[OA045] Robustness of textural features in ADC magnetic resonance imaging maps among different contouring for prostate cancer patients. (August 2018)

• Record Type: Journal Article
• Authors: Menghi, Enrico; Feliciani, Giacomo; Ravaglia, Valentina; Mazzotti, Giorgio; Scabbia, Francesca; Barbieri, Simone; Parenti, Giancarlo; Duca, Massimo Del; Sarnelli, Anna
• Is Part Of: Physica medica. Volume 52(2018)Supplement 1
• Page Start: 19
• Publication Date: 2018-08
• Subjects: Medical physics -- Periodicals; Biophysics -- Periodicals; Biophysics -- Periodicals; Imagerie médicale -- Périodiques; Radiothérapie -- Périodiques; Rayons X -- Sécurité -- Mesures -- Périodiques; Physique -- Périodiques; Médecine -- Périodiques; 610.153
• DOI: 10.1016/j.ejmp.2018.06.117
• Languages: English
• ISSNs: 1120-1797

Abstract

Abstract : Purpose: Among Radiomics studies MR ones are the most challenging. This is due to variability in acquisition and reconstructions methods. In this study we proof robustness of features in prostate cancer patients which predict gleason score from ADC maps employing 5 different contouring methods. Methods: In our study 25 patients diagnosed with prostate cancer from biopsy and staged by gleason score were enrolled for MR scan and subsequent radiotherapy. A DW echo planar imaging sequence with b-values of 0–1000 s/mm 2, and TR/TE = 3000/120 ms provided data for Apparent Diffusion Coefficient (ADC) maps. 41 lesions were contoured and labeled according to respective patient gleason score. Radiomics features extracted from 2 manual contours and 3 thresholds contouring algorithm (range 120–140% of the minimum value) were correlated with gleason score. A predictive model was developed by selecting features able to discriminate gleason 6 vs higher grade lesions among all the 5 contouring styles. Results: Correlation with gleason 6 versus higher grade was found for one feature for the manual contouring. Four more feature agreed with the main endpoint employing the automatic contouring in the threshold range. A model based on the feature selected by manual contouring scored an AUC of 0.79 whereas the model based on threshold algorithm an AUC of 0.85. Conclusions: Our study underlined the critical importance in the choice of contouring style. Agreement of 4 feature correlation with gleason score was confirmed for semi automatic algorithm in a wide threshold range scoring a high predictive power. Validation with 8 external patient is ongoing.