Abstract
BACKGROUND AND PURPOSE: The risk of genitourinary late effects is a major dose-limiting factor in radiotherapy for prostate cancer. By using shape analysis and machine learning, the aim of this study was to evaluate whether bladder shape descriptors from the first week of treatment could identify patients experiencing genitourinary late effects. MATERIAL AND METHODS: From a cohort of 258 prostate cancer patients treated with daily cone-beam computed tomography (CBCT)-guided radiotherapy (prescription doses of 77.4-81.0 Gy), 7 pre-treatment asymptomatic cases experiencing RTOG genitourinary late effects ≥Grade 2 and 21 matched controls were selected. The bladder was manually contoured on each CBCT, and a 17-D vector comprising shape descriptors was used for patient clustering, focusing on bladder contours from the first week of treatment. ANOVA was used to test statistical significance of descriptors across and within clusters. RESULTS: Of the contours from the first week of treatment, 84 % could be classified in two main clusters with distinct bladder shape characteristics. This cluster stratification remained identical when bladder contours from the entire course of treatment were used. Convexity, elliptic variance and compactness were significantly different between patients with vs. without genitourinary late effects ≥Grade 2 (p < 0.05). Dice Coefficients between predictive models using descriptors of the first week and the voxels' probability of belonging to the bladder were above 93 ± 6 % (median ± interquartile range). CONCLUSION: Bladder shape descriptors in the first week of treatment showed potential to predict the risk of developing genitourinary late effects after radiotherapy for prostate cancer.