---
title: "Fraunhofer MEVIS presents sparse autoencoders for H\u0026E patches at ECP 2026, Stockholmsmässan, Stockholm; improves survival risk estimation (c-index 0.71)"
sdDatePublished: "2026-08-13T13:07:00Z"
source: "https://www.mevis.fraunhofer.de/de/fairs-and-conferences/2026/ecp-2026.html"
topics:
  - name: "artificial intelligence"
    identifier: "medtop:20001298"
  - name: "medical research"
    identifier: "medtop:20000737"
locations:
  - "Stockholm"
  - "Sweden"
---


Fraunhofer MEVIS presents sparse autoencoders for H&E patches at ECP 2026, Stockholmsmässan, Stockholm; improves survival risk estimation (c-index 0.71)

ECP 2026 - Fraunhofer MEVIS

Veranstaltungsdetails Veranstaltungsort Stockholmsmässan Datum 12. September 2026 - 16. September 2026 Diesen Termin als iCal herunterladen

12. September 2026 - 16. September 2026

The European Society of Pathology (ESP) and the Swedish Society of Pathology warmly invite you to the 38th European Congress of Pathology (ECP 2026), taking place at Stockholmsmässan in Stockholm from 12–16 September 2026.

ECP 2026 will bring together more than 5,000 pathologists, clinicians, researchers, molecular biologists, bioinformaticians, medical scientists, and industry partners for professional development, networking, and knowledge exchange. Under the motto “Innovation Driving Excellence in Pathology,” the congress will highlight the impact of innovation on diagnostics, prognosis, therapy response, and clinical practice. The programme will feature keynote lectures, slide seminars, symposia, and interactive workshops led by internationally renowned experts, with a special focus on digital pathology, AI, molecular pathology, and emerging technologies.

Hosted in Sweden, a country with a strong tradition in medical science and internationally recognised expertise in molecular pathology and precision oncology, ECP 2026 promises an inspiring meeting where science, innovation, and culture come together.

This year, Fraunhofer MEVIS is represented with the following contributions:

003 "Decomposing H&E-patches into interpretable features using sparse autoencoders"

Sparse autoencoders decompose pathology foundation model embeddings into more disentangled features (only 0.5% vs. 18.2% strongly correlated dimension pairs) without sacrificing classification accuracy (~99% and ~89% on two patch datasets). These sparse representations also improve patient-level survival risk estimation compared to standard dense representations (c-index of 0.71 vs. 0.68 for NSCLC and 0.70 vs. 0.60 for HNSCC) and enable identification of biomarker-relevant dimensions correlating with specific nuclear morphologies (e.g., r = 0.69 for neoplastic cells).

(Authors:Till Nicke, Jan-Raphael Schäfer, Eike Petersen, Fleming Wolf-Eli Michelsen, WeberSteinhilber, Louisa Areta Flach, Leonard Simon Brandenburg, Joachim Georgii, Henning Höfener, Johannes Lotz)

Session: Computational Pathology Symposium, CP-01 AI essentials for pathologists

Sunday, September 13, 10:00 – 10:10 | K1

12. September 2026 - 16. September 2026

Diesen Termin als iCal herunterladen