Heterogeneity analysis of 18 F-FDG PET imaging in oncology: clinical indications and perspectives

Abstract

Purpose

To evaluate the performances and perspectives of heterogeneity analysis of FDG PET imaging in oncology.

Methods

We performed a review of the literature using PubMed/Medline and Google scholar, with multiple research keywords for each organ accompanying the terms ''radiomics'', ''texture'', ''heterogeneity'', ''FDG'', ''PET'' and ''PET/CT''. The review considers shape, histogram and textural analysis of FDG PET. The references of the retrieved articles were also considered to identify additional articles. The search was limited to English language. Preoperative workup exploration, preclinical and animal studies were not included in the review. A total of 128 original articles were considered for the review.

Results

Many publications have explored the use of radiomics for multiple cancer sites, especially lung, head and neck, oesophageal and breast cancers. The results show variable levels of correlation with pathological characteristics and prediction of tumoral staging. Although initial results are promising, few studies have explored PET radiomics for delineating the radiotherapy target volume. Studies have predominantly investigated the prognostic value of radiomics for identifying failure of tumour response to treatment, high risk of recurrence and high mortality. Results are for the most part encouraging although rarely properly validated. The studies also exhibit a great methodological diversity for extracting and analysing features extraction, as various statistical approaches have been used, with the increasing help of machine learning tools.

Conclusion

PET radiomics show their potential to non-invasively improve tumoral characterisation and identification of aggressive pattern complementary to other omics. To be implemented in clinic, these promising results nevertheless need further validation, which must go through a standardisation of their extraction conditions. Machine learning, especially deep learning, could allow the empowerment of the selection of features and the creation of powerful prognostic models including clinical, pathological and iconographic parameters.