Covid-19-20 - basic method of network segmentation based on vnet3d

Today we will share COVID The complete implementation process of binary segmentation of , In order to facilitate everyone to learn and understand the whole process , The whole process steps are sorted out , And give the detailed results of the steps . If you are interested, please give it a try .

One 、 Image analysis and preprocessing

（1）、199 example COVID Of CT The original data, annotation data and training label files can be downloaded and obtained on the official website . The validation data are 50 example CT data , You can also download it on the official website .

（2）、 Deal with data in a simple and crude way , Because the training graphics card is 1080ti Of 11G The memory , So first, the training original image and annotation image are scaled to a fixed size , The concrete is 160x160x64, If your graphics memory is large enough , You can change the size a little larger , If the video memory is small , You can reduce the size .

（3）、 Set the window width and level of the scaled original image , take -1000 To 0 Between Hu Value is reserved , Other ranges are truncated .

（4）、 Then the average value is 0, The variance of 1 Normalize the original image .

（5）、 In order to prevent the model from over fitting , The processing process of data enhancement is also added , Use rotation , translation , Flipping and other methods are used to expand the image and annotation data at the same time , Expanded here 20 times .

Two 、 Divide the network

（1）、 build VNet3d Model , The network input size is （160,160,64）. Specific implementation can refer to ​ ​Tensorflow Introductory tutorial （ nineteen ）—— be based on VNet The case of prostate segmentation ​​.

（2）、loss It's binary dice function . Specific implementation can refer to ​ ​Tensorflow Introductory tutorial （ Thirty-four ）—— Two kinds of image segmentation loss functions are commonly used ​​.

（3）、 The segmentation loss results and accuracy are shown in the following figure .

3、 ... and 、 Verify the data results

stay 50 Test on the data , The results of validation data rank in 81 name , The result doesn't look so bad .

Detailed evaluation index values can be seen on the submission results , With dice For example, indicators , Good result dice The value can reach 0.8 above , Poor results dice Only value 0.

Four 、 Direction of improvement

1、 Considering the limited computing resources , We can segment and extract the lung tissue region first , Then the lung tissue is segmented and extracted , Improve segmentation accuracy

2、 consider Vnet2d and Vnet3d Divide the network for combination segmentation , Improve segmentation accuracy .

3、 Data preprocessing part , Adopt appropriate strategies , For example, the setting of window width and window level , Image size setting, etc .

---

​