3D BRAIN TUMOR DETECTION
The findings show that pretraining models using 3D proxy tasks for various self-supervised learning approaches generates more rich semantic representations and enables completing downstream tasks more accurately and quickly compared to training the models from scratch and pretraining them on 2D slices.
Abstract
The efficacy of self-supervised learning techniques in a variety of application fields has helped them become more popular recently. In this work, we give 3D proxy tasks for five various self-supervised learning approaches using these principles. By enabling neural network feature learning from unlabelled 3D images, we hope to minimise the cost of expert annotation. The techniques include Relative 3D Patch Location, 3D Exemplar Networks, 3D Rotation Prediction, 3D Jigsaw Puzzles, and 3D Contrastive Predictive Coding. Our findings show that pretraining models using our 3D tasks generates more rich semantic representations and enables completing downstream tasks more accurately and quickly compared to training the models from scratch and pretraining them on 2D slices.