neuluna / evonms

Evolutionary Normalization Optimization Boosts Semantic Segmentation Network Performance

This repository contains the source code for the project to determine how layer-specific normalization methods can influence the segmentation result of an U-Net by using an evolutionary algorithm approach.

Structure of the Repository

data_generator.py

Loads the dataset and normalizes the images/masks to 0 to 1 if it is a binary task. If it is a multi-class segmentation, the masks contain the class labels from 0 to num_classes. Returns the dataset split into the tensors train, val and test for further processing.

main.py

• main(): runs the U-Net and the evaluation of the different individuals of a generation over multiple generations and sorting the models for the first and the last generation
• select(): Runs the selection and breeding of the new population

metrics.py

• dice_coefficient(): calculates the Dice Similarity Coefficient (DSC)
• dice_coef_loss(): calculates the loss based on the DSC
• get_flat(): flattens the predicted masks
• draw_bb(): draws a minimal rectangle based on the predictions
• get_bb(): calls the draw_bb() function
• bb_IoU(): calculates the IoU score of the predicted bounding boxes
• IoU(): calculates the Intersection over Union score
• hd_95_monai(): calculates the Hausdorff Distance 95

unet.py

• conv_layer(): builds one layer of the U-Net with the given settings and can be variable including normalization, activation and filter size.
• unet(): builds the U-Net architecture by using 4 Layers for encoding and Decoding by using a normal up-sampling.

Util Files

To install this project, follow these steps:

• constants.json: defines all genes which can be chosen to build one U-Net individual
• datasets.json: gives an overview of the datasets and its structure
• env.yml: yaml-file to create an environment to run the code

License

This repository is licensed under the terms of the MIT License.

Citation

Please site the usage of the software as follows:

Neubig, L., Kist, A.M. (2023). Evolutionary Normalization Optimization Boosts Semantic Segmentation Network Performance. In: Greenspan, H., et al. Medical Image Computing and Computer Assisted Intervention – MICCAI 2023. MICCAI 2023. Lecture Notes in Computer Science, vol 14223. Springer, Cham. <https://doi.org/10.1007/978-3-031-43901-8_67>