Cell Detection Using Extremal Regions in a Semisupervised Learning Framework

Nisha Ramesh, Ting Liu, Tolga TaşdizenJournal of Healthcare Engineering 2017 · DOI

TL;DR

This paper introduces a semisupervised learning framework for cell detection that significantly reduces the need for extensive manual annotation. By leveraging a hierarchical representation of image regions and a novel unsupervised loss function, the method achieves high accuracy with minimal "dot" annotations, making cell detection more accessible and efficient.

What Problem Does This Paper Address?

Accurately detecting cells in microscopy images is crucial for various biological studies, but it's a challenging task. Traditional supervised methods require a large amount of meticulously annotated training data (e.g., precise cell masks), which is time-consuming and labor-intensive to create. This paper addresses the problem of reducing this annotation burden by developing a semisupervised approach that can learn effectively from a small set of simple annotations (just a dot inside each cell) and a larger pool of unlabeled data.

Methods

The core of the proposed method lies in a semisupervised learning framework built upon a hierarchical representation of image regions. Here's a breakdown of the key techniques:

  • Extremal Region Generation: The process starts by using the Maximally Stable Extremal Region (MSER) detector on probability maps (generated by ILASTIK, a pixel classifier trained on minimal annotations) to identify candidate cell regions. MSER is effective at finding regions with distinct intensity contrasts.
  • Hierarchical Representation: These MSER regions are organized into a hierarchical tree structure, where nested regions represent different levels of detail. This tree captures the spatial relationships and potential overlaps between candidate regions.
  • Semisupervised Learning Framework: The learning process combines both supervised and unsupervised components within a Bayesian framework.
  • Supervised Loss: This component uses the simple "dot" annotations (cell centers) to train a classifier (an Artificial Neural Network - ANN) to predict the probability of a region being a cell.
  • Unsupervised Loss (Path Consistency Constraint): This is the novel contribution. The authors develop a differentiable unsupervised loss term that enforces the constraint that at most one cell should be detected within any given path in the MSER hierarchical tree. This directly addresses the issue of overlapping detections and ensures a more biologically plausible segmentation. This loss is formulated using DeMorgan's laws and products to make it differentiable.
  • Parameter Estimation: Model parameters (weights of the ANN and standard deviations of the loss terms) are learned using maximum a posteriori (MAP) estimation, minimizing a cost criterion.
  • Inference: After training, an inference procedure (either greedy or a bottom-up/top-down dynamic programming approach) is used on the MSER tree to select a set of non-overlapping regions that best represent the detected cells, guided by the learned probabilities and the path consistency constraint.

Datasets

The authors evaluated their framework on four distinct datasets, each featuring different cell types and imaging characteristics:

  1. Phase-contrast images of cervical cancer HeLa cells: This dataset is used for detailed comparisons and visualizations.
  2. Synthetic cells: A dataset of artificially generated cells.
  3. Drosophila melanogaster Kc167 cells: Cells from fruit flies.
  4. Bright-field images of fission yeast cells: Yeast cells imaged using bright-field microscopy.

The paper notes that for the fission yeast dataset, they did not have access to the complete data and could only obtain 9 images. They also mention using ILASTIK with minimal training data (two images) to generate pixel probability maps for each dataset.

Key Insights and Ideas

  • The Power of Minimal Annotations: The most significant insight is that a semisupervised approach, combined with a clever unsupervised loss, can achieve high performance with extremely sparse annotations (just dots). This drastically lowers the barrier to entry for cell detection tasks.
  • Hierarchical Structure for Constraint Enforcement: Utilizing the MSER tree structure is key. It provides a natural way to represent candidate regions and their relationships, which is then exploited by the unsupervised loss to enforce non-overlapping constraints.
  • Differentiable Unsupervised Loss: Developing a differentiable unsupervised loss for the non-overlapping constraint is a crucial technical innovation. This allows the entire framework to be trained end-to-end using gradient-based optimization.
  • Unlabeled Data is Valuable: The experiments demonstrate that incorporating unlabeled data (from both training and testing sets) significantly improves performance, especially when labeled training data is scarce.

Key Results

  • Improved Accuracy with Less Data: The proposed semisupervised method consistently outperforms the baseline method by Arteta et al. [1] (which also uses dot annotations) across all four datasets, achieving better F-scores.
  • Effectiveness of Unsupervised Loss: The addition of the unsupervised path consistency loss demonstrably improves detection scores compared to using only supervised loss.
  • ILASTIK Integration Boosts Performance: Using pixel probability maps from ILASTIK (especially for low-contrast images) as input to the MSER detector leads to better results than using MSER directly on intensity images.
  • Competitive Performance: For the phase-contrast HeLa cells, the method achieves detection scores comparable to state-of-the-art methods, while also yielding a better DICE score.
  • Training Efficiency: The paper notes that training is relatively fast, with supervised training taking approximately 48.76 seconds and semisupervised training taking 96.59 seconds for a specific dataset on a 3.5 GHz Intel CPU.

Limitations and Future Work

  • Limitations:
    • The paper acknowledges that for some datasets, they did not have access to complete ground truth segmentation masks, limiting the evaluation of certain metrics (like DICE coefficient) to only one dataset.
  • The performance of the MSER detector on intensity images is limited in low-contrast scenarios, necessitating the use of probability maps.
  • Future Work:
    • The authors suggest extending the framework to handle time-lapse microscopy images, where the goal would be to learn classifiers that can predict cell lineages.
  • Further exploration of incorporating unlabeled data from testing sets to improve performance, especially in scenarios with very limited training data.

Summary updated

Abstract

This paper discusses an algorithm to build a semisupervised learning framework for detecting cells. The cell candidates are represented as extremal regions drawn from a hierarchical image representation. Training a classifier for cell detection using supervised approaches relies on a large amount of training data, which requires a lot of effort and time. We propose a semisupervised approach to reduce this burden. The set of extremal regions is generated using a maximally stable extremal region (MSER) detector. A subset of nonoverlapping regions with high similarity to the cells of interest is selected. Using the tree built from the MSER detector, we develop a novel differentiable unsupervised loss term that enforces the nonoverlapping constraint with the learned function. Our algorithm requires very few examples of cells with simple dot annotations for training. The supervised and unsupervised losses are embedded in a Bayesian framework for probabilistic learning.
BibTeX
@article{ramesh2017cell,
  author = {Nisha Ramesh and Ting Liu and Tolga Taşdizen},
  title = {{Cell Detection Using Extremal Regions in a Semisupervised Learning Framework}},
  journal = {Journal of Healthcare Engineering},
  year = {2017},
  doi = {10.1155/2017/4080874},
}
Figure