ESTRO 2024 - Abstract Book

S573

Clinical - Breast

ESTRO 2024

RT. The Clinical Target Volume (CTV) segmentations were available from RT plans and CTV Hounsfield Units (HU) histograms were extracted using a Python code. Patients with a silicon implant were excluded. Only the range [- 200HU, 70 HU] was investigated to avoid clips and other high-density materials. Parameters as volume, mean/median HU, SD, kurtosis, skewness, 10th/25th/75th and 90th percentiles etc were extracted. Fatty and fibroglandolar tissues were defined with the range [-200HU, -100HU] and [-100HU, 50HU] as reported in literature [1]. Absolute and percentual volume of fatty and fibroglandolar tissues together with other histogram based parameters were also extracted. Based on a previously validated methodology, the extracted parameters were combined into Multivariate Cox Regression (MCR) using minimum redundancy on the entire population due to the limited number of events despite the large cohort. An internal validation procedure was performed by repeating the MCR fit for 1000 data sets obtained by bootstrapping the original cohort. Median and inter-quartile (IQR) ranges of p-values, hazard ratios (HRs) and AUC values obtained by the procedure were reported and compared with the results from the original cohort, as a measure of the results’ robustness. A prognostic index (PI) of risk was computed as linear combination of variables and their b coefficients for each model and Kaplan-Meier test using the ROC Youden index as criterion was used to stratify patients with higher and lower risk.

Results:

With a median follow-up of 6 years, IQR=[4, 8], the LP/DP/death rates were 2.3%/4.1%/7.0% with 26/46/80 events out of 1127 patients respectively.

The resulting Cox model for LPFS included two parameters: percentual volume of fat (VFAT%) and 10th percentile of HU-distribution (P10) (p=0.03, IQR=[0.001, 0.05], AUC=0.62, IQR=[0.60, 0.65]), showing worse outcomes for patients with lower VFAT% and P10. The Kaplan-Meier test applied to the computed PI-index (PI_LPFS) stratified patients according to their risk of LP with very high ability (p=0.003, HR=5.91, IQR=[1.81, 19.26]) as shown in Figure 1a. The resulting Cox model for DPFS included two parameters: Median and Kurtosis of HU-distribution (p=0.08, IQR=[0.002, 0.12], AUC=0.60, IQR=[0.55, 0.62]), showing worse outcomes for patients with lower Median and Kurtosis. The Kaplan-Meier test applied to the computed PI-index (PI_DPFS), stratified patients well enough according to their risk of DP (p=0.026, HR=1.93, IQR=[1.08, 3.54]) as shown in Figure 1b. Interestingly the median value of total distribution was -102HU, IQR=[-114HU, -75HU] showing that patients with median values lower than the higher HU-threshold of fat region were prone to have a better prognosis in distant relapse. The resulting Cox model for OS included two parameters: percentual volume of fat (VFAT%) and 75th percentile of HU-distribution (P75) (p=0.0002, IQR=[0.00004, 0.003], AUC=0.63, IQR=[0.60, 0.66]), showing worse outcomes for patients with lower VFAT% and P75. The Kaplan-Meier test applied to the computed PI-index (PI_OS), stratified patients according to their risk of death (p<0.0001, HR=3.02, IQR=[1.83, 4.96]) as shown in Figure 1c. Interestingly, adding the patient age as variable, the AUC increased, as expected, up to 0.70, IQR=[0.67, 0.72], without changing the impact of the VFAT% to the model showing that age and VFAT% are independently and strongly associated to OS (p<0.0001), unveiling reduced survival for patients with lower VFAT% and higher age. As expected, the computed PI-index (PI_OS_age_corrected) in Kaplan-Meier test better stratified patients according to their risk of death (p<0.0001, HR=5.76, IQR=[3.42, 9.71]) as shown in Figure 1d. In Figure 2, as example, are reported one-hundred HU histograms of patients, with the fatty and fibroglandolar regions separated by dot lines.