plan for spine stereotactic body radiotherapy (SBRT). Yet, this manual delineation process remains a laborious task that is also prone to considerable variations among clinicians. The automation of target volume localization and delineation remains an urgent need. In this study, we proposed a deep learning (DL)-driven framework to enable the automatic localization and delineation of target volumes in spine SBRT, aiming to enhance the spine SBRT workflow efficiency. Materials/

Methods: We collected 56 spine SBRT cases with metastasis limited to single vertebral level in the thoracic or lumbar spine from our institutional database. They were partitioned into 3 groups: 40 for training, 8 for validation and 8 for independent testing. Each case includes CT, T1c, and T2 MRI data, alongside corresponding clinical charts. The framework comprises three steps: (1) Whole spine vertebrae segmentation and labeling, (2) Large language model (LLM)-assisted target localization, and (3) Target volume segmentation. Firstly, the multi-modal images will be co-registered and resampled into the same resolution and spatial coordinate as the CT. The full spine vertebrae will be segmented and labeled according to the spine atlas from CT using a TotalSeg-based DL model. Subsequently, for each case, the clinical chart is imported into the platform, and its anonymized text will be input into a LLM to identify the lesion location on the vertebral level. In our current preliminary approach, the anonymized chart text was fed into GPT-3.5-Turbo model through application programming interface (API) to obtain a string of the lesion location (e.g., “T3”) as output. After localizing the targeted vertebral level, the corresponding vertebra mask generated from step 1 will be utilized to crop the multi-modal images into small volumes of interest (VOI) centered at the targeted vertebra. Lastly, a 3D nnUNet-based segmentation model was employed to segment the target volumes using the localized multi-modal crops. Additionally, this framework is supported by a comprehensive web platform facilitating automatic execution of procedures upon data reception, alongside user interactions like review and revise, while maintaining a patient database.
Results: The implementation of our proposed framework on the independent testing set has demonstrated promising results. It can accurately identify lesion locations from clinical charts consistently across all evaluated cases via LLM. It achieves a mean Dice score of 0.86 and 95th percentile Hausdorff distance (HD95) of 4.93mm for clinical target volume (CTV) segmentation.
Conclusion: Our proposed framework enables artificial intelligence-powered target volume localization and delineation for spine SBRT while providing comprehensive database management. Preliminary results demonstrate promising performance, indicating its potential to streamline the spine SBRT workflow.