UNet segmentation model

UNet segmentation model

General information

This item contains the necessary components to segment the muscle layers of rat uterine uCT images using a UNet model.

Contents

In this segmentation archive:

• The unet-container.def is the Dockerfile to create the environment in which the model was trained. The location of the files needs to be changed before creating the container.
• The requirements.txt is a file containing the Python packages to install.
• The unet-container.sif is the the Singularity version of the Docker image.
• The unet-model_v6.h5 is the model weights after training.

Model information

The model uses the code from GitHub user Karol Żak. The code is available under the MIT license and can be found here. The model has an input size of (512, 512, 1) with 32 filters, and 4 layers. It was trained for 100 epochs in batches of 250 with a dropout of 0.3. The model used the Adam optimizer and a binary cross-entropy loss function. The metric used was the intersection over union (IOU). The trained model obtained a validation IOU of 0.88.