Oriented Cell Dataset: A Dataset and Benchmark for Oriented Cell Detection and Applications

Lucas Kirsten, Angelo Angonezi, Jose Marques, Fernanda Oliveira, Juliano Faccioni, Camila Cassel, Débora de Sousa, Samlai Vedovatto, Guido Lenz, Claudio JungWACV 2025

TL;DR

This paper introduces the Oriented Cell Dataset (OCD), a novel benchmark for cell detection in bright-field microscopy images using Oriented Bounding Boxes (OBBs). By demonstrating that OBBs offer a more accurate representation of cell shape than traditional Horizontal Bounding Boxes (HBBs) with minimal annotation overhead, the authors pave the way for improved automated analysis of crucial biological processes like cell confluence and polarity.

What Problem Does This Paper Address?

Accurately detecting and analyzing cells in microscopy images is fundamental for numerous biological and medical applications, including studying cell growth, migration, and disease progression. While bright-field microscopy is advantageous due to its non-invasiveness and affordability, it presents challenges in automated analysis due to low contrast and potential artifacts. Existing methods often rely on Horizontal Bounding Boxes (HBBs), which are simple to annotate but provide a poor representation of cell shape, especially for elongated or oriented cells. Full segmentation masks offer better shape representation but are prohibitively labor-intensive to create. This paper addresses the need for a more accurate yet efficient cell representation and a corresponding dataset to benchmark detection methods.

Methods

The core contribution of this work lies in the introduction of Oriented Bounding Boxes (OBBs) as a superior representation for cell shapes in microscopy images compared to HBBs. OBBs, which include an angle parameter, capture the orientation and aspect ratio of cells more effectively. The authors also explore Oriented Ellipses (OEs) as a further refinement. To evaluate these representations and benchmark detection algorithms, they introduce the Oriented Cell Dataset (OCD). The paper then benchmarks 14 state-of-the-art oriented object detectors on OCD, exploring their performance on two key biological applications: cell confluence and cell polarity estimation. They also conduct an inter-annotator assessment (IAA) to establish a suitable Intersection over Union (IoU) threshold for evaluating OBB detectors.

Datasets

  • Oriented Cell Dataset (OCD): A newly introduced dataset specifically designed for oriented cell detection. It contains 160 bright-field microscopy images of cancerous human cells, annotated with OBBs. The dataset is split into training (120 images), testing (30 images), and an annotator's comparison set (10 images). The annotator's comparison set features annotations from five independent experts to assess inter-annotator agreement.
  • Cell Tracking Challenge (CTC) datasets: Used for evaluating the representation comparison between HBBs, OBBs, and OEs against segmentation masks.

Key Insights and Ideas

  • OBBs offer a better shape representation: OBBs provide a significantly more accurate representation of cell shape than HBBs, leading to higher IoU scores when compared to segmentation masks. This is particularly important for elongated or irregularly shaped cells.
  • Minimal annotation cost for OBBs: Annotating OBBs requires only a slight increase in effort (an estimated "extra click") compared to HBBs, making them a practical choice for large-scale datasets.
  • Oriented Ellipses (OEs) further improve overlap: OEs, derived from OBBs, show even better overlap with segmentation masks, suggesting they are a highly effective representation for cell shapes, especially for rounder cells.
  • Human variability necessitates careful IoU thresholding: The inter-annotator assessment revealed that human annotators can have significant disagreements, highlighting the need for robust evaluation metrics and appropriate IoU thresholds. A threshold of 0.5 is proposed as a suitable starting point for OBB detection in this context.
  • OBBs enable new biological analyses: The ability of OBBs to capture cell orientation is crucial for applications like cell polarity estimation, which is not feasible with HBBs.

Key Results

  • Representation comparison: OBBs achieved 20.7% higher IoU than HBBs, and OEs achieved 37.7% higher IoU than HBBs when compared to segmentation masks on CTC datasets.
  • Inter-annotator assessment: Krippendorff's Alpha scores indicated good to very good agreement among human annotators (0.760 for class-aware and 0.794 for class-agnostic). The analysis suggested an IoU threshold of 0.55 as a reasonable upper bound for OBB detectors.
  • Object detection benchmark: Oriented RepPoints achieved the highest Average Precision (AP) in both class-aware (86.28%) and class-agnostic (78.98%) evaluations on the OCD dataset.
  • Biological applications:
    • Confluence estimation: ReDet and RoI Transformer achieved the lowest Mean Relative Error (MRE) below 10%.
  • Polarity estimation: Oriented RepPoints achieved the best result, closely followed by RoI Transformer and ReDet, with a mean $\chi^2$ distance compatible with human variability.

Limitations and Future Work

  • Dataset size: While OCD is valuable, it is relatively small compared to some other benchmarks. The authors suggest that larger datasets would be beneficial for training more robust models.
  • Annotation variability: The study acknowledges that human annotator variability can impact evaluation metrics. While they propose an IoU threshold, further research into robust evaluation methods that account for this variability is warranted.
  • Generalizability: The current study focuses on cancerous human cells. Future work could explore the applicability of OBB detection to a wider range of cell types and imaging conditions.
  • Exploration of other representations: While OBBs and OEs show promise, further investigation into other shape representations and their trade-offs with annotation effort could be valuable.
  • Real-time performance: While some methods show good accuracy, achieving real-time performance for live cell tracking applications remains an ongoing challenge.

Summary updated

BibTeX
@inproceedings{kirsten2025oriented,
  author = {Lucas Kirsten and Angelo Angonezi and Jose Marques and Fernanda Oliveira and Juliano Faccioni and Camila Cassel and Débora de Sousa and Samlai Vedovatto and Guido Lenz and Claudio Jung},
  title = {{Oriented Cell Dataset: A Dataset and Benchmark for Oriented Cell Detection and Applications}},
  booktitle = {IEEE/CVF Winter Conference on Applications of Computer Vision},
  year = {2025},
}
Figure