Semantic segmentation with spreading scribbles

Conference item

Hand-annotating medical images with segmentation masks requires an immense amount of time and effort from clinical experts. Replacing full masks with a simpler annotating gesture can mitigate annotation costs. This can come in the form of a scribble, and leads to weakly supervised training scenarios. Scribble-supervised segmentation typically utilises advanced neural architectures to compensate for the limited training data. Instead of just relying strictly on the pixels from each scribble, we also enhance each scribble by spreading, i.e. propagating, annotation labels through the image. We use a hierarchical partitioning of the image, produced with watershed/waterfall transforms, and propagate the individual pixel labels through the waterfall regions. We propose that a semantic label can be propagated to all other pixels in the same waterfall region. This increases the number of pixels that can be used for training supervision. We show experimentally that this technique greatly boosts the performance of established neural architectures on public semantic segmentation datasets like ACDC and MSCMRseg.

 

Authors: Gabrielyan, Y; Yeghiazaryan, V; Voiculescu, I

• Publication status: Published
• Peer review status: Peer reviewed
• Publisher: Springer Nature
• Host title: Medical Image Understanding and Analysis
• Volume: 15918
• Series: Lecture Notes in Computer Science
• Publication date: 2025-07-15
• Acceptance date: 2025-05-15
• Event title: 29th UK Conference on Medical Image Understanding and Analysis (MIUA) 2025
• Event location: Leeds, UK
• Event website: https://conferences.leeds.ac.uk/miua/
• Event start date: 2025-07-15
• Event end date: 2025-07-17
• DOI: 10.1007/978-3-031-98694-9_6
• EISSN: 1611-3349
• ISSN: 0302-9743
• EISBN: 9783031986949
• ISBN: 9783031986932
• Language: English
• Keywords: waterfall transform; spreading scribbles; scribble-supervised segmentation