Expert Branches: Module Diversity for Stronger Feature Learning in Laparoscopic Segmentation

Lin Guo, Chiara Camerota, Mohammad Mahmoud, Flavio Esposito

Keywords: Network design, Multi-Branch Network, Segmentation, Surgical Imaging.

TL;DR: Unlike networks that rely on large models or heavy pretraining, we verified the design of diverse modules as independent Expert Branches, can enable richer task-specific features with fewer parameters.

Abstract: Module diversity fundamentally enhances a model’s ability to learn geometric structure by enabling a broader and more expressive set of feature representations. While many architectures improve performance by scaling parameters or relying on large-scale pretraining, these strategies make it difficult to identify which design principles truly enhance feature learning capability, especially in challenging domains with limited data such as laparoscopic surgical segmentation. This work investigates a parameter-constrained, no-pretraining setting to isolate the intrinsic feature learning capability of different module configurations. We introduce expert branches, a design concept that assigns different module families to their own independent pathways rather than mixing all features within a single stream. This separation encourages branch-specific specialization (Experts), reduces parameters, and avoids the entanglement that commonly obscures each module’s contribution. We test this idea with TriEB, a UNet-based model incorporating CNN, deformable-convolution, and dynamic-snake branches with less total parameters. TriEB surpasses the vanilla UNet, the non-diverse TriCNN counterpart, and transformer-based models including SegFormer and Swin on the DSAD laparoscopic dataset. These results demonstrate that expert branches offer a more effective design principle for extracting diverse features from surgical imagery. The study highlights module diversity as a promising, architecture-agnostic framework for building efficient, interpretable, and data-adaptive feature extractors.

Primary Subject Area: Segmentation

Secondary Subject Area: Segmentation

Registration Requirement: Yes

Reproducibility: https://github.com/LinG16pr/ExpertBranches

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: 378