Department of Medical Biophysics, University of Western Ontario, London, Ontario, Canada.
Int J Radiat Oncol Biol Phys. 2013 Jan 1;85(1):95-100. doi: 10.1016/j.ijrobp.2011.07.046. Epub 2012 May 8.
To perform a rigorous technological assessment and statistical validation of a software technology for anatomic delineations of the prostate on MRI datasets.
A 3-phase validation strategy was used. Phase I consisted of anatomic atlas building using 100 prostate cancer MRI data sets to provide training data sets for the segmentation algorithms. In phase II, 2 experts contoured 15 new MRI prostate cancer cases using 3 approaches (manual, N points, and region of interest). In phase III, 5 new physicians with variable MRI prostate contouring experience segmented the same 15 phase II datasets using 3 approaches: manual, N points with no editing, and full autosegmentation with user editing allowed. Statistical analyses for time and accuracy (using Dice similarity coefficient) endpoints used traditional descriptive statistics, analysis of variance, analysis of covariance, and pooled Student t test.
In phase I, average (SD) total and per slice contouring time for the 2 physicians was 228 (75), 17 (3.5), 209 (65), and 15 seconds (3.9), respectively. In phase II, statistically significant differences in physician contouring time were observed based on physician, type of contouring, and case sequence. The N points strategy resulted in superior segmentation accuracy when initial autosegmented contours were compared with final contours. In phase III, statistically significant differences in contouring time were observed based on physician, type of contouring, and case sequence again. The average relative timesaving for N points and autosegmentation were 49% and 27%, respectively, compared with manual contouring. The N points and autosegmentation strategies resulted in average Dice values of 0.89 and 0.88, respectively. Pre- and postedited autosegmented contours demonstrated a higher average Dice similarity coefficient of 0.94.
The software provided robust contours with minimal editing required. Observed time savings were seen for all physicians irrespective of experience level and baseline manual contouring speed.
对一种用于 MRI 数据集前列腺解剖勾画的软件技术进行严格的技术评估和统计学验证。
采用了 3 阶段验证策略。第 I 阶段包括使用 100 个前列腺癌 MRI 数据集构建解剖图谱，为分割算法提供训练数据集。在第 II 阶段，2 名专家使用 3 种方法（手动、N 点和感兴趣区域）对 15 例新的前列腺癌 MRI 病例进行了轮廓勾画。在第 III 阶段，5 名具有不同 MRI 前列腺轮廓勾画经验的新医生使用 3 种方法（手动、无编辑的 N 点和允许用户编辑的完全自动分割）对相同的 15 例第 II 阶段数据集进行了分割。使用传统描述性统计、方差分析、协方差分析和合并学生 t 检验对时间和准确性（使用 Dice 相似系数）终点进行了统计分析。
在第 I 阶段，2 名医生的总轮廓和每片轮廓时间的平均值（标准差）分别为 228（75）秒、17（3.5）秒、209（65）秒和 15 秒（3.9）秒。在第 II 阶段，根据医生、轮廓类型和病例顺序观察到医生轮廓时间存在统计学显著差异。与最终轮廓相比，N 点策略的初始自动分割轮廓具有更高的分割准确性。在第 III 阶段，根据医生、轮廓类型和病例顺序再次观察到轮廓时间的统计学显著差异。与手动轮廓相比，N 点和自动分割的平均相对节省时间分别为 49%和 27%。N 点和自动分割策略的平均 Dice 值分别为 0.89 和 0.88。预编辑和后编辑的自动分割轮廓的平均 Dice 相似系数更高，为 0.94。
该软件提供了具有最小编辑需求的稳健轮廓。所有医生（无论经验水平和基线手动轮廓速度如何）都观察到了时间节省。
