Novel Dose Criteria for Lung Cancer SBRT to Improve Local Control in Patients Treated to 50 Gy/5 Fractions Using Deep Learning Methods and Explainability Techniques

Abstract

To some radiation oncologists, 50 Gy/5 fractions has been considered controversial, as they feel the nominal BED of 100 Gy might be too low for long-term local control of some lesions. We analyzed a large cohort of these patients using a deep learning model to predict local recurrence (LR) and used explainability techniques to extract new dose features important to the model's prediction. Subsequently, we determined optimal cut-points for the most significant metrics to provide actionable criteria for treatment planning in these patients. A total of 535 SBRT lung cancer patients treated between 2009 and 2017 were retrospectively analyzed using a deep learning approach. All patients had NSCLC and all of them were treated with 50 Gy in 5 fractions (100 Gy BED, α/β = 10). Mean clinical maximum tumor diameter was 2.2 cm. There were 31 LR in the dataset with mean follow-up time of 28 months. Mean age was 75 years. CT images, 3D dose distribution and patient demographic details were used to train a deep learning survival model to predict time to failure and probability of local control. Validation, training, and testing were in accordance with TRIPOD criteria. 80 % of the data were used for 5-fold cross-validation (10 iterations) and 20 % was held for independent testing. The Grad-CAM method was applied to identify regions of the dose distribution that are the most significant to the model's decision-making. Based on the results, appropriate dose metrics were proposed, and optimal cut-points were determined to distinguish between lower and higher LR-risk patients. The model has an acceptable performance (c-index: 0.72, 95% CI: 0.68-0.75); the testing c-index was 0.69. Grad-CAM showed that the model's spatial attention was mostly concentrated in the tumor's "PTV-GTV" region. Statistically significant criteria are in Table 1. A novel deep learning model for prediction of LR, incorporating 3D dose data, CT images and patient demographics, was developed and tested. Grad-CAM demonstrated superior significance of peripheral (PTV-GTV) dose features. Subsequently determined optimal cut-points have significant prognostic power (log rank, p<0.001) and could be used as additional criteria in treatment planning. While these data have repercussions in treatment planning, they do not suggest that a significantly higher BED for the prescription dose is necessary for tumor control in NSCLC. Nevertheless, it might be effective to slightly elevate the prescribed dose, i.e., from 100 Gy BED to 104 Gy BED.