BACKGROUND AND PURPOSE Manual delineation of clinical target volumes (CTVs) and organs at risk (OARs) is time-consuming, and automatic contouring tools lack clinical validation. We aimed to construct and validate the use of convolutional neural networks (CNNs) to set better contouring standards for rectal cancer radiotherapy. MATERIALS AND METHODS We retrospectively collected and evaluated computed tomography (CT) scans of 199 rectal cancer patients treated at our hospital from February 2018 to April 2019. Two CNNs-DeepLabv3+ for extracting high-level semantic information and ResUNet for extracting low-level visual features-were used for the CTV and small intestine contouring, and bladder and femoral head contouring, respectively. Contouring quality was compared using the paired t test. Five-point objective grading was performed independently by two experienced radiation oncologists and verified by a third. The CNN manual correction time was recorded. RESULTS CTVs calculated using DeepLabv3+ (CTVDeepLabv3+) had significant quantitative parameter advantages over CTVResUNet (volumetric Dice coefficient, 0.88 vs 0.87, P = 0.0005; surface Dice coefficient, 0.79 vs 0.78, P = 0.008). Among 315 graded cases, DeepLabv3+ obtained the highest scores with 284 cases, consistent with the objective criteria, whereas CTVResUNet had the minimum mean manual correction time (7.29 min). DeepLabv3+ performed better than ResUNet for small intestine contouring and ResUNet performed better for bladder and femoral head contouring. The manual correction time for OARs was <4 min for both models. CONCLUSION CNNs at various feature resolution levels well delineate rectal cancer CTVs and OARs, displaying high quality and requiring shorter computation and manual correction time.

中文翻译：

---

直肠癌术后放疗深度学习自动勾画临床靶区及危险器官

背景和目的 手动勾画临床靶区 (CTV) 和危险器官 (OAR) 非常耗时，而且自动勾画工具缺乏临床验证。我们旨在构建和验证卷积神经网络 (CNN) 的使用，以便为直肠癌放射治疗设置更好的轮廓标准。材料与方法 我们回顾性收集并评估了 2018 年 2 月至 2019 年 4 月在我院治疗的 199 名直肠癌患者的计算机断层扫描 (CT) 扫描结果。两个用于提取高级语义信息的 CNNs-DeepLabv3+ 和用于提取低级视觉特征的 ResUNet -分别用于 CTV 和小肠轮廓以及膀胱和股骨头轮廓。使用配对 t 检验比较轮廓质量。五点客观分级由两名经验丰富的放射肿瘤学家独立进行，并由第三名进行验证。记录CNN人工校正时间。结果 使用 DeepLabv3+ (CTVDeepLabv3+) 计算的 CTV 与 CTVResUNet 相比具有显着的定量参数优势（体积骰子系数，0.88 对 0.87，P = 0.0005；表面骰子系数，0.79 对 0.78，P = 0.008）。在 315 个分级案例中，DeepLabv3+ 获得最高分 284 个，符合客观标准，而 CTVResUNet 的平均手动校正时间最短（7.29 分钟）。DeepLabv3+ 在小肠轮廓绘制方面的表现优于 ResUNet，而在膀胱和股骨头轮廓方面，ResUNet 表现更好。两种模型的 OAR 手动校正时间均小于 4 分钟。