基于改进的YOLOv3-DN对脑胶质瘤检测分级

Detection and classification of glioma based on improved YOLOv3-DN

探讨一种基于改进的YOLOv3-DN计算机辅助诊断系统对脑部胶质瘤核磁共振成像术前检测和分级的价值。根据收集的低级别胶质瘤30例和多形性胶质母细胞瘤30例,对其进行数据增强,通过N4ITK偏场校正及归一化预处理后,将2400幅图像用于改进的YOLOv3-DN网络进行训练,后在测试数据集上检测,通过5折交叉验证及ROC评估性能。实验结果表明,YOLOv3-DN优于原始的YOLOv3和YOLO,能够以93.71%的整体准确度检测肿瘤位置,以94.49%的总准确率区分肿瘤级别,对胶质瘤术前分级具有一定的预测价值。

To explore the value of a computer-aided diagnosis system based on improved YOLOv3-DN in the preoperative detection and grading of brain glioma MRI,30 cases of low-grade gliomas and 30 cases of glioblastoma multiforme were collected and the data were enhanced.After N4ITK bias correction and normalization preprocessing,2400 images were used to train the improved YOLOv3-DN network,and then they were detected on the test data set.The performance was evaluated by 5 fold cross validation and ROC.The results show that YOLOv3-DN is superior to the original YOLOv3-DN and YOLOv3,which can detect the tumor location with 93.71%overall accuracy,and distinguish the tumor level with 94.49%overall accuracy.It has a certain predictive value for the preoperative grade of glioma.