Beyond Visible Boundaries: Benchmarking Foundation Models for Overlapping Cell Segmentation in Microscopic Imaging

Vo Minh Khoi Nguyen, Khanh Nguyen Vo Ngoc, Vi Vu, Thien Nguyen, Thanh-Huy Nguyen, Min Xu

Keywords: Benchmark, Overlapping cell segmentation, Segment Anything Model

TL;DR: We introduce a controlled benchmark for overlapping cell segmentation with sub-region decomposition and a new metric, revealing that foundation models fail at recovering occluded regions despite strong full-mask performance.

Abstract: Overlapping cell segmentation remains a critical bottleneck in computational pathology, yet existing benchmarks evaluate models exclusively on clean, non-overlapping structures, and standard metrics conflate visible and hidden anatomy into a single score, masking fundamental amodal completion failures. In this work, we present an algorithm for synthesising overlapping cell occlusion that is applicable across diverse cell imaging datasets. Building on this, we construct the first controlled benchmark, to the best of our knowledge, for stress-testing foundation models under overlapping cell occlusion, generating images from the ISBI-2014 (cervical cytology) and SegPC-2021 (plasma-cell microscopy) datasets across three severity levels: Light, Medium, and Heavy. We further propose a decomposed evaluation protocol that partitions each cell mask into three regions: the full mask, the non-overlap sub-region, and the overlap sub-region. Standard metrics such as Dice and Precision are inappropriate in this setting due to mis-penalization across sub-regions, leading to systematic bias. To address this, we propose a fixed-prior recall-weighted F-measure that computes the weighted harmonic mean between sub-region recall and full-mask precision. Experiments with SAM, SAM2, MedSAM, and MedSAM2 show that medical SAM-based models are more effective at recovering the overlap sub-region while remaining competitive on the non-overlap sub-region.

Email Sharing: We authorize the sharing of all author emails with Program Chairs.

Data Release: We authorize the release of our submission and author names to the public in the event of acceptance.

Submission Number: 33