Avoiding overfitting using unetLayers

Hello!

I'm trying to get a unetLayers-network to work on a binary segmentation problem. I'm training the network on patches of CT-images where the goal is to segment the bone pixels from background pixels. Since bones in CT-images have high numerical values compared to background, this shouldn't be a difficult problem. However, my network keeps overfitting after approximately 1-2 epochs each time I train.

I have approximately 23000 training patches. There is a slight class imbalance in the extracted patches. The pixel count is:

{'background'} : 1.6879e+08

{'bone' } : 2.2309e+07

What I've tried so far to avoid overfitting:

• I tried adding augmentation according to: aug = imageDataAugmenter('RandRotation', [-20, 20], 'RandScale', [0.7 1.5]); as well as other types of augmentation.
• I tried adding different amounts of 'L2Regularization'.
• I tried making the 'EnoderDepth' smaller in order to reduce the complexity of the model.

What more can I try to remove the overfitting when I'm training my network on this data using the unetLayers-architecture?

Answer 1

[Srivardhan Gadila]

The following are some suggestions:

1. Since there is a slight class imbalance, try the following: Balance Classes Using Class Weighting
2. Include dropoutLayer and batchNormalizationLayer in your architecture
3. Refer to Deep Learning Tips and Tricks