Edge-Based Self-supervision for Semi-supervised Few-Shot Microscopy Image Cell Segmentation

Youssef Dawoud, Katharina Ernst, Gustavo Carneiro, Vasileios Belagiannis

Abstract

Deep learning models for microscopy image segmentation typically require extensive manual annotation, which is time-consuming and requires domain expertise. This paper introduces a novel semi-supervised framework that utilizes a simple edge-detection proxy task—based on the Canny filter—to provide self-supervision on unlabeled images. By employing a multi-task learning architecture with a shared encoder, the model learns robust morphological features that facilitate high-performance segmentation even with limited labeled data. Our results demonstrate that this approach, using only 10% of available labels, can match the performance of fully supervised models and significantly outperforms standard self-supervised methods in few-shot adaptation tasks.

Introduction

Microscopy image analysis is a cornerstone of biological research, yet the transition to deep learning-based segmentation is often hindered by the "data hunger" of neural networks. Generating thousands of pixel-level annotations is an expensive and tedious process. Furthermore, the high variability between microscopy modalities—such as the difference between fluorescent staining and electron microscopy—means that models trained on one domain often fail to generalize to another.

This creates a critical need for "few-shot" adaptation: the ability to take a model and quickly specialize it for a new imaging environment using only a handful of labeled examples (e.g., 1 to 10 images). This paper addresses this challenge by leveraging the vast amounts of unlabeled microscopy data typically available in laboratory settings through a targeted self-supervised task.

Methodology

The Edge-Based Proxy Task

The authors propose that for biological cells, shape and boundary information are the most salient features for segmentation. Unlike general computer vision tasks that might use arbitrary self-supervision (like image rotation or jigsaw puzzles), this method uses a Canny edge detector to automatically generate edge maps from unlabeled images. These maps act as "pseudo-labels" for a self-supervised training phase.

Multi-Task Architecture

The framework utilizes a Fully Convolutional Regression Network (FCRN) with a specific multi-task configuration:

  1. Shared Encoder: A single encoder extracts features from the input images.
  2. Segmentation Decoder: This branch is trained on a small set of manually labeled images to predict precise cell masks.
  3. Edge Decoder: This branch is trained on the larger pool of unlabeled images to reconstruct the Canny edge maps.

By training these components simultaneously, the encoder is forced to learn features that are useful for both identifying boundaries and defining whole-cell regions. This joint training regularizes the model and prevents it from overfitting to the small labeled dataset.

Few-Shot Adaptation

Once the joint training is complete, the edge decoder is discarded. The remaining encoder and segmentation decoder are then fine-tuned on a "target" dataset consisting of very few labeled images (the "shots"). This allows the model to adapt its learned morphological knowledge to the specific textures and contrasts of a new microscopy modality.

Experimental Results

Performance Parity with 10% Labels

A key finding of the study is the efficiency of the edge-based task. On the TNBC (Triple Negative Breast Cancer) histology dataset, the proposed method achieved a Mean Intersection over Union (IoU) of approximately 49.5% using only 10 labeled images. This performance is nearly identical to (and in some cases exceeds) the ~48.5% IoU achieved by a fully supervised model trained on the entire labeled set. This suggests that the self-supervised edge task captures most of the structural information necessary for accurate segmentation.

Comparison with Other SSL Methods

The researchers compared their edge-based approach against popular self-supervised learning (SSL) techniques, including SimCLR (contrastive learning) and Rotation Prediction. The edge-based task consistently outperformed these methods, particularly in 1-shot and 3-shot scenarios. This highlights the importance of choosing a proxy task that is domain-relevant; while rotation might help a model understand natural objects, edge detection is more directly aligned with the requirements of biological image segmentation.

The Over-Regularization Phenomenon

Interestingly, the study observed that using 100% of the available unlabeled data was not always superior to using a smaller subset (e.g., 60%). If the proxy task (edge detection) becomes too dominant during training, it can act as an over-regulator, potentially distracting the model from the primary goal of segmentation. This suggests that the ratio of labeled to unlabeled data is a critical hyperparameter in semi-supervised workflows.

Conclusion and Future Work

This research demonstrates that edge-based self-supervision is a powerful tool for reducing the annotation burden in microscopy. By focusing on boundaries—a fundamental characteristic of cells—the model achieves high-quality results with a fraction of the human effort.

However, the method's reliance on the Canny filter remains a limitation, as it may struggle with extremely noisy or low-contrast images where the filter fails to produce meaningful pseudo-labels. Future research could explore more sophisticated morphological proxy tasks or adaptive weighting mechanisms to better balance the segmentation and edge-detection objectives across different datasets.

Summary updated

BibTeX
@inproceedings{dawoud2022edge,
  author = {Youssef Dawoud and Katharina Ernst and Gustavo Carneiro and Vasileios Belagiannis},
  title = {{Edge-Based Self-supervision for Semi-supervised Few-Shot Microscopy Image Cell Segmentation}},
  booktitle = {MOVI @ MICCAI},
  year = {2022},
  doi = {10.1007/978-3-031-16961-8\_3},
}
Figure