Sep 30

QP 10 - DHI 2: Quick Pitch: The Digital Revolution in Radiation Oncology: AI Models for Enhanced Patient Care

1056 - Vision-Language AI Model for Detecting PET/CT-Occult Lymph Node Metastasis in Early-Stage NSCLC Treated with SABR to Prevent Regional Recurrence

01:00pm - 01:05pm PT

Room 20/21

Presenter(s)

Hui Xu, PhD - The University of Texas MD Anderson Cancer Center, Houston, TX

H. Xu¹, X. Xu^1,2, K. Zhang¹, J. Lin³, M. B. Saad¹, G. Eapen³, J. Zhang⁴, D. L. Gibbons⁵, J. Heymach⁴, A. A. Vaporciyan⁶, J. Roth⁶, R. Mehran⁶, P. Balter⁷, J. M. Pollard⁷, D. C. Qian⁸, S. H. Lin², S. Gandhi², Z. Liao⁸, J. Wu¹, and J. Y. Chang²; ¹Department of Imaging Physics, The University of Texas MD Anderson Cancer Center, Houston, TX, ²Department of Radiation Oncology, The University of Texas MD Anderson Cancer Center, Houston, TX, ³Department of Pulmonary Medicine, The University of Texas MD Anderson Cancer Center., Houston, TX, ⁴Department of Thoracic/Head and Neck Medical Oncology, The University of Texas MD Anderson Cancer Center, Houston, TX, ⁵Department of Thoracic-Head & Neck Medical Oncology, The University of Texas MD Anderson Cancer Center, Houston, TX, ⁶Department of Thoracic and Cardiovascular Surgery, The University of Texas MD Anderson Cancer Center, Houston, TX, ⁷Department of Radiation Physics, The University of Texas MD Anderson Cancer Center, Houston, TX, ⁸Department of Thoracic Radiation Oncology, The University of Texas MD Anderson Cancer Center, Houston, TX

Purpose/Objective(s):

Although PET/CT has improved the accuracy of staging, occult lymph node metastasis (OLNM) remains a challenge, often leading to regional recurrence in early-stage NSCLC patients treated with stereotactic ablative radiotherapy (SABR). Herein, we introduce LymphDetect, a vision-language model designed to predict OLNM and regional recurrence in this clinical setting.

Materials/Methods:

1,428 pathologically confirmed, chest CT and PET or PET/CT staged NSCLC patients (T = 7cm, N0M0) were analyzed. The discovery cohort (n = 569) included patients who underwent surgical resection and lymph node dissection or endobronchial ultrasound (EBUS), with 122 of 569 patients (21%) diagnosed with OLNM. Tumor descriptors, including standard radiomics, regional variation, and lung vessel characteristics, were extracted from high-quality CT images to characterize the tumor morphology, texture, and its surrounding microenvironment. Beyond that, deep embedding features from our in-house CT foundation model were included, to capture additional underlying imaging patterns. A vision-language architecture was developed using radiology reports to contextualize the predictive model’s learning, with Bidirectional Encoder Representations from Transformers (BERT) applied on raw radiology reports. The LymphDetect model was constructed on the discovery cohort, and its performance was evaluated through a five-fold independent testing scheme. Further, three independent cohorts, including pre-registered SABR program (n = 645), prospective I-SABR (n = 135) and STARS (n = 79) trials, were used to test its ability to predict time to regional recurrence as the first recurrence (TTRR) and time to cumulative regional recurrence (TTCRR).

Results:

During model construction and testing, LymphDetect achieved a mean area under the curve (AUC) of 0.67 (range: 0.62-0.72) for OLNM prediction, outperforming classical machine learning methods such as Logistic Regression (AUC: 0.59 ± 0.07), Random Forest (AUC: 0.61 ± 0.04), and XGBoost (AUC: 0.59 ± 0.05), as well as the tabular foundation model TabPFN (AUC: 0.61 ± 0.04). In the SABR program, 35% (226 of 645) of patients were identified as high-risk for regional recurrence, with significantly worse TTRR (HR: 1.9, 95% CI: [1.1-3.1], p = .012) and TTCRR (HR: 1.8, 95% CI: [1.2-2.9], p = .009) compared to the low-risk group. In the I-SABR and STARS trials, 16% and 35% of patients were flagged as high risk, respectively, with worse TTRR, (I-SABR: HR: 2.9, 95% CI: [0.9-9.8], p = .063; STARS: HR: 10, 95% CI: [1.3-78], p = .007) and worse TTCRR ( STARS: HR: 9.6, 95%CI: [1.2-75], p = .008). The vision-language model also outperformed models fitted on a single modality, demonstrating its advantage in integrating imaging and textual data.

Conclusion:

The vision-language LymphDetect model demonstrated promise in predicting OLNM in early-stage NSCLC treated with SABR, and if validated, could serve as a powerful tool to tailor the treatment for reducing the risk of regional recurrence.