How do I create a custom Dice pixel classification layer in MATLAB R2022a using the Computer Vision Toolbox?

I am using the information on the "3-D Brain Tumor Segmentation Using Deep Learning" tutorial page to carry out training on three-dimensional scans in images:

https://www.mathworks.com/help/releases/R2022a/vision/ug/segment-3d-brain-tumor-using-deep-learning.html

However, I am having difficulty in implementing a custom Dice layer in MATLAB R2022a using this tutorial page above as guidance.

Are there any other resources that would help me to implement a custom Dice layer?

Answer 1

[MathWorks Support Team]

For a detailed MATLAB implementation on building a custom Dice layer, please refer to the "Define Custom Pixel Classification Layer with Tversky Loss" documentation page:

https://www.mathworks.com/help/releases/R2022a/deeplearning/ug/define-custom-pixel-classification-layer-with-tversky-loss.html

The Tversky Loss differs from the Dice Loss by having different weights for false positives (FPs) and false negatives (FNs) in the loss function, unlike the Dice Loss which weights both FPs and FNs equally. Given the default settings in the example code in the Tversky Loss documentation page linked above, alpha=0.5 and beta=0.5, this will simplify to the Dice score. For further reference on the Tversky Loss, please consult Section 2.2 of the following publication:

https://arxiv.org/pdf/1706.05721