EJNMMI Physics (Sep 2022)

Principal component analysis of texture features derived from FDG PET images of melanoma lesions

  • DeLeu Anne-Leen,
  • Sathekge Machaba,
  • Maes Alex,
  • De Spiegeleer Bart,
  • Beels Laurence,
  • Sathekge Mike,
  • Pottel Hans,
  • Christophe Van de Wiele

DOI
https://doi.org/10.1186/s40658-022-00491-x
Journal volume & issue
Vol. 9, no. 1
pp. 1 – 10

Abstract

Read online

Abstract Background The clinical utility of radiomics is hampered by a high correlation between the large number of features analysed which may result in the “bouncing beta” phenomenon which could in part explain why in a similar patient population texture features identified and/or cut-off values of prognostic significance differ from one study to another. Principal component analysis (PCA) is a technique for reducing the dimensionality of large datasets containing highly correlated variables, such as texture feature datasets derived from FDG PET images, increasing data interpretability whilst at the same time minimizing information loss by creating new uncorrelated variables that successively maximize variance. Here, we report on PCA of a texture feature dataset derived from 123 malignant melanoma lesions with a significant range in lesion size using the freely available LIFEx software. Results Thirty-eight features were derived from all lesions. All features were standardized. The statistical assumptions for carrying out PCA analysis were met. Seven principal components with an eigenvalue > 1 were identified. Based on the “elbow sign” of the Scree plot, only the first five were retained. The contribution to the total variance of these components derived using Varimax rotation was, respectively, 30.6%, 23.6%, 16.1%, 7.4% and 4.1%. The components provided summarized information on the locoregional FDG distribution with an emphasis on high FDG uptake regions, contrast in FDG uptake values (steepness), tumour volume, locoregional FDG distribution with an emphasis on low FDG uptake regions and on the rapidity of changes in SUV intensity between different regions. Conclusions PCA allowed to reduce the dataset of 38 features to a set of 5 uncorrelated new variables explaining approximately 82% of the total variance contained within the dataset. These principal components may prove more useful for multiple regression analysis considering the relatively low numbers of patients usually included in clinical trials on FDG PET texture analysis. Studies assessing the superior differential diagnostic, predictive or prognostic value of principal components derived using PCA as opposed to the initial texture features in clinical relevant settings are warranted.

Keywords