A Multicenter Benchmarks of Multiple Instance Learning Models for Lymphoma Subtyping from HE Whole Slide Images

Rao Muhammad Umer, Daniel Sens, Jonathan Noll, Christian Matek, Lukas Wolfseher, Rainer Spang, Ralf Huss, Johannes Raffler, Sarah Reinke, Wolfram Klapper, Katja Steiger, Kristina Schwamborn, Carsten Marr

Keywords: Multicenter Lymphomas Benchmark, Multiple Instance Learning, Whole Slide Images, Pathology Foundation Models

Abstract:

Timely and accurate lymphoma diagnosis is essential for guiding treatment. In standard diagnostic practice, the initial assessment of hematoxylin and eosin (HE)-stained whole slide images is followed by integration of immunohistochemistry (IHC) stained sections, flow cytometry, and molecular genetic tests to determine the specific lymphoma subtype. These auxiliary tests require costly equipment, expensive reagents, skilled personnel, and often lead to treatment delays. Automated deep learning-based computational methods offer an opportunity to assist pathologists by extracting maximal diagnostic information from routinely available HE-stained whole slides. While previous approaches have shown promise in analyzing HE-stained images and achieving high accuracies for a limited set of lymphoma subtypes, they have not yet demonstrated sufficient clinical utility for real-world diagnostic workflows.
In this work, we extend previous approaches by presenting the first multicenter lymphoma benchmarking dataset, covering four common lymphoma subtypes and healthy control tissue. We systematically compare attention-based and transformer-based multiple instance learning models with public pathology foundation models across multiple image magnifications. On in-distribution test sets, our attention-based models achieve multiclass accuracies exceeding 80% across all magnifications; on out-of-distribution test sets, performance remains around 60%. To support further research and robust model evaluation, we provide an automated benchmarking pipeline for external use.

Primary Subject Area: Application: Histopathology

Secondary Subject Area: Foundation Models

Registration Requirement: Yes

Visa & Travel: Yes

Read CFP & Author Instructions: Yes

Originality Policy: Yes

Single-blind & Not Under Review Elsewhere: Yes

LLM Policy: Yes

Submission Number: 39