Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

CyclePose -- Leveraging Cycle-Consistency for Annotation-Free Nuclei Segmentation in Fluorescence Microscopy

Jonas Utz, Stefan Vocht, Anne Tjorven Buessen, Dennis Possart, Fabian Wagner, Mareike Thies, Mingxuan Gu, Stefan Uderhardt, Katharina Breininger

TL;DR

CyclePose introduces an annotation-free nuclei segmentation framework by embedding a Cellpose-based segmentation model within a CycleGAN architecture and using cycle-consistency as self-supervision. The method replaces the typical two-stage workflow of synthetic data generation followed by segmentation training with a unified, unsupervised training loop augmented by Perlin and mask-to-image losses. Empirical results on BBBC039/U-2 OS and BBBC038/DSB2018 show CyclePose achieving state-of-the-art performance among unsupervised methods and competitive results relative to supervised baselines, with notable stability from the proposed losses. The approach reduces computational overhead and brings annotation-free segmentation closer to real-world applicability, though it relies on ellipse-based mask syntheses and modest annotated data for model selection. Future directions include extending to 3D segmentation and exploring richer shape priors beyond ellipses to generalize to diverse cell types.

Abstract

In recent years, numerous neural network architectures specifically designed for the instance segmentation of nuclei in microscopic images have been released. These models embed nuclei-specific priors to outperform generic architectures like U-Nets; however, they require large annotated datasets, which are often not available. Generative models (GANs, diffusion models) have been used to compensate for this by synthesizing training data. These two-stage approaches are computationally expensive, as first a generative model and then a segmentation model has to be trained. We propose CyclePose, a hybrid framework integrating synthetic data generation and segmentation training. CyclePose builds on a CycleGAN architecture, which allows unpaired translation between microscopy images and segmentation masks. We embed a segmentation model into CycleGAN and leverage a cycle consistency loss for self-supervision. Without annotated data, CyclePose outperforms other weakly or unsupervised methods on two public datasets. Code is available at https://github.com/jonasutz/CyclePose

CyclePose -- Leveraging Cycle-Consistency for Annotation-Free Nuclei Segmentation in Fluorescence Microscopy

TL;DR

CyclePose introduces an annotation-free nuclei segmentation framework by embedding a Cellpose-based segmentation model within a CycleGAN architecture and using cycle-consistency as self-supervision. The method replaces the typical two-stage workflow of synthetic data generation followed by segmentation training with a unified, unsupervised training loop augmented by Perlin and mask-to-image losses. Empirical results on BBBC039/U-2 OS and BBBC038/DSB2018 show CyclePose achieving state-of-the-art performance among unsupervised methods and competitive results relative to supervised baselines, with notable stability from the proposed losses. The approach reduces computational overhead and brings annotation-free segmentation closer to real-world applicability, though it relies on ellipse-based mask syntheses and modest annotated data for model selection. Future directions include extending to 3D segmentation and exploring richer shape priors beyond ellipses to generalize to diverse cell types.

Abstract

In recent years, numerous neural network architectures specifically designed for the instance segmentation of nuclei in microscopic images have been released. These models embed nuclei-specific priors to outperform generic architectures like U-Nets; however, they require large annotated datasets, which are often not available. Generative models (GANs, diffusion models) have been used to compensate for this by synthesizing training data. These two-stage approaches are computationally expensive, as first a generative model and then a segmentation model has to be trained. We propose CyclePose, a hybrid framework integrating synthetic data generation and segmentation training. CyclePose builds on a CycleGAN architecture, which allows unpaired translation between microscopy images and segmentation masks. We embed a segmentation model into CycleGAN and leverage a cycle consistency loss for self-supervision. Without annotated data, CyclePose outperforms other weakly or unsupervised methods on two public datasets. Code is available at https://github.com/jonasutz/CyclePose

Paper Structure

This paper contains 11 sections, 4 equations, 2 figures, 2 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Methods
  4. Evaluation and Results
  5. Datasets
  6. Training
  7. Compared Methods
  8. Evaluation Metrics
  9. Results
  10. Discussion
  11. Acknowledgments.

Figures (2)

  • Figure 1: Overview of the cycles in CyclePose: Synthetic masks from rotated ellipses undergo elastic deformation to create training masks. Adding Perlin noise-based texture, Gaussian blur, and Poisson noise produces Perlin images, which serve as auxiliary inputs for Cellpose alongside fake and real images. Discriminators are omitted for brevity. During inference only $S$ is used.
  • Figure 2: Qualitative comparison of CyclePose against top-performing unsupervised and supervised baseline methods. Methods trained with supervision are marked with $^\star$.