Auto-segmentation variability of organs at risk in patients with nasopharyngeal carcinoma and its dosimetric impacts
10.3760/cma.j.cn112271-20231220-00216
- VernacularTitle:基于鼻咽癌危及器官自动分割可变性及剂量学影响
- Author:
Liyuan ZHANG
1
;
Jinyan HU
;
Shiyong GU
;
Xiaping WEI
Author Information
1. 广州中医药大学金沙洲医院肿瘤放射治疗中心，广州 510168
- Keywords:
Organ at risk;
Auto-segmentation;
Dose difference;
Overlap volume histogram
- From:
Chinese Journal of Radiological Medicine and Protection
2024;44(11):944-952
- CountryChina
- Language:Chinese
-
Abstract:
Objective:To explore the adjustment ranges of auto-segmentation contours for organs at risk (OAR) in patients with nasopharyngeal carcinoma and assess the dosimetric impacts of the contours from varying sources on radiotherapy plans.Methods:Twenty-five patients with early-stage nasopharyngeal carcinoma were investigated. Through expert delineation, deep learning-based automatic delineation, and atlas-based automatic delineation of their spinal cord, brainstem, optic nerves, optic chiasm, parotid glands, oral cavity, hypopharynx, and mandible, as well as expert correction of these automatic delineations, five structure sets were formed. Moreover, the contours delineated by experts (also referred to as the expert contours) of the target volumes and other OARs were copied into the images for subsequent research. The Dice similarity coefficients (DSCs) of the structure sets were calculated. Using the radiotherapy plans optimized based on expert contours as templates, the radiotherapy plans and dose distributions of all the structure sets were established. The expert contours and contours determined using automatic delineation and corrected by experts (also referred to as the corrected contours) were defined as clinical contours. Then, three research objectives were set: the dosimetric effects of inter-observer clinical contour variations, the impacts of contour variations on plan optimization, and the impacts of contour variations on plan evaluation.Results:The average DSC of the visual pathway was 0.62±0.10, lower than that of other OARs (0.86±0.04). After expert correction, the DSCs of contours obtained using deep learning- and atlas-based automatic delineation increased by 7.61% and 10.69%, respectively. For the dosimetric effects of inner-observer contour variations, the Dmax of the optic chiasm was the maximum (3.96±6.02)Gy, while the Dmean of the hypopharynx was the minimum (0.81±0.55 Gy). When the impacts of contour variations on plan optimization were assessed based on expert contours, the dose differences (Δ D) exceeding ±3 Gy accounted for 22%, 14%, 46%, and 42%, respectively for the spinal cord, brainstem, optic nerve, and optic chiasm and accounted for only 2% for other OARs. After expert correction, the Δ D between automatic and expert contours decreased, with Δ D exceeding ±3 Gy decreased by 16% and 14%, respectively for the optic nerves and optic chiasm. When the average distance of the overlap volume histogram (OVH) exceeded 3.5 cm, all Δ Dmax fell within ±3 Gy. When the average distance of OVH was greater than 1.5 cm, all Δ Dmean fell within ±2 Gy. For contours obtained using deep learning and atlas-based automatic delineation, the doses of 50.0%±17.3% and 52.6%±19.3% of patients fell within the dose ranges of clinical contours, respectively. The numbers of patients for whom the Dmax of the spinal cord, optic nerve, optic chiasm and the D1 cm 3 of the mandible in the two types of automatic contours fell within the dose ranges of clinical contours were statistically different ( t = -4.24, -3.99, -3.16, 3.51, P < 0.05). Conclusions:After expert correction, the automatic delineation results from different sources exhibited certain geometric differences. The expert correction reduced the impacts of automatic contours on plan optimization. The average distance of OVH is identified as an important feature used to determine dose differences. For small-volume serial organs close to the target volumes, meticulous corrections are required before applying auto-segmentation to clinical practice.
