Fuzzy Radiomics: A novel approach to minimize the effects of target delineation on radiomic models

Ziel/Aim:

The characterization of tumours with radiomics from PET imaging has become fashionable. However, the effect of different lesion delineation approaches on radiomic variations has beed widely discussed. Our goal was to integrate fuzzy logics into the process of radiomics feature extraction in order to handle uncertainties arising from variations in lesion delineation.

Methodik/Methods:

Reconstructed NEMA IQ PET images acquired in 13 PET/CT systems with the same physical phantom [1] were included in our study. The largest sphere (37 mm) was delineated with two automated approaches: the first method built on dichotomized fuzzy clustering, while the second method applied gradient maximization. In addition, a fuzzy delineation was executed, which provided a non-binary probability mask for the lesions. Twelve radiomic features with known minimal multi-centric variations [2] were extracted from the 37 mm sphere of the 13 PET acquisitions with the two binary as well as with the fuzzy masks. The binary masks were used for classic radiomics that considered binary membership of voxels, while the fuzzy mask was used for modified radiomics calculations, handling non-binary probability membership values. Coefficients of variation (CV) were calculated for each feature and delineation method over the 13 PET acquisitions. Finally, the mean feature CVs across features were calculated to describe the average multi-centric variation of each delineation method.

Ergebnisse/Results:

The average CV of the fuzzy mask with fuzzy radiomics was 21%, while the average CV for the dichotomized fuzzy clustering and gradient maximization approaches were 37% and 63%, respectively.

Schlussfolgerungen/Conclusions:

We demonstrated that performing fuzzy radiomics on fuzzy probability maps can minimize variations of radiomic features in a multi-centric environment, thus, pointing towards reproducible quantitative radiomics.

Literatur/References:

[1] Rausch I, et al; Variation of system performance, quality control standards and adherence to international FDG-PET/CT imaging guidelines. A national survey of PET/CT operations in Austria. Nuklearmedizin 2014;53(6):242 – 248

[2] Papp L, et al:Optimized feature extraction for radiomics analysis of 18F-FDG-PET imaging. JNM 2018 (in press).