ESTRO 2024 - Abstract Book

S4448

Physics - Machine learning models and clinical applications

ESTRO 2024

Material/Methods:

The patient dataset consisted of 478 NSCLC early stage patients treated with SBRT (VMAT-323/IMRT-155) between 2010 and 2017. A machine learning model, incorporating various neural networks for multi-modal input data (planning CT images, planning dose distribution and patient clinical/demographic details), was built and trained in accordance with TRIPOD criteria, and a Grad-CAM method was applied to show which part of the input data is most relevant for decision-making. Subsequently, Cox regression was applied and the most significant cut-points were tested using Kaplan-Meier (KM) analysis and log-rank tests. The cut-points were studied in patient subsets stratified according to different confounding variables, such as treatment technique, tumor size (i.e. staging) or tumor shape, which helped to reproduce and understand previous results in this area.

Results:

The outcome model exhibited an acceptable performance (c-index: 0.60±0.02), considering the event imbalance (DM/DM-free = 91/387) and limited number of training data points. The grad-CAM showed a superior importance of dose outside the PTV, which aligns with the findings from previous studies. In the Cox regression, dosiomic features in a 3 cm thick shell around the PTV (PTV3cm), treatment technique and tumor size/shape were identified as the strongest predictors of DM (p<0.001). However, the relationship between the mean dose to PTV3cm and risk of DM was opposite to Diamant’s results and in agreement with Lalonde’s findings (HR<1, p<0.001), who also used a dataset of mixed (IMRT/VMAT) treatment techniques. The treatment technique showed a significant correlation with the mean dose to PTV3cm, thus, it was found as a highly confounding factor that requires accounting for. After stratification by treatment technique, there was no significant correlation between the mean dose to PTV3cm in IMRT/VMAT patients and risk of DM. The same results were reported by Hughes et al., who also analyzed only patients treated with one treatment technique (VMAT). Finally, the VMAT cohort of patients was further stratified according to tumor sphericity, excluding all patients with tumor sphericity below 0.5. The analysis of this sub-cohort showed a significant correlation of the mean dose to PTV3cm with the risk of DM (HR>1, p=0.004). The same was reported by Diamant et al. [1], who analyzed patients with generally smaller tumors volumes (median: 18.5 cm3, IQR: 5.7-51.5 cm3). The median tumor volume in our dataset was 32.1 cm3 (IQR: 19.7-56.7 cm3), though it was comparable with Diamant’s dataset after stratification by treatment technique and tumor sphericity (median: 18.9 cm3, IQR:7.9-40.3 cm3).

Conclusion:

In this work we have shown that the main reason of conflicting conclusions in previously published studies was inconsistent datasets and insufficiently considered confounding variables. This work reproduced results of all previous publications by properly stratifying the patient dataset according to treatment technique and tumor size/shape. In conclusion, there is no general correlation between the risk of DM and dose outside the PTV. However, the mean dose to PTV3cm can be a significant predictor of DM in small spherical tumor volumes treated with VMAT, which might clinically imply considering larger PTV margins for smaller and more spherical tumors (e.g., if GTV > 2 cm, then margin ≤ 7 mm, else margin>7 mm).

Keywords: Non-small Cell Lung Cancer, Distant Metastasis

References: