2990 - Artificial Intelligence and Radiomic-Based Predictors of Response to Neoadjuvant SABR in Early-Stage Breast Cancer: Insights from the SPORT-DS Trial

03:00pm - 04:00pm PT

Hall F

Screen: 11

POSTER

Presenter(s)

Felipe Restini, MD - McGill University Health Center, Montreal, QC

F. Restini¹, T. Hijal², P. Vavassi³, G. Jarry⁴, E. Natier⁵, C. Lambert⁶, and M. Yassa⁷; ¹Department of Radiation Oncology, McGill, Montreal, Canada, ²McGill University Health Centre, Division of Radiation Oncology, Montreal, QC, Canada, ³Department of Radiation Oncology, CR-HMR, HMR - Hôpital Maisonneuve, Montreal, QC, Canada, ⁴Hopital Maisonneuve-Rosemont, Montreal, QC, Canada, ⁵Radiation Oncology MUHC, Montreal, QC, Canada, ⁶McGill University Health Centre, Department of Radiation Oncology, Montreal, QC, Canada, ⁷CIUSSS de L'Est-de-I'lle-de Montreal Hopital Maisonneuve-Rosemont, Montreal, QC, Canada

Purpose/Objective(s):

Recent trials, such as SPORT-DS and ABLATIVE, suggest single-fraction SABR with delayed surgery might be a potential new indication for early-stage breast cancer in the near future. SPORT-DS results are promising, with 84.6% of patients showing a strong response (99% median tumor reduction), while 15.4% of patients had no response. The reason for this variability remains unclear, remaining unexplainable based on clinical features. This study aims to identify planning CT patterns by extracting radiomic features (RadF) that may explain this dichotomous response and refine patient selection for SABR.

Materials/Methods:

We analyzed clinical data and planning images from SPORT-DS patients, classifying them as responders (R) or non-responders (NR) based on pathological response. RadFs were extracted and integrated with clinical data. Using Recursive Feature Elimination (RFE), we identified the 20 most relevant features. Their correlation with the binary outcome was then assessed using T-test, Mann-Whitney U test, and chi-square test, with False Discovery Rate (FDR) correction applied for multiple comparisons.

Results: Out of 13 patients, two exhibited 0% cellular response and were classified as NR, while 11 had responses of 99%. The GTV had a mean volume of 0.89 cm³ (95% CI: 0.36–1.42 cm³), while the PTV eval averaged 71.67 cm³ (95% CI: 60.19–83.15 cm³). The D95 PTV had a mean of 19.73 Gy (95% CI: 19.57–19.89 Gy), and the Dmax PTV was 21.39 Gy (95% CI: 21.14–21.64 Gy).

No significant correlation was observed between clinical features and the outcome, GTV (p = 0.091), PTV eval (p = 0.608), Dmax PTV (p = 0.289), Dmin PTV - D99 (p = 0.065), D95 PTV (p = 0.056), and V95 PTV (p = 0.397). However, after extracting nearly 1.800 RadF from each case. The FDR correction to control for Type I error found 12 RadF significantly associated with NR (p < .05). They were LargeDependenceEmphasis (p = 0.025), LongRunEmphasis (p = 0.032), RunLengthNonUniformity (p = 0.012), RunLengthNonUniformityNormal (p = 0.025), RunPercentage (p = 0.010), RunVariance (p = 0.032), ShortRunEmphasis (p = 0.017), RunLengthNonUniformity (p = 0.032), LargeDependenceEmphasis (p = 0.015), RunPercentage (p = 0.015), ShortRunEmphasis (p = 0.012), and DependenceVariance (p = 0.012). All identified features belong to Gray Level analysis, which quantifies textural patterns within the tumor volume.

Conclusion:

Despite the limited number of events, our rigorous statistical analysis identified RadF associated with poor response to single-fraction neoadjuvant SABR in early-stage breast cancer within the SPORT-DS trial. As this approach emerges as a potential new indication for SBRT, integrating Artificial Intelligence could enhance patient selection and treatment personalization. Further validation with larger datasets is needed, and this model will be tested in the ongoing SPORT-DNS trial, which employs the same single fraction with an extended delay before surgery.