王芳,王熙,兰文娟,杨智超,孙碧婷,刘颖,谷宇,白洁,唐思源.基于知识蒸馏改进U-Net网络模型用于分割CT图像中的口腔颌面部肿瘤[J].中国医学影像技术,2024,40(4):591~597 |
基于知识蒸馏改进U-Net网络模型用于分割CT图像中的口腔颌面部肿瘤 |
Improved U-Net network model based on knowledge distillation for segmenting oral and maxillofacial tumor on CT images |
投稿时间:2023-12-20 修订日期:2024-01-28 |
DOI:10.13929/j.issn.1003-3289.2024.04.024 |
中文关键词: 口腔疾病 颌面部疾病 肿瘤 神经网络,计算机 体层摄影术,X线计算机 知识蒸馏 |
英文关键词:mouth diseases maxillofacial diseases neoplasms neural networks, computer tomography, X-ray computed knowledge distillation |
基金项目:包头市卫生健康科技项目计划(wsjkwkj021)、包头医学院科学研究基金(BYJJ-ZRQM202009)。 |
|
摘要点击次数: 747 |
全文下载次数: 535 |
中文摘要: |
目的 观察基于知识蒸馏改进U-Net网络模型用于分割CT图像中的口腔颌面部肿瘤的价值。方法 收集2个医疗中心121例口腔颌面部肿瘤患者共609幅CT图像;于公开数据集HECKTOR2020搜集254例口腔颌面部肿瘤患者共1 977幅CT图像。向U-Net网络模型中引入多尺度和注意力机制,加入残差网络,建立改进U-Net模型;采用知识蒸馏技术生成学生模型,观察模型分割CT图像中的口腔颌面部肿瘤的效能。结果 改进U-Net模型大小为89.30 MB,参数数量为17.82 M,计算量为22.13 GFlops;其分割CT所示口腔颌面部肿瘤的精确率(Precision)、召回率(Recall)、戴斯相似系数和交并比分别为0.835、0.787、0.812及0.761,优于既往结合常规损失函数(Dice Loss function)所获模型及未改进模型;且除Precision之外,学生模型与教师模型差异较小。结论 基于知识蒸馏改进U-Net网络模型用于分割CT图像中的口腔颌面部肿瘤具有较高价值。 |
英文摘要: |
Objective To observe the value of an improved U-Net network model based on knowledge distillation for segmenting oral and maxillofacial tumors on CT images. Methods Totally 609 CT images of 121 patients with oral and maxillofacial tumors from 2 medical centers were collected. Meanwhile, 1 977 CT images of 254 patients with oral and maxillofacial tumors in public dataset HECKTOR2020 were selected. The multi-scale and attention mechanisms were introduced into U-Net network model to establish an improved U-Net model combining with residual network. Knowledge distillation technology was used to generate student models. The efficacy of the improved U-Net model for segmenting oral and maxillofacial tumor on CT images was observed. Results The improved U-Net model had a size of 89.30 MB, a parameter count of 17.82 M and a computational load of 22.13 GFlops. The Precision, Recall, Dice similarity coefficient and intersection over union of the improved U-Net for segmenting oral and maxillofacial tumors on CT images was 0.835, 0.787, 0.812, and 0.761, respectively, superior to those of models established with previous methods combined with conventional Dice Loss function and unimproved model. Except for Precision, the model had relatively small difference with its teacher model. Conclusion Improved U-Net network model based on knowledge distillation was valuable for segmenting oral and maxillofacial tumors on CT images. |
查看全文 查看/发表评论 下载PDF阅读器 |
|
|
|