Open AccessAssociation of visual and quantitative heterogeneity of 18F-FDG PET images with treatment response in locally advanced rectal cancer: A feasibility study
Author(s) -
Paula Martin-Gonzalez,
Estibaliz Gomez de Mariscal,
María Elena Martino,
Pedro M. Gordaliza,
Isabel Peligros,
José Carreras,
Felipe A. Calvo,
Javier Pascau,
Manuel Desco,
Arrate MuñozBarrutia
Publication year - 2020
Publication title -
plos one
Language(s) - English
Resource type - Journals
SCImago Journal Rank - 0.99
H-Index - 332
ISSN - 1932-6203
DOI - 10.1371/journal.pone.0242597
Subject(s) - logistic regression , medicine , multivariate statistics , area under the curve , receiver operating characteristic , colorectal cancer , artificial intelligence , cancer , nuclear medicine , oncology , statistics , computer science , mathematics
Background and purpose Few tools are available to predict tumor response to treatment. This retrospective study assesses visual and automatic heterogeneity from 18 F-FDG PET images as predictors of response in locally advanced rectal cancer. Methods This study included 37 LARC patients who underwent an 18 F-FDG PET before their neoadjuvant therapy. One expert segmented the tumor from the PET images. Blinded to the patient´s outcome, two experts established by consensus a visual score for tumor heterogeneity. Metabolic and texture parameters were extracted from the tumor area. Multivariate binary logistic regression with cross-validation was used to estimate the clinical relevance of these features. Area under the ROC Curve (AUC) of each model was evaluated. Histopathological tumor regression grade was the ground-truth. Results Standard metabolic parameters could discriminate 50.1% of responders (AUC = 0.685). Visual heterogeneity classification showed correct assessment of the response in 75.4% of the sample (AUC = 0.759). Automatic quantitative evaluation of heterogeneity achieved a similar predictive capacity (73.1%, AUC = 0.815). Conclusion A response prediction model in LARC based on tumor heterogeneity (assessed either visually or with automatic texture measurement) shows that texture features may complement the information provided by the metabolic parameters and increase prediction accuracy.