| 姜佳丽,李小强,邹童,陈兆学.基于深度学习的自动分割与多模态特征融合分型架构用于预测肾肿瘤类型[J].中国医学影像技术,2026,42(5):759~763 |
| 基于深度学习的自动分割与多模态特征融合分型架构用于预测肾肿瘤类型 |
| Deep learning-based automatic segmentation and multimodal feature fusion classification architecture for predicting type of renal tumor |
| 投稿时间：2025-05-27 修订日期：2025-11-30 |
| DOI：10.13929/j.issn.1003-3289.2026.05.022 |
| 中文关键词: 肾肿瘤 体层摄影术,X线计算机 深度学习 影像组学 |
| 英文关键词:kidney neoplasms tomography, X-ray computed deep learning radiomics |
| 基金项目:北京市中医药薪火传承"3+3"工程李景元基层老中医传承工作室建设项目(2023-JC-02)。 |
|
| 摘要点击次数: 88 |
| 全文下载次数: 51 |
| 中文摘要: |
| 目的 构建基于深度学习的自动分割与多模态特征融合分型模型,观察其预测肾肿瘤类型的价值。方法 纳入2023年肾脏和肾脏肿瘤分割挑战赛(KiTS23)数据集中419例肾肿瘤的CT图像,包括310例肾透明细胞癌(ccRCC)、45例乳头状肾细胞癌(pRCC)、38例肾嫌色细胞癌(chRCC)及26例嗜酸细胞腺瘤(ONC)。利用数据增强将样本量扩充至528例,并按7∶1∶2比例划分训练集、测试集及验证集。采用基于UNet的改良3D nnUNet算法框架自动分割CT图像,提取肿瘤影像组学特征和深度学习特征;将二者融合并引入多模态融合模块,构建分型模型3DResnet18_A。采用戴斯相似系数(DSC)评估3D nnUNet模型分割效能;以准确率(Acc)、精确率(Pre)、召回率(Rec)及F1-score比较3DResnet18_A模型与其他模型的分型效能。结果 3D nnUNet模型分割CT中的肾脏及肾肿瘤的DSC分别为0.964及0.923。3DResnet18_A模型基于3D nnUNet分割标签及原始人工标注标签预测肾肿瘤类型的Acc、Pre、Rec及F1-score均优于其他模型,其最大值分别为83.02%、84.31%、0.830及0.823。结论 基于深度学习的自动分割与多模态特征融合分型架构预测肾肿瘤类型效能较佳。 |
| 英文摘要: |
| Objective To construct a deep learning-based automatic segmentation and multimodal feature fusion classification model, and to observe its value for predicting the type of renal tumor. Methods CT images of 419 cases of renal tumor in kidney and kidney tumor segmentation challenge 2023 (KiTS23) dataset were enrolled, including 310 cases of clear cell renal cell carcinoma (ccRCC), 45 cases of papillary renal cell carcinoma (pRCC), 38 cases of chromophobe renal cell carcinoma (chRCC) and 26 cases of oncocytoma (ONC). The sample size was expanded to 528 cases using data augmentation, and then were divided into training set, testing set and validation set at the ratio of 7∶1∶2. The images were automatically segmented using improved 3D nnUNet algorithm framework based on UNet. Subsequently, radiomics features and deep learning features of renal tumor were extracted and fused, and a multimodal fusion module was introduced to construct classification model 3DResnet18_A. The segmentation performance of 3D nnUNet model was evaluated using Dice similarity coefficient (DSC), while its classification performance was evaluated using accuracy (Acc), precision (Pre), recall (Rec) and F1-score through comparison with other models. Results DSC of 3D nnUNet model for segmentation of kidney and renal tumor in CT images was 0.964 and 0.923, respectively. Based on 3D nnUNet segmentation labels and original manually annotated labels, 3DResnet18_A model demonstrated Acc, Pre, Rec and F1-score for predicting the type of renal tumor all superior to those of the other models, with the maximum value of 83.02%, 84.31%, 0.830 and 0.823, respectively. Conclusion Deep learning-based automatic segmentation and multimodal feature fusion classification architecture performed well in predicting the type of renal tumor. |
| 查看全文 查看/发表评论 下载PDF阅读器 |
|
|
|
