Med

Detecting Brain Lesions in Multiple Sclerosis Patients with Deep Learning

from scipy.signal import convolve2d# vertical line detector
kernel = np.array([
[0,0,0,1,0,0,0],
[0,0,0,1,0,0,0],
[0,0,0,1,0,0,0],
[0,0,0,1,0,0,0],
[0,0,0,1,0,0,0],
[0,0,0,1,0,0,0],
[0,0,0,1,0,0,0]
])
image_v = convolve2d(image,kernel,mode='same')

The resulting image is the one in the middle. Note how vertical lines are more pronounced than in the original. Horizontal lines can be amplified with another kernel, generating the image on the right.

# horizontal line detector
kernel = np.array([
[0,0,0,0,0,0,0],
[0,0,0,0,0,0,0],
[0,0,0,0,0,0,0],
[1,1,1,1,1,1,1],
[0,0,0,0,0,0,0],
[0,0,0,0,0,0,0],
[0,0,0,0,0,0,0]
])
image_h = convolve2d(image,kernel,mode='same')

Usually, many filters are used within a single layer in a CNN which are learned during the training (inference) phase, i.e. weights are set in a way that regions of interest are more pronounced than irrelevant ones. In some cases, a pooling layer is applied after a convolutional layer. In a CNN, sequential mappings of convolutional layers and pooling layers are calculated, which amplify certain patterns over the course of the network as the output of each operation is a again feature map. Feature maps are getting smaller in pixels, which is how more complex structures can be learned by the filters.

Note that for volumetric images, three-dimensional filters are used. As mentioned above, a volume consists of several millions of voxels. This makes it currently unfeasible to process the whole volume, and subsamples have to be used instead (e.g. 80x80x80 voxels). This provides less context in border regions which will results in worse performance, but it need less computational power. There are different architectures depending on the task (e.g. if a classification or segmentation is performed) and image tensor rank; U-Net, 3D-U-Net, V-Net are among the better known ones for segmentation tasks, whereas ResNet proved to be high-performing in classification tasks. Current top accuracies in lesion segmentation are in the range of 91.7–93.4%.


Source link
Tags
Back to top button
close
Thanks !

Thanks for sharing this, you are awesome !

Pin It on Pinterest

Share This

Share this post with your friends!

Close