Disconnect to Connect: A Data Augmentation Method for Improving Topology Accuracy in Image Segmentation

Juan Miguel Valverde, Maja Østergaard, Adrian Rodriguez-Palomo, Peter Alling Strange Vibe, Nina Kølln Wittig, Henrik Birkedal, Anders Dahl

Keywords: topology, segmentation, data augmentation, inpainting

Abstract:

Accurate segmentation of thin, tubular structures remains a significant challenge for deep neural networks, where minor misclassifications can lead to broken connections and topologically incorrect results. Current approaches for improving topology accuracy, such as topology loss functions, require training labels that are both precise and topologically accurate. However, accurate labels are hard to obtain since image annotation is laborious, time-consuming, and the images' low resolution and low contrast can make tubular structures appear disconnected. We present CoLeTra, a data augmentation method that improves topology accuracy, even on datasets with topologically inaccurate training labels. CoLeTra achieves this by artificially breaking the structures in the images while maintaining the original labels, teaching models that structures can be continuous even if they appear disconnected. We evaluated CoLeTra on three datasets with six loss functions and two architectures, demonstrating that CoLeTra generally improved topology accuracy while often improving the Dice coefficient and clDice. We also release a dataset specifically suited for image segmentation methods with a focus on topology accuracy.
CoLetra's code can be found at https://github.com/jmlipman/CoLeTra.

Primary Subject Area: Segmentation

Secondary Subject Area: Application: Other

Registration Requirement: Yes

Visa & Travel: No

Read CFP & Author Instructions: Yes

Originality Policy: Yes

Single-blind & Not Under Review Elsewhere: Yes

LLM Policy: Yes

Submission Number: 15