Prediction of Regional and Distant Failure After Definitive Thoracic Stereotactic Body Radiation Therapy (SBRT) Using Pre-Treatment CT-Based Radiomic Analysis

Abstract

<h3>Purpose/Objective(s)</h3> Patients with early-stage non-small cell lung cancer (NSCLC) have the option of either lobectomy or definitive SBRT, both of which provide excellent local control. However, a tangible portion still suffer regional failure (RF) in the lymph nodes (LNs), or distant failure (DF) outside the thorax. Those receiving SBRT are at a further disadvantage as a LN dissection can identify occult nodal spread at the time of surgery. Thus predicting the risk of RF/DF, particularly for those treated with SBRT, remains an essential issue. Here we seek to assess whether CT-based radiomics (quantitative, sub-visual cues) could provide such insight. <h3>Materials/Methods</h3> We identified patients who received definitive lung SBRT between 2014-2019, had a pre-treatment CT chest (CT1) within 3 months prior, and had at least 12 months of follow-up. Charts were assessed for RF and DF. A pre-trained U-Net model was used to segment right and left lungs from CT1 for radiomic feature extraction. 304 radiomic signatures (Gabor wavelet, Haralick, and CoLlAGe) were extracted, filtered, and used to train Linear Discriminant Analysis (LDA) machine learning classifiers for RF and DF prediction tasks. Features from the ipsilateral (Ip) and contralateral (Con) lung were analyzed separately to assess for their comparative utility. Leave-one-out cross validation method was utilized. To serve as a control, a baseline LDA classifier was trained using canonical clinical features (Clin) including performance status, tumor stage, age, and tumor location. <h3>Results</h3> 89 stage I-II NSCLC patients with 90 lesions who received definitive SBRT were included. Median dose was 48 Gy (range: 40-54 Gy) in 3-5 fractions, prescribed to the 100% isodose line. RF was identified in 12 (13%) lesions, while DF was found in 21 (23%) lesions. Radiomic features significantly outperformed the baseline approach in both RF (AUC=0.65 vs. 0.46) and DF (AUC=0.66 vs. 0.57). There was no significant difference between features from the ipsilateral vs. contralateral lung. <h3>Conclusion</h3> CT-based radiomic analysis predicted RF and DF more accurately than clinical features alone. The similar performance between the ipsilateral and contralateral lung suggests there may be features intrinsic to the parenchyma contributing to disease recurrence. Further studies should assess whether different ROIs, including the hilum/mediastinum or peritumoral region, are more informative.