Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Data Efficiency and Transfer Robustness in Biomedical Image Segmentation: A Study of Redundancy and Forgetting with Cellpose

Shuo Zhao, Jianxu Chen

TL;DR

The paper tackles data efficiency and transfer robustness in biomedical image segmentation by analyzing training data redundancy with Dataset Quantization (DQ) and catastrophic forgetting in cross-domain finetuning using Cellpose. It shows that 10–40% of the Cyto training data suffices for near-saturation performance, while naive cross-domain finetuning causes substantial forgetting; a lightweight replay strategy reintroducing 5–10% of source data mitigates forgetting and supports better generalization across domains. It also reveals that multi-stage transfer order significantly shapes retention and final performance, with curricula that start from diverse or intermediate domains yielding better balance between generalization and retention. The findings advocate data-centric training and curriculum-aware transfer as practical paths toward robust, scalable biomedical segmentation pipelines and provide actionable baselines for replay-based forgetting mitigation.

Abstract

Generalist biomedical image segmentation models such as Cellpose are increasingly applied across diverse imaging modalities and cell types. However, two critical challenges remain underexplored: (1) the extent of training data redundancy and (2) the impact of cross domain transfer on model retention. In this study, we conduct a systematic empirical analysis of these challenges using Cellpose as a case study. First, to assess data redundancy, we propose a simple dataset quantization (DQ) strategy for constructing compact yet diverse training subsets. Experiments on the Cyto dataset show that image segmentation performance saturates with only 10% of the data, revealing substantial redundancy and potential for training with minimal annotations. Latent space analysis using MAE embeddings and t-SNE confirms that DQ selected patches capture greater feature diversity than random sampling. Second, to examine catastrophic forgetting, we perform cross domain finetuning experiments and observe significant degradation in source domain performance, particularly when adapting from generalist to specialist domains. We demonstrate that selective DQ based replay reintroducing just 5-10% of the source data effectively restores source performance, while full replay can hinder target adaptation. Additionally, we find that training domain sequencing improves generalization and reduces forgetting in multi stage transfer. Our findings highlight the importance of data centric design in biomedical image segmentation and suggest that efficient training requires not only compact subsets but also retention aware learning strategies and informed domain ordering. The code is available at https://github.com/MMV-Lab/biomedseg-efficiency.

Data Efficiency and Transfer Robustness in Biomedical Image Segmentation: A Study of Redundancy and Forgetting with Cellpose

TL;DR

The paper tackles data efficiency and transfer robustness in biomedical image segmentation by analyzing training data redundancy with Dataset Quantization (DQ) and catastrophic forgetting in cross-domain finetuning using Cellpose. It shows that 10–40% of the Cyto training data suffices for near-saturation performance, while naive cross-domain finetuning causes substantial forgetting; a lightweight replay strategy reintroducing 5–10% of source data mitigates forgetting and supports better generalization across domains. It also reveals that multi-stage transfer order significantly shapes retention and final performance, with curricula that start from diverse or intermediate domains yielding better balance between generalization and retention. The findings advocate data-centric training and curriculum-aware transfer as practical paths toward robust, scalable biomedical segmentation pipelines and provide actionable baselines for replay-based forgetting mitigation.

Abstract

Generalist biomedical image segmentation models such as Cellpose are increasingly applied across diverse imaging modalities and cell types. However, two critical challenges remain underexplored: (1) the extent of training data redundancy and (2) the impact of cross domain transfer on model retention. In this study, we conduct a systematic empirical analysis of these challenges using Cellpose as a case study. First, to assess data redundancy, we propose a simple dataset quantization (DQ) strategy for constructing compact yet diverse training subsets. Experiments on the Cyto dataset show that image segmentation performance saturates with only 10% of the data, revealing substantial redundancy and potential for training with minimal annotations. Latent space analysis using MAE embeddings and t-SNE confirms that DQ selected patches capture greater feature diversity than random sampling. Second, to examine catastrophic forgetting, we perform cross domain finetuning experiments and observe significant degradation in source domain performance, particularly when adapting from generalist to specialist domains. We demonstrate that selective DQ based replay reintroducing just 5-10% of the source data effectively restores source performance, while full replay can hinder target adaptation. Additionally, we find that training domain sequencing improves generalization and reduces forgetting in multi stage transfer. Our findings highlight the importance of data centric design in biomedical image segmentation and suggest that efficient training requires not only compact subsets but also retention aware learning strategies and informed domain ordering. The code is available at https://github.com/MMV-Lab/biomedseg-efficiency.

Paper Structure

This paper contains 22 sections, 2 equations, 4 figures, 8 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Generalist Biomedical Image Segmentation Models
  4. Dataset Quantization and Coreset Selection
  5. Catastrophic Forgetting in Biomedical Transfer
  6. Methodology
  7. Datasets
  8. Dataset Quantization for Coreset Selection
  9. Training and Preprocessing Details
  10. Zero-Shot Generalization
  11. Cross Domain Finetuning and Forgetting Analysis
  12. Multi-Stage Sequential Transfer
  13. Simple Replay Strategy for Forgetting Mitigation
  14. Results and Analysis
  15. Redundancy in Training Data
  16. ...and 7 more sections

Figures (4)

  • Figure 1: Image segmentation performance metrics under DQ across increasing training data percentages (1%–100%) in Cyto.
  • Figure 2: Example image patches selected by dataset quantization (left) and random sampling (right) at 1% selection rate in Cyto trainset. DQ patches exhibit greater diversity in morphology and contrast.
  • Figure 3: t-SNE projection of MAE features from Cyto trainset, grouped into 5 bins. DQ samples (red) show broader coverage in the latent feature space than Random (golden), indicating higher visual diversity.
  • Figure 4: Performance degradation before and after cross-domain finetuning across six evaluation metrics. Catastrophic forgetting is observed in Cyto-to-Histo transfer, with source domain metrics dropping sharply. The reverse transfer results in milder forgetting.