A CNN to segment X-Ray images
XNet is a Convolutional Neural Network designed for the segmentation of X-Ray images into bone, soft tissue and open beam regions. Specifically, it performs well on small datasets with the aim to minimise the number of false positives in the soft tissue class.
Built on a typical encoder-decoder architecture as inspired by SegNet.
Additional feature extraction stage, with weight sharing across some layers.
Fine and coarse grained feature preservation through concatenation of layers.
L2 regularisation at each of the convolutional layers, to decrease overfitting.
The architecture is described in the
XNet outputs a mask of equal size to the input images.
XNet is trained on a small dataset which has undergone
augmention. Examples of this augmentation step can be found in the
augmentations.ipynb notebook in the
Augmentations folder. Similarly the
Training folder contains python scripts that perform the necessary augementations.
train.py from the
Training folder calls various other scripts to perform one of two possible ways of augmenting the images:
'On the fly augmentation' where a new set of augmentations is generated at each epoch.
XNet was benchmarked against two of the leading segmentation networks:
We trained on a dataset of:
150 X-Ray images.
No scatter correction.
.tifimage downsampled to 200x200
20 human body part classes.
As this work grew out of work with a corporation we are sadly unable to share the propriatory data we used.
For more information and context see the conference poster
Please note that some of the path variables may need to be corrected in order to utilise the current filing system. These are planned to be updated in the future.
Download the file for your platform. If you're not sure which to choose, learn more about installing packages.
Hashes for durham_XNet-0.0.4-py3-none-any.whl