This is the official Tensorflow implementation of ARACHNET: INTERPRETABLE SUB-ARACHNOID SPACE SEGMENTATION USING AN ADDITIVE CONVOLUTIONAL NEURAL NETWORK.
Considering the growing demand for interpretable deep learning models in medical imaging, this paper introduces ArachNet, a novel Convolutional Neural Network (CNN) architecture tailored for spinal sub-arachnoid space segmentation in MR images. Unlike recent relevant approaches, ArachNet is an inherently interpretable model based on the EPU-CNN interpretable framework. The model necessitates an initial image decomposition into perceptually meaningful components representing structural and anatomical features of the spinal region. Each perceptual component is propagated to a dedicated sub-network that outputs a perceptually meaningful 2D representation, enabling the interpretation of the overall inferred segmentation. The experimental evaluation indicates that ArachNet has great application prospects in the navigation of minimally invasive surgical tools, such as vine robot endoscopes, in the spine.
ArachNet, requires an initial decomposition of the input image into perceptual image components, referred to as Perceptual Feature Maps (PFMs). For the interpretations to be meaningful, each of these components needs to render a perceptual quantity, e.g., texture or a range of values corresponding to different tissues. ArachNet, utilizes a set of CNN sub-networks, each of which is dedicated to processing a different perceptual component. Then, the outputs provided by each CNN sub-network are aggregated through summation, and subsequently propagated to an activation function for the construction of the final segmentation. Since the final segmentation is the result of a known aggregation function, the user can easily derive the contribution of each PFM to that segmentation.
To use ArachNet, you need to provide a base CNN architecture and a perceptual feature extractor. In this repository we provide a Unet base CNN architectures and perceptual feature extractors used in our experiments. An example of usage is shown below:
from models.arachnet import ArachNet
from models.subnetworks import UnetSubnet
# Initialize the ArachNet model
arachnet = ArachNet(init_size=64, subnet_act="tanh", arach_act="sigmoid", features_num=2,
subnet=UnetSubnet)The subnet defines the base CNN architecture. In our implementation init_size defines the number of
features in the first convolutional layers that is incremented by a factor of 2 in each subsequent convolutional
block of the base CNN architecture. num_pfms defines the number of perceptual feature maps used as input for a particular
application.
The subnet_act defines the output activation function of the base CNN architecture and the arach_act defines the inverse
of the link function g used to provide the final output of the EPU-CNN model.
Detailed info to be added soon.
Detailed info to be added soon.
Detailed info to be added soon.
- Add references to all the paper datasets
- Complete README.md
- Add requirements.txt
- Implement interpretation visualizations in a nice format
- Add evaluation code
- Add inference code
- Add Wavelet PFM extraction
- Add Multiclass Training and Evaluation code
- Add the ISBI2024 Reference
This work has been funded by the European Union, under grant agreement No 101099145, project SoftReach.
If you find this work useful, please cite our paper:
@inproceedings{dimas2024arachnet,
title = {ArachNet: Interpretable Sub-Arachnoid Space Segmentation Using an Additive Convolutional Neural Network},
author = {Dimas, George and Kalozoumis, Panagiotis G and Vartholomeos, Panagiotis and Iakovidis, Dimitris K},
booktitle = {2024 IEEE International Symposium on Biomedical Imaging (ISBI)},
pages = {1--5},
year = {2024},
organization = {IEEE}
}
- George Dimas
- Panagiotis Kalozoumis
- Panagiotis Vartholomeos
- Dimitris Iakovidis