Predictive value of automatic organs-at-risk segmentation geometric metrics for dose variation in nasopharyngeal carcinoma radiotherapy planning
Objective To explore the correlation between geometric segmentation metrics and dosimetric indicators of or-gans at risk(OAR),and to study the impact of geometric indicator changes on dose indicators through simulation experi-ments,as well as the feasibility of predicting dose variations.Methods A retrospective analysis of 20 nasopharyngeal carcinoma patients were selected from January 01,2023 to June 30,2023 in Jinshazhou Hospital of Guangzhou University of Chinese Medicine.There were 13 males and 7 females.Age ranged from 38 to 74 years,with a median age of 53.8 years.Clinical experts delineated OARs as reference contours and optimized into reference plans,while deep learning software delineated OARs as alternative contours.Using the reference plan as a template and maintaining consistent planning pa-rameters,each corresponding alternative contour was re-optimized to form an alternative plan.The Dice similarity coeffi-cient(DSC)and organ delta dose(△D)of the contours of the reference and alternative two-group plans were statistically calculat-ed,and the Pearson correlation coefficient r between DSC and △D was calculated.Additionally,simulation experiments involved simple random sampling of a nasopharyngeal carcinoma CT image to simulate the delineation of a U-shaped target and four OARs:the spinal cord,parotid glands,and oral cavity.The study focused on assessing the dosimetric impact of different distances,the shape and volume of the spinal cord contour relative to the target contour,and changes in its posi-tion relative to reference contour.Results In clinical case,the DSC for the brainstem reached 0.85,while the DSC for the parotid glands and oral cavity was 0.78-0.80.The DSC for the visual pathway(optic nerve and optic chiasm)had an average of 0.55 with a significant standard deviation.The mean and standard range between the reference and alternative plan for the brainstem maximum dose and the left and right parotid gland mean dose were(-1.26±1.65),(0.99±1.60)and(0.72±2.54)Gy respectively,with a statistically significant difference(t values were-4.700,3.818 and 1.831,P values were＜0.001,＜0.001 and 0.041).The correlation coefficients between brainstem DSC and PGTV homogeneity index,and between the right parotid gland DSC and PTV homogeneity index were|r|＞0.4,indicating a moderate correlation,and the correlation coefficients between other metrics and △D were all small.The simulation experi-ments found that the shape,volume,and position of the OAR contours were all characteristics affecting dose variations.H owever,a reduction in the distance between the OAR contour and the target area did not necessarily lead to a greater dose effect.Conclusions Geometric metrics lack dose prediction features.When evaluating dose effects of geometric met-rics,the shape,size,and location of contours must be carefully considered.
nasopharyngeal carcinomaorgans at riskautomatic segmentationsegmentation metricdosimetric metric
NETL
NSTL
万方数据
