Rubik’s Cube
Due to the annotation of 3D medical data being hard to acquire, the number of annotated 3D images for training is often not enough.
Self-supervised learning deeply exploiting the information of raw data can be a solution to loose requirements for training.
This paper proposes a method like solving a Rubik’s Cube to apply self-supervised learning to learn relevant features.
Architecture
First, the 3D input is partitioned into 2x2x2(=8) cubes. Then permute the cubes with random rotations. Finally, like playing a Rubik’s cube, the model will try to recover the original configuration. In order to recover the cube, there is a loss for ordering and a loss for orientation.
Reference: Self-supervised Feature Learning for 3D Medical
Images by Playing a Rubik’s Cube
