Funded by: Afeyan Family Foundation
PIs: Varduhi Yeghiazaryan, Irina Voiculescu
The project focused on scribble-supervised medical image segmentation through a hybrid approach, combining deep learning models with image processing techniques from classical computer vision. The proposed approach enhanced minimally labeled training images by propagating annotation labels through the image, keeping the annotation process cost-efficient, while increasing the amount of supervision data available for training.
The approach utilized a hierarchical partitioning of the image produced with watershed/waterfall transforms. Semantic labels from scribbles were propagated to other pixels within the same waterfall region, substantially increasing the number of labeled pixels that could be used for training supervision. The project included experiments on public semantic segmentation datasets, including ACDC and MSCMRseg, to evaluate the effectiveness of the annotation enhancement approach. The work has been built on a previous collaboration with an aim to further develop the proposed methodology and disseminate the results through scientific publications.