Exploratory analysis of using supervised machine learning in [18F] FDG PET/CT images to predict for recurrence and suvival in cervical cancer

387 Aim: The aim of this study was to identify relevant features on 2-deoxy-2-(18F)fluoro-D-glucose PET/CT ([18F] FDG-PET/CT) to predict for recurrence (R) and overall survival (OS) in cervical cancer patients. Methods: 63 treatment naive cervical cancer patients, who had a positive [18F] FDG-PET/CT from 12/2008 to 12/2015 were included in this analysis. The primary tumours were delineated on the PET images using semi-automatic VOIs, followed by feature extraction. Each tumour was characterized by 118 features including in vivo intensity, histogram, shape, textural and joint fusion features. Identification of highly-correlating features was performed by the utilization of ensemble machine learning approaches in a multi-fold training scheme. Overall 150 Monte Carlo (MC) folds were established. In each MC fold 80% of the original data was randomly selected. In each MC fold 8 machine learning (ML) exploratory analysis was performed as presented in Papp et al. The individual datasets for these ML executions was selected from the given MC subset by bootsrapping. The final feature weights were determined by averaging the 1200 (150x8) weights determined by ML. Results: In the studied cohort 22 patients had a recurrence, 12 died. Mean time to treatment failure (TTF) was 14.3 months (range: 0-73 mo) and mean OS was 40.6 mo (range: 0-100 mo). The three highest weighted parameters were the CT-based textural features Low gray level zone emphasis (GLZSM; 0.083) and Small zone low gray emphasis (GLZSM; 0.080) as well as the joint fusion features Sum entropy (0.057) when predicting recurrence. For survival prediction the three highest weighted parameters were CT-based textural features maximum probability and Sum entropy of Gray-level co-occurrence matrix (GLCM-MP: 0.179; GLCM-SE: 0.10), as well as the PET-based minimum intensity feature (0,057). Conclusions: These preliminary results of our exploratory analysis demonstrate that textural and joint fusion features obtained by supervised ML are a valuable option for predicting recurrence and overall survival in cervical cancer. However further analysis with a bigger patient population is needed and still ongoing.