结合MRI多模态信息和3D-CNNs特征提取的脑肿瘤分割研究

Brain Tumor Segmentation Based on MRI Multimodal Information and 3D-CNNs Feature Extraction

目的探究结合MRI多模态信息和3D-CNNs特征提取对于脑肿瘤分割的价值。方法分析相比于未加入多模态3D-CNNs特征的方法,并对比2D-CNNs特征方法和3D-CNNs特征方法分割的结果,主要参考dice系数,假阳性率和sensitibity。结果在加入多模态3D-CNNs特征之后,患者的dice系数均有不同程度的提高,sensitibity系数也有改变,假阳性率显著得到改善;加上多模态3D-CNNs特征提取后,dice系数变为(88.26±4.65)%,显著优于多模态2D-CNNs特征提取的(83.67±4.22)%。而多模态2D-CNNs特征提取的运用甚至比单独使用灰度邻域结合haar小波低频系数的分割结果。结论基于多模态3D-CNNs特征提取的MRI脑肿瘤分割准确度高,适应不同患者不同模态之间的多变性和差异性,值得参考。

Objective Explore the value of combining MRI multimodal information and 3D-CNNs feature extraction for brain tumor segmentation.Methods Compared with the method of adding multimodal 3D-CNNs,and comparing the results of 2D-CNNs feature method and 3D-CNNs feature method,the main reference is the dice coefficient,false positive rate and sensitibity.Results After adding the multimodal 3D-CNNs features,the patient’s dice coefficient increased to varying degrees,the sensitibity coefficient also changed,and the false positive rate was significantly improved.After the multimodal 3D-CNNs feature extraction,the dice coefficient became (88.26±4.65)%,significantly better than the multimodal 2D-CNNs feature extraction (83.67±4.22)%.The application of multi-modal 2D-CNNs feature extraction is even more than the segmentation result of the haar wavelet low-frequency coefficient combined with the gray neighborhood.Conclusion MRI brain tumor segmentation based on multimodal 3D-CNNs feature extraction has high accuracy and adapts to the variability and difference between different modes of different patients.It is worthy of reference.