Gradual Learning for Semi-Supervised 3D Segmentation

Johann Christopher Engster, Nele Blum, Laura Hellwege, Thorsten Buzug, Maik Stille

Keywords: 3D Segmentation, Semi-Supervised Learning, Few-Shot Learning

Abstract:

We propose an extension of the recently published Gradual Learning, a semi-supervised method for segmentation of slice-stacks. While the original Gradual Learning is based on 2D slices to leverage the high similarity within the local neighborhood, the extension utilizes 3D subvolumes instead.
Thus, a 3D segmentation network is trained on initial subvolumes and the corresponding ground truth. Afterward, pseudo labels of the expanded subvolumes are generated, which are reused for training. This process is repeated a set number of times.
The approach results in improved segmentation quality without the need for large expert-labeled data sets.
The method was evaluated on head magnetic resonance imaging scans for brain segmentation but can be easily transferred to other modalities.
The results showed high gains in Intersection over Union scores on a separate test data set (depending on the number of used subvolumes $n$: $n$=2: $0.30 \rightarrow 0.58$, $n$=3: $0.41 \rightarrow 0.63$, $n$=10: $0.55 \rightarrow 0.76$, training with full volume: $0.86$).

Submission Number: 13