Towards annotation-efficient segmentation via image-to-image translation.

Area of Science:

• Medical Imaging
• Deep Learning
• Computational Biology

Background:

• Deep learning for medical image segmentation requires extensive annotated data, which is costly and time-consuming to acquire.
• Tumor segmentation, in particular, demands expert labeling of every slice in volumetric scans, posing a significant bottleneck.

Purpose of the Study:

• To introduce a novel semi-supervised framework, GenSeg, for training segmentation models with minimal annotated data.
• To leverage weak labels (presence of a tumor) and generative domain translation to improve tumor segmentation accuracy.

Main Methods:

• GenSeg disentangles tumor information from healthy tissue in latent representations.
• It performs diseased-to-healthy image translation and simultaneously outputs tumor segmentation masks.
• Implicit data augmentation is achieved through healthy-to-diseased image translation.

Main Results:

• GenSeg trained U-Net-like architectures achieved superior performance on synthetic, brain tumor (MRI), and liver metastasis (CT) datasets.
• Improvements of 8-14% Dice score on brain tumors and 5-8% on liver tumors were observed with only 1% annotated data.
• Outperformed baseline semi-supervised and supervised methods in low-annotation scenarios.

Conclusions:

• GenSeg effectively addresses the challenge of training deep segmentation models with large amounts of unlabeled and unpaired medical data.
• The framework is particularly suitable for tumor segmentation tasks where annotated data is scarce.