search

IAmSuyogJadhav / 3d-mri-brain-tumor-segmentation-using-autoencoder-regularization

Keras implementation of the paper "3D MRI brain tumor segmentation using autoencoder regularization" by Myronenko A. (https://arxiv.org/abs/1810.11654).
MIT License
369 stars 113 forks source link

Two things to mention #29

Closed Crispy13 closed 4 years ago

Crispy13 commented 4 years ago

Hi. I'm using this source code to build a model for my project.

I found several things which are different from the paper or need to be fixed.

  1. VAE branch is not trained.

    • When creating model using keras.models.Model, the output argument has not the VAE branch output.

  2. Dropout was used instead of spatial dropout.

    We also use the spatial dropout with a rate of 0.2 after the initial encoder convolution.

  3. When evaluating the model, dice score varys batch by batch. Current dice score is not the mean of the scores(e.g. (sample 1 + sample 2) / 2, ) but the function calculates the score like the samples of a batch are one sample.

IAmSuyogJadhav commented 4 years ago

Hello @Crispy13, I am currently involved with a lot of other work and do not have spare time to focus on past projects. I see you have created your own fork implementing fixes for some of the issues you faced. If you manage to fix any more and test them out for correctness, feel free to submit a PR here for other users to benefit from.