ESTRO 2021 Abstract Book

S1419

ESTRO 2021

10 for validation, and 10 for testing. Our proposed method consists of three main steps: (1) the first step is using the Greedy Closed Principal Curve method (GCPC) to obtain the data sequences by using a few points of prostate ROI as the approximate initialization, (2) the second step is using the Memory-based Adaptive Differential Evolution (MADE) to obtain the initial optimal parameters of the Caputo Fractional-order Backpropagation Training algorithm (CFBT), and (3) we use the CFBT’s parameters to denote the interpretable mathematical expression of prostate contour. Results As shown in Fig. 1 and Fig. 2, we observed a high correlation between automatic and manual contours, where the average results were obtained for the prostate with a Dice Similarity Coefficient (DSC) of 96.5%, Jaccard Similarity Coefficient (Ω) of 95.1%, and Accuracy (ACC) of 96.3% on 10 testing patients. Furthermore, even the Salt and pepper noise’s Signal Noise Rate SNR=0.6, the DSC, Ω, and ACC of the proposed method are as high as 91.2%, 90.1%, and 91.1%, respectively. The results show that our method has good robustness, where the DSC, Ω, and ACC fluctuate about 5.5% at most during SNR decreases from 1 to 0.6

.