ESTRO 2024 - Abstract Book

S5087

Physics - Radiomics, functional and biological imaging and outcome prediction

ESTRO 2024

features (RFs) Morphology_COMshift (MCS) and Statistical_percentile10 (SP10) were extracted from 18F Fluorodeoxyglucose (18F-FDG) PET images from two different scanners also used in the referenced study (Discovery ST, Discovery-STE, General Electric Medical Systems, Milwaukee, WI, USA) using the IBSI consistent SPAARC software [2] (with the same procedure of the referenced article, i.e.: voxel size 3mmx3mmx3mm, bilinear interpolation). The RFs were used in a Cox model to predict distant progression. Model coefficients were MCS: −0.3479, SP10: 0.0001222. We predicted the prognostic index (P-index) of the model (linear combination of RFs and their respective model coefficients) and tested if it was statistically significant (p value test). Distant progression time was computed as the difference between the clinical finding of distant progression and the date of the PET image used for radiomics features extraction. Afterwards we trained a new model with the same RFs, plotting the Kaplan-Meier curve, dividing the population into two sub-populations (high risk and low risk). High risk individuals were the ones having a P-index higher than the median value of the population. The distribution of RFs of the 36 patients was also compared to the one of the 176 patients’ study, in order to assess whether the two populations have different distributions or not.

Results:

The validation of the model showed no statistical significance (Table 1: b = -0.22, p = 0.61). However, training a new Cox model with the two RFs shows that MCS was still predictive (p value of MCS = 0.02), with good separation between high risk and low risk individuals (Figure 1). The resulting best-fit value of HR (i.e.: Exp(b) in Table 2) showed a value very near to the one obtained in the original model (2.72 vs 2.53). Additionally, features distributions in Figure 2 are similar in terms of median and shape (excluding outliers), considering that the 36 patients’ dataset has a small population. Moreover, histogram shapes are very similar, e.g.: all SP10 histograms show a two-peaked distribution