SEG4SEG accepted as a MIDL 2026 spotlight

Our paper on discovering systematically underperforming subgroups in medical image segmentation was accepted as a spotlight at MIDL 2026.

How can we identify samples that systematically underperform in medical image segmentation tasks?

We address this question in our MIDL 2026 spotlight paper SEG4SEG: Identifying Systematic Failure Modes in Segmentation by Subgroup Discovery Methods, led by Nina Weng during her research stay in our group last summer.

Segmentation models can achieve great overall metrics but still fail systematically on hidden subgroups. For example, a segmentation algorithm might silently underperform on a group of images with a previously unknown artefact, or on a group of images that followed a different annotation style. SEG4SEG allows us to identify such underperforming groups by extending Slice Discovery Methods, which so far have been applied almost exclusively to classification, to segmentation.

Key contributions:

• SEG4SEG, an algorithm that can reliably identify various types of hidden subgroups in a wide variety of segmentation tasks.
• A comprehensive taxonomy of potential systematic failure modes in medical image segmentation.
• Principled success criteria for evaluating whether a subgroup discovery method actually found a meaningful subgroup.

Four categories of systematic failure modes in medical image segmentation covered by our taxonomy.

Joint work with Nina Weng, Eike Petersen, Alceu Bissoto, Susu Sun, Lisa M. Koch, Aasa Feragen and Siavash Bigdeli.

Paper: https://openreview.net/forum?id=dbZEIwiAon
Project page: https://nina-weng.github.io/seg4seg.github.io
Code: https://github.com/nina-weng/seg4seg