Valdo2021 - vascular space segmentation in vascular disease detection challenge (I)

Today we will share Segmentation of cerebral vascular space Of The first step is the complete implementation process of thermodynamic diagram regression detection , 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 、 data Analysis and pretreatment ​

First, the effective data of intracranial vascular space area in training is extracted , There are some data without vascular space areas , Not as training data , Only tag values are analyzed here 1, Other labels are 0. In total 40 Example data , There are data of vascular space 22 example .

Analyze this 22 Basic information of example data ： Average image size [243.27777778, 298.5, 168.16666667], Images Spacing Average size [0.63042518,0.63042518,0.85555538], Average size of vascular space [5.05882353,5.62745098,3.70588235].

Then according to the boundingbox Center and size of , Generate a Gaussian thermal map at this point , The center of the Gauss thermogram is the coordinate of the center point of the vascular space area , gaussian Sigma The value is boundingbox The maximum of , If there are multiple vascular spaces on an image , Add the Gauss thermogram of all vascular spaces . The effect is shown below , The left picture is the original picture of vascular space and mask chart , The picture on the right is Gauss thermodynamic diagram .

2、 Get ready 3d Training data

In order to input the whole image into the network , The original image and thermal map need to be scaled , Because the graphics card is 1080TI Of 11G The size of the video memory , So the three modal images are scaled to a fixed size （128,128,96）, On the image （5,95） The mean of 0, The variance of 1 Normalization of .

3、 Build a return network structure

The model uses VNet3d Main structure , The activation function of the last layer is set to linear and non activation function , The loss function uses L2-loss, The learning rate is 0.001,droupout yes 0.5, The number of iterations is 600, The training data totaled 42 example , The number of iterations is 600epochs,batchsize yes 1, The optimizer is RMSPropOptimizer.

4、 Network training results

3d The regression network loss function curve is shown below .

5、 Network prediction results

stay 1 Test on the data ,3d The results of the thermodynamic diagram are as follows , You can see that there are many candidate areas , Next, we need to classify these candidate regions , Further reduce false positive areas .