Research existing methods for performing image segmentation of neuroimaging data

[mcmasterai]

Research existing methods for performing image segmentation of neuroimaging data

[dufaultc]

Highly recommend reading this paper . Provides recent and clearly explained overview of current research in using CNN's for tumour segmentation This paper is also worth reading to get more general overview of tumour segmentation methods

Summary

• Three ways to perform tumour segmentation: Manual, Semi-Automatic, Fully-Automatic

• Manual is relatively accurate but suffers from high variability in results between radiologists and is time-consuming. However, manual segmentation is used to evaluate other methods

• Semi-Automatic methods are less time-consuming than manual but still suffer from high variability between radiologists

• Fully-Automatic methods can be divided into those defining custom features and CNNS

• CNN based methods have proven to be the most accurate

• One tumours are segmented they can be further investigated to determine response to therapy, type of cancer, etc

• CNNs use convolutional layers to apply filters to input image which serve to highlight certain features of the image

• There are a few common methods used to stop the overfitting of CNNs to the input data

  • Data Augmentation: Changing the orientation and zoom of input images to stop CNN from relying on artifacts of input data

  • Dropout: Dropping nodes of the CNN to introduce network imprecision

  • Batch Normalization: Stabilizes CNN training by normalizing the affect individual nodes of model have; helpful as some may become highly biased

  • Pooling: Downsampling the image to force CNN to learn more imprecise features of the input dataset

• Output