BioImage Computing Workshop at ECCV 26 · 2026

Unsupervised Visual Concept Bottlenecks for Interpretable Bioimage Classification

Linghui Liu, Henrike Stephani, Jördis Sieburg-Rockel, Stephanie Helmling, Andrea Olbrich, Janis Keuper

Why this publication matters

Biologists need to see which structures support an image classifier’s answer. This study discovers visual concepts without manually labeling each concept, then checks them against expert judgments and biological attributes across three microscopy datasets. The models remain competitive at prediction while exposing evidence that often matches recognizable biological structures.

Abstract

Concept Bottleneck Models (CBM) have recently shown promising results towards more interpretable image classification. However, the application of this explainable learning method in bioimage analysis has so far been limited by the cost of expert concept annotation. We therefore investigate whether unsupervised visual concept bottlenecks can offer an interpretable alternative for microscopy image classification by discovering concepts from local image regions rather than relying on manually defined labels. Our study covers three bioimage datasets representing softwood anatomical features, white blood cell morphology, and colorectal histopathology. We compare task-trained representations with frozen foundation-model features adapted through a lightweight projection layer, and evaluate both dense and sparse bottlenecks. Beyond classification performance, we assess the biological quality of the discovered concepts using human expert verifications and available attribute annotations. Our results show that, across datasets, unsupervised concept bottlenecks retain competitive predictive performance while producing inspectable visual concepts, which often align with known biological structures.

Abstract source ↗

Figures

Microscopy examples from softwood anatomy, white blood cells, and colorectal tissue datasets. — **Figure 1.** Examples from the three microscopy datasets. Each column shows two examples from one class. The Softwood dataset contains five anatomical feature classes, and all patches are shown at the same scale. WBCAtt contains five white blood cell types, while NCT-CRC contains nine colorectal tissue classes. View in source ↗

Discovered softwood concepts at different image grid sizes, with expert-assessed quality differences. — **Figure 2.** Softwood concept visualizations across grid sizes. Each block shows the top-12 activating patches of one discovered concept, as inspected by the domain experts. Patches are resized to a common display size. All examples are from the bordered class, where bordered pits appear as small ring-like patterns around a bright center. The 2×2 and 3×3 examples are expert-good concepts, while the 4×4 examples are weak concepts, often capturing only partial or blurred bordered-pit morphology. View in source ↗

Cite this paper

Download .bib

@inproceedings{liu2026unsupervisedvisualconcept97,
  title = {{Unsupervised Visual Concept Bottlenecks for Interpretable Bioimage Classification}},
  author = {Linghui Liu and Henrike Stephani and Jördis Sieburg-Rockel and Stephanie Helmling and Andrea Olbrich and Janis Keuper},
  booktitle = {BioImage Computing Workshop at ECCV 26},
  year = {2026},
  url = {https://openreview.net/pdf?id=hois3HoeYg}
}

Figures and abstract are reproduced from the linked research sources. Credit remains with the authors and publishers.