Generative augmentations for cardiac ultrasound segmentation - real-time demo

One of the main challenges in current research on segmentation in cardiac ultrasound is the lack of large and varied labeled datasets and the differences in annotation conventions between datasets. This makes it difficult to design robust segmentation models that generalize well to external datasets. This work utilizes diffusion models to create generative augmentations that can significantly improve diversity if the dataset and thus the generalizability of segmentation models without the need for labeled data. The augmentations are applied in addition to regular augmentations. A visual Turing test survey shows that experts can not distinguish between real and synthetic images. Using the proposed generative augmentations, Dice score was improved by up to 14\% on an external dataset and the limits of agreement of automatic ejection fraction estimation was improved by by up to 6\%. These improvements come exclusively from the improved diversity and realism of the training data, without modifying the underlying machine learning model.

ArXiv preprint coming soon!