Cross-Modal Tumor Segmentation Using Generative Blending Augmentation and Self-Training.

Area of Science:

• Medical Imaging
• Computer Vision
• Machine Learning

Background:

• Data scarcity and domain shifts create biased training sets, failing to represent real-world deployment conditions.
• Cross-modal image segmentation aims to segment unlabeled images using labeled data from different imaging modalities.

Purpose of the Study:

• To develop a novel cross-modal segmentation method addressing data limitations.
• To improve the generalization of segmentation models in diverse imaging contexts.

Main Methods:

• Proposed a cross-modal segmentation approach using image synthesis enhanced by Generative Blending Augmentation (GBA).
• Utilized a SinGAN model within GBA to learn generative features from a single image, diversifying tumor appearances and mitigating synthesis errors.
• Integrated GBA with an iterative self-training procedure using pseudo-labels to further boost segmentation model generalization.

Main Results:

• The proposed method achieved first place in vestibular schwannoma (VS) segmentation at the MICCAI CrossMoDA 2022 challenge.
• Demonstrated superior performance with the best mean Dice similarity and average symmetric surface distance measures on validation and test sets.

Conclusions:

• Generative Blending Augmentation (GBA) effectively enhances segmentation model performance by improving data diversity and compensating for image synthesis limitations.
• The combination of local contrast alteration and iterative self-training shows promise for improving segmentation across various medical imaging applications.