- Author: Jesus Ramseths Echeverria
This project is for educational purposes, if you want to take it to production environments 300, 000 official MRI images are required, in addition to evaluate the metrics of each model implemented in this project.
The objective is to understand and extract information from images at the pixel level. It is used for object localization and recognition, some applications are medical imaging and autonomous cars. A neural network is trained to produce a pixel mask.
State-of-the-art techniques are based on a deep learning approach that makes use of common architectures such as CNN, TCN and deep encoder-decoders. The ResUNet architecture was used in this project.
As the convolutional neural networks become deeper, gradient fading occurs and affects the performance of the network.
The residual neural network includes the "skip by connection" feature that allows training of 152 layers without problems such as gradient fading.
ResNet works by adding identity assignments in the main part of the convolutional neural network.
