[Paper Review] Enhanced Center Coding for Cell Detection with Convolutional Neural Networks

Haoyi Liang, Aijaz Ahmad Naik+3 more|arXiv (Cornell University)|Apr 18, 2019

Cell Image Analysis Techniques34 references4 citations

TL;DR

This paper proposes repel coding, a novel center coding scheme for cell detection using convolutional neural networks that improves training stability and detection accuracy by suppressing responses between closely spaced cell centers. It achieves up to 12% higher counting accuracy and 14% higher detection accuracy compared to Gaussian and proximity coding on four cell types and two network architectures.

ABSTRACT

Cell imaging and analysis are fundamental to biomedical research because cells are the basic functional units of life. Among different cell-related analysis, cell counting and detection are widely used. In this paper, we focus on one common step of learning-based cell counting approaches: coding the raw dot labels into more suitable maps for learning. Two criteria of coding raw dot labels are discussed, and a new coding scheme is proposed in this paper. The two criteria measure how easy it is to train the model with a coding scheme, and how robust the recovered raw dot labels are when predicting. The most compelling advantage of the proposed coding scheme is the ability to distinguish neighboring cells in crowded regions. Cell counting and detection experiments are conducted for five coding schemes on four types of cells and two network architectures. The proposed coding scheme improves the counting accuracy versus the widely-used Gaussian and rectangle kernels up to 12%, and also improves the detection accuracy versus the common proximity coding up to 14%.

Motivation & Objective

• Address the challenge of suboptimal performance in learning-based cell detection due to poor coding of raw dot labels.
• Improve training stability and prediction robustness by designing a coding scheme that balances entropy and reversibility.
• Enable better distinction of neighboring cells in crowded regions, a persistent issue in cell counting and detection.
• Develop a coding method that outperforms widely used schemes like Gaussian and rectangle kernels across diverse cell types and imaging conditions.

Proposed method

• Propose two evaluation criteria for center coding: entropy (measuring training suitability) and reversibility (measuring robustness in label recovery).
• Design repel coding to minimize response in the middle of two cell centers while boosting responses near actual centers, using a repulsion mechanism.
• Implement repel coding as a learnable or fixed transformation of raw dot labels into a density map that enhances feature learning.
• Train two CNN architectures—FCNN-A and a residual-block-enhanced version—on four microscopy datasets with five coding schemes for comparison.
• Use F1 score and mean absolute error (MAE) to evaluate detection and counting performance across all configurations.
• Apply a matching criterion based on distance less than the average cell radius to assess detection accuracy, ensuring one-to-one pairing of predicted and ground truth centers.

Experimental results

Research questions

• RQ1 How does the proposed repel coding scheme compare to existing coding methods (Gaussian, rectangle, proximity, raw dot labels) in terms of training stability and prediction accuracy?
• RQ2 To what extent does repel coding improve the detection of closely spaced or overlapping cells in crowded regions?
• RQ3 Does the balance between entropy and reversibility in repel coding lead to better generalization across diverse cell types and imaging modalities?
• RQ4 Can repel coding reduce training divergence issues observed with raw dot labels, especially in datasets with small image sizes and sparse labels?
• RQ5 How does the inclusion of residual convolution blocks affect the performance of different coding schemes?

Key findings

• Repel coding achieves the highest F1 score in 38 out of 40 sub-experiments across four datasets and two network architectures.
• On the DG dataset with FCNN-A, repel coding improves F1 score by 6.96% over proximity coding and 7.95% over Gaussian kernel coding.
• On the HBM dataset with residual blocks, repel coding achieves an F1 score of 0.9011, outperforming proximity coding (0.8933) by 0.78% and Gaussian kernel (0.8628) by 3.83%.
• Training with raw dot labels frequently diverges when batch size is below eight on the Adip and Vgg datasets, while repel coding avoids this issue due to better label distribution.
• In partially occluded regions (Vgg dataset), repel coding produces more stable predictions than raw dot labels, which tend to generate duplicated or unstable center detections.
• The repel coding scheme demonstrates superior reversibility, enabling more accurate and robust recovery of original cell center locations from network outputs.