ИСТИНА |
Войти в систему Регистрация |
|
ИСТИНА ИНХС РАН |
||
Target delineation is an important step in radiosurgery (RS) treatment planning. Routinely the targets are delineated through slice-by-slice manual segmentation on MR images. This process is time-consuming, operator - dependent and could lead to treatment delays. The aim of this study was to investigate the speed up of the tumor delineation within the radiosurgery treatment planning using contours generated by a deep convolutional neural network (CNN). The MR images of ten patients treated with Gamma Knife RS were selected from routine clinical practice. The dataset consisted of four cases of meningioma, two cases of vestibular schwannoma and four cases of multiple brain metastases. We compared the times needed for two contouring techniques: manual delineation of the tumors and a user adjustment of the CNN generated contours of the tumors. The time spent on each task was recorded. The tasks were performed in Leksell Gamma Plan (version 11.1, Elekta AB) and iPlan (version 4.5, BrainLab) by four experts. The 3D - Unet architecture with residual connections, trained with custom loss function and sampling procedure [Krivov et al, 2018], optimized for metastases segmentation was used for automatic brain tumor segmentation. The automatic contours were generated within five seconds. The time required to import these contours to the treatment planning systems was less than one minute. The generated contours were acceptable with no or minor corrections.The total median time needed to delineate a tumor manually was 9.15 min. (ranged from 3.15 min. to 29.18 min). The median times saved were 6.54 min. (range 40 sec. - 17.06 min.), 2.16 min. (range 48 sec.- 8.20 min.), 9 min. (range 1 min. - 26 min.), 5.27 min. (range 3 min - 17.35 min) for User 1, User 2, User 3, User 4 respectively. The Wilcoxon signed-rank test was used to compare results (p < 0.05, r > 0.6). On average, the automatic algorithm speeds up the process of the delineation in 2.30 times. The usage of deep learning generated contours accelerates delineation more than twofold. Though the automatically generated contours were almost identical to the manual ones, further investigation is needed to quantify these differences and compare it with inter-rater reliability. 1. Krivov E. et al. Tumor Delineation For Brain RS by a ConvNet and Non-Uniform Patch Generation // 3rd Int. Workshop Patch - MI . 2018. 8p.