School of Medicine, Queen's University, 88 Stuart Street, Kingston, ON, K7L 3N6, Canada.
Department of Pathology and Molecular Medicine, Queen's University, Kingston, ON, Canada.
Int J Comput Assist Radiol Surg. 2022 Sep;17(9):1663-1672. doi: 10.1007/s11548-022-02658-4. Epub 2022 May 19.
Ultrasound-based navigation is a promising method in breast-conserving surgery, but tumor contouring often requires a radiologist at the time of surgery. Our goal is to develop a real-time automatic neural network-based tumor contouring process for intraoperative guidance. Segmentation accuracy is evaluated by both pixel-based metrics and expert visual rating.
This retrospective study includes 7318 intraoperative ultrasound images acquired from 33 breast cancer patients, randomly split between 80:20 for training and testing. We implement a u-net architecture to label each pixel on ultrasound images as either tumor or healthy breast tissue. Quantitative metrics are calculated to evaluate the model's accuracy. Contour quality and usability are also assessed by fellowship-trained breast radiologists and surgical oncologists. Additionally, the viability of using our u-net model in an existing surgical navigation system is evaluated by measuring the segmentation frame rate.
The mean dice similarity coefficient of our u-net model is 0.78, with an area under the receiver-operating characteristics curve of 0.94, sensitivity of 0.95, and specificity of 0.67. Expert visual ratings are positive, with 93% of responses rating tumor contour quality at or above 7/10, and 75% of responses rating contour quality at or above 8/10. Real-time tumor segmentation achieved a frame rate of 16 frames-per-second, sufficient for clinical use.
Neural networks trained with intraoperative ultrasound images provide consistent tumor segmentations that are well received by clinicians. These findings suggest that neural networks are a promising adjunct to alleviate radiologist workload as well as improving efficiency in breast-conserving surgery navigation systems.
超声引导是保乳手术中很有前景的方法，但肿瘤轮廓勾画通常需要在手术时由放射科医生来完成。我们的目标是开发一种实时自动基于神经网络的肿瘤轮廓勾画方法，用于术中指导。通过像素级指标和专家视觉评分来评估分割准确性。
本回顾性研究纳入了 33 例乳腺癌患者的 7318 例术中超声图像，随机分为 80:20 用于训练和测试。我们采用 u 型网络架构对超声图像上的每个像素进行标记，将其分为肿瘤或健康乳腺组织。计算定量指标以评估模型的准确性。 fellowship培训的乳腺放射科医生和外科肿瘤医生还评估了轮廓质量和可用性。此外，通过测量分割帧率来评估将我们的 u 型网络模型应用于现有手术导航系统的可行性。
我们的 u 型网络模型的平均骰子相似系数为 0.78，受试者工作特征曲线下面积为 0.94，灵敏度为 0.95，特异性为 0.67。专家的视觉评分是积极的，93%的受访者认为肿瘤轮廓质量在 7/10 或以上，75%的受访者认为轮廓质量在 8/10 或以上。实时肿瘤分割实现了 16 帧/秒的帧率，足以满足临床应用。
用术中超声图像训练的神经网络提供了一致的肿瘤分割，得到了临床医生的认可。这些发现表明，神经网络是一种很有前途的辅助手段，可以减轻放射科医生的工作量，并提高保乳手术导航系统的效率。
