Preserving Marker Specificity Preserving Marker Specificity with Lightweight Channel-Independent Representation Learning

Simon Gutwein, Arthur Longuefosse, Jun Seita, Sabine Taschner-Mandl, Roxane Licandro

Keywords: Multiplex Imaging, Representation Learning, Cell Phenotyping

TL;DR: Small channel-isolated neural networks match large-scale foundation models for multiplex tissue imaging while delivering biologically accurate and interpretable cell phenotyping.

Registration Requirement: Yes

Abstract: Multiplex tissue imaging measures dozens of protein markers per cell, yet most deep learning models still apply early channel fusion, a design choice that erroneously imposes strong shared structure across weakly correlated channels. We present a systematic study of channel fusion as an architectural design choice for self-supervised representation learning. Using a Hodgkin lymphoma CODEX dataset with 145,000 cells and 49 markers, we demonstrate that preserving channel isolation yields substantially stronger representations than early fusion. Notably, a shallow channel-isolated architecture with only 5.5K parameters matches state-of-the-art models. We further demonstrate that channel-isolated embeddings enable interpretable phenotyping via marker attribution. Blind expert validation confirms that the proposed method produces annotations that respect biological constraints better than conventional pipelines (82.1\% vs. 54.1\% agreement). Our results show that delayed channel fusion can substitute for model scale in multiplex imaging, yielding interpretable representations that align more closely with biological ground truth and enable reliable expert-validated phenotyping. Code is available at https://github.com/SimonBon/CIM-S.

Reproducibility: Code is available at https://github.com/SimonBon/CIM-S

Visa & Travel: No

Read CFP & Author Instructions: Yes

Originality Policy: Yes

Single-blind & Not Under Review Elsewhere: Yes

LLM Policy: Yes

Submission Number: 43