PT  - JOURNAL ARTICLE
AU  - PIERPAOLO ALONGI
AU  - Alessandro Stefano
AU  - Albert Comelli
AU  - Riccardo Laudicella
AU  - Stefano Barone
AU  - Giorgio Russo
TI  - New Artificial intelligence model for 18F-Choline PET/CT in evaluation of high-risk prostate cancer outcome: texture analysis and radiomics features classification for prediction of disease progression
DP  - 2020 May 01
TA  - Journal of Nuclear Medicine
PG  - 1303--1303
VI  - 61
IP  - supplement 1
4099  - http://jnm.snmjournals.org/content/61/supplement_1/1303.short
4100  - http://jnm.snmjournals.org/content/61/supplement_1/1303.full
SO  - J Nucl Med2020 May 01; 61
AB  - 1303Aim: The aim of the study was to investigate the potential application of texture analysis of Cho-PET/CT images in prostate cancer and to propose a model incorporating a new machine-learning radiomics model to select PET imaging features able to predict disease progression in prostate cancer (PCa) in patients with same class of risk at re-staging. Materials and Methods: We retrospectively analyzed 88 high-risk PC patients who underwent restaging Cho-PET/CT after first-line therapy. Follow-up data about clinical, laboratory and radiological exams were recorded for a time of 24 months after PET. PET studies and related structures containing volumetric segmentation using an active contour method were imported in LifeX toolbox to extract 51 imaging features from each lesion (primary or local relapse, nodal or bone metastasis). In addition, PSA and Gleason score features were considered (51+2=53 features). Due to the redundancy, heterogeneity and uncertainty of the information represented by radiomics features, a novel statistical system based on correlation matrix and point-biserial correlation coefficient has been implemented for feature reduction and selection, while Discriminant Analysis (DA) was used as a method for feature classification in a whole sample (all lesions in 88patients) and in sub-group analysis for T, N and M. Results: The most Cho-PET avid lesions (for T, N and M=128 lesions) were examined for texture analysis in the whole group (N=88): 53 features were extracted for each lesion for a total of 6784. After statistical reduction and selection, 4 features (SUVmin; HIST=_Energy Uniformity; GLRLM_SRLGE; GLZLM_SZLGE) resulted able to discriminate the occurrence of disease progression at follow-up, obtaining the best performance in DA classification and improving the FU status assessment ( Sensitivity 72%, Specificity 68%, and Accuracy 69%). In the sub-group analysis the same methods demonstrates the best performance in DA classification for T (N=38) in 3 features (SUVmin; SHAPE_Sphericity; GLCM_Correlation= Sensitivity 91,6%; Specificity 84%; Accuracy 87%)) and for N (N=44) in 2 features (HIST=_Energy Uniformity; GLZLM_SZLGE= Sensitivity 68%; Specificity 91% Accuracy 83%). No results were obtained for M due to an unbalanced number of cases with disease progression compared to stable disease. Conclusions: the presented artificial intelligence model demonstrated to be feasible and able to select Cho-PET features for T and N with valuable association with High-Risk PCa patients’ outcomes through an accurate stratification of patients with disease progression.