目的:各种AI软件之间存在明显的自动分割不一致。在这项研究中,我们开发了一种新颖的卷积神经网络(CNN)微调工作流程,以实现精确和鲁棒的局部分割。
方法:数据集包括湖北省肿瘤医院数据集,西妥昔单抗头颈部公共数据集,和魁北克公共数据集。七个危险器官(OAR),包括脑干,左腮腺,食道,左视神经,视神经交叉,下颌骨,和咽部收缩,被选中。首先将来自四个商业AI软件的自动分割结果与手动描绘进行比较。然后在湖北省肿瘤医院的40个样本和10个样本上训练并测试了带有注意力模块的新的多尺度轻量级残差CNN模型(命名为HN-Net),分别。为了提高网络的准确性和泛化能力,微调工作流程利用不确定性估计方法,从西妥昔单抗头颈公共数据集中自动选择具有价值的候选样本,以进行进一步培训.在湖北省肿瘤医院数据集和/或整个魁北克公共数据集上评估分割性能。
结果:通过四个AI软件观察到七个OAR的平均Dice值和Hausdorff距离值的最大差异为0.13和0.7mm。拟议的HN-Net实现了比AI软件高0.14的平均Dice值,它的表现也优于其他流行的CNN模型(HN-Net:0.79,U-Net:0.78,U-Net++:0.78,U-Net-Multi-scale:0.77,AI软件:0.65)。此外,使用本地数据集和外部公共数据集的HN-Net微调工作流进一步改进了自动分割,平均Dice值增加了0.02。
结论:商业AI软件的描述需要仔细审查,本地化的进一步培训对于临床实践是必要的。通过使用本地数据集和外部公共数据集,可以合理地采用所提出的微调工作流程来实现准确且可靠的自动分割模型。
OBJECTIVE: Obvious inconsistencies in auto-segmentations exist among various AI software. In this study, we have developed a novel convolutional neural network (CNN) fine-tuning workflow to achieve precise and robust localized segmentation.
METHODS: The datasets include Hubei Cancer Hospital dataset, Cetuximab Head and Neck Public Dataset, and Québec Public Dataset. Seven organs-at-risks (OARs), including brain stem, left parotid gland, esophagus, left optic nerve, optic chiasm, mandible, and pharyngeal constrictor, were selected. The auto-segmentation results from four commercial AI software were first compared with the manual delineations. Then a new multi-scale lightweight residual CNN model with an attention module (named as HN-Net) was trained and tested on 40 samples and 10 samples from Hubei Cancer Hospital, respectively. To enhance the network\'s accuracy and generalization ability, the fine-tuning workflow utilized an uncertainty estimation method for automatic selection of candidate samples of worthiness from Cetuximab Head and Neck Public Dataset for further training. The segmentation performances were evaluated on the Hubei Cancer Hospital dataset and/or the entire Québec Public Dataset.
RESULTS: A maximum difference of 0.13 and 0.7 mm in average Dice value and Hausdorff distance value for the seven OARs were observed by four AI software. The proposed HN-Net achieved an average Dice value of 0.14 higher than that of the AI software, and it also outperformed other popular CNN models (HN-Net: 0.79, U-Net: 0.78, U-Net++: 0.78, U-Net-Multi-scale: 0.77, AI software: 0.65). Additionally, the HN-Net fine-tuning workflow by using the local datasets and external public datasets further improved the automatic segmentation with the average Dice value by 0.02.
CONCLUSIONS: The delineations of commercial AI software need to be carefully reviewed, and localized further training is necessary for clinical practice. The proposed fine-tuning workflow could be feasibly adopted to implement an accurate and robust auto-segmentation model by using local datasets and external public datasets.