Stress-Testing Pelvic Autosegmentation Algorithms Using Anatomical Edge Cases

<p>In this single institution retrospective study, we reviewed 950 consecutive patients with prostate adenocarcinoma receiving definitive radiotherapy between 2011 and 2019, and identified among them 112 patients with anatomic variations (edge cases) seen on simulation CT and/or MRI imaging. These variations included hip arthroplasty, prostate median lobe hypertrophy, so-called “droopy” seminal vesicles, presence of a urinary catheter, and others. A separate cohort of 19 “normal” cases were randomly selected for inclusion. Prostate, rectum, bladder, and bilateral femoral heads were manually segmented on all CT simulation images (where present) and were ultimately used clinically for radiation treatment planning.</p><p>We leveraged this imaging dataset to assess the comparative performance of deep learning, atlas-based, and model-based autosegmentation methods across both normal and edge case cohorts: <a href="https://doi.org/10.1016/j.phro.2023.100413">https://doi.org/10.1016/j.phro.2023.100413</a>.  In this paper and in the figure on the right, we show the Cross-sectional CT-based anatomy and autosegmentation performance for representative edge cases.</p><p>A) Hypertrophic prostate edge case. Each panel depicts a focused excerpt from a single CT scan, centered about two different structures (prostate, bladder) in three different planes (axial, sagittal, coronal). Clinician-delineated “ground truth” contours (MD) for each structure are shown in red, while atlas-based (AB), model-based (MB), and deep-learning based (DL) autosegmented contours are depicted in green, orange, and blue, respectively. Numerical values represent DSC for the corresponding autosegmented volumes compared to MD volumes.</p><p>B) So-called “droopy” seminal vesicles edge case. Each panel depicts a focused excerpt from a single CT scan, centered about the prostate in two different planes (axial, sagittal). All colors and labeling are as in Panel A).</p>

在本院回顾性研究中，我们对2011年至2019年间接受确定性放射治疗的950例连续前列腺腺癌患者进行了审查，并在其中识别出112例在模拟CT和/或MRI成像中发现的解剖学变异（边缘案例）的患者。这些变异包括髋关节置换术、前列腺中叶肥大、所谓“下垂”的精囊、存在导尿管等情况。另外随机选取了19例“正常”病例进行纳入。在所有CT模拟图像（如有）上手动分割了前列腺、直肠、膀胱和双侧股骨头，并最终用于临床放射治疗计划。本研究利用这一影像数据集，评估了深度学习、基于图谱和基于模型的自分割方法在正常和边缘案例群体中的比较性能：[链接](https://doi.org/10.1016/j.phro.2023.100413)。在本论文及右侧图示中，我们展示了基于横断面CT的代表性边缘案例的解剖结构和自分割性能。

A) 前列腺肥大边缘案例。每个面板展示了单个CT扫描的聚焦摘录，以三个不同的平面（轴位、矢状位、冠位）中两种不同的结构（前列腺、膀胱）为中心。每个结构由临床医师绘制的“地面真实”轮廓（MD）以红色表示，而基于图谱（AB）、基于模型（MB）和基于深度学习（DL）的自分割轮廓分别以绿色、橙色和蓝色表示。数值代表与MD体积对应的自分割体积的DSC值。

B) 所称“下垂”的精囊边缘案例。每个面板展示了单个CT扫描的聚焦摘录，以两个不同的平面（轴位、矢状位）中的前列腺为中心。所有颜色和标签与面板A相同。