A Multi-Scale Inception-UNet with Structure-Aware Evaluation for Branch-Preserving Segmentation of Organoids

Sandra H. Andrusca, Christopher D. Kießling, Andreas R. Bausch

Keywords: semantic segmentation, organoid imaging, PDAC, multi-scale architectures, UNet, topology preservation, topology-aware loss, clDice, brightfield microscopy, multi-seed evaluation, branch morphology

TL;DR: We propose a multi-scale Inception-UNet for segmenting complex branched organoids and demonstrate statistically significant improvements over classical UNet-based architectures, particularly in preserving structural continuity.

Abstract:

Branched organoids exhibit increasingly complex morphologies as they progress from simple
spheroid states to highly ramified structures, making topology-preserving segmentation
essential for quantitative biological analysis. Capturing thin protrusions and maintaining
branch continuity remains challenging for classical UNet-based architectures, particularly
in brightfield imaging where fine structures are easily blurred or disconnected.
In this work, we present a multi-scale Inception-UNet designed to capture the het-
erogeneous spatial scales of branched organoids through parallel convolutional paths with
complementary receptive fields. As a model system, we analyze brightfield pancreatic duc-
tal adenocarcinoma (PDAC) organoids, a system known for strong morphological hetero-
geneity and invasive branching behavior (Randriamanantsoa, 2022), cultured using high-
throughput Patternoid assays (Kurzbach, 2025) that enable standardized imaging and ro-
bust quantitative analysis.
To assess segmentation quality beyond region overlap, we combine Dice with the
structure-aware clDice metric that directly probes branch integrity and topological conti-
nuity. Across deterministic seeds and strictly separated organoid positions, the Inception-
UNet achieves the highest region-based Dice (0.868 ± 0.062) and clDice (0.545 ± 0.123), and
most importantly, the strongest preservation of branch continuity compared to UNet and
UNet++. These improvements become increasingly pronounced with growing morpholog-
ical complexity.
Overall, our results demonstrate that multi-scale feature extraction combined with
topology-aware evaluation substantially improves segmentation of branched organoids and
provides a robust foundation for downstream morphological and invasion-related analyses.

Primary Subject Area: Segmentation

Secondary Subject Area: Learning with Noisy Labels and Limited Data

Registration Requirement: Yes

Visa & Travel: No

Read CFP & Author Instructions: Yes

Originality Policy: Yes

Single-blind & Not Under Review Elsewhere: Yes

LLM Policy: Yes

Midl Latex Submission Checklist: Ensure no LaTeX errors during compilation., Replace NNN with your OpenReview submission ID., Includes \documentclass{midl}, \jmlryear{2026}, \jmlrworkshop, \jmlrvolume, \editors, and correct \bibliography command., Did not override options of the hyperref package., Did not use the times package., Use the correct spelling and format, avoid Unicode characters, and use LaTeX equivalents instead., Any math in the title and abstract must be enclosed within $...$., Did not override the bibliography style defined in midl.cls and did not use \begin{thebibliography} directly to insert references., Avoid using \scalebox; use \resizebox when needed., Included all necessary figures and removed *unused* files in the zip archive., Removed special formatting, visual annotations, and highlights used during rebuttal., All special characters in the paper and .bib file use LaTeX commands (e.g., \'e for é)., No separate supplementary PDF uploads., Acknowledgements, references, and appendix must start after the main content.

Submission Number: 394