Automated Individualization of Size-Varying and Touching Neurons in Macaque Cerebral Microscopic Images

Zhenzhen You, Yaël Balbastre, Clément Bouvier, Anne‐Sophie Hérard, Pauline Gipchtein, Philippe Hantraye, Caroline Jan, Nicolas Souedet, Thierry DelzescauxFrontiers in Neuroanatomy 2019 · DOI

TL;DR

This paper introduces a novel automated method for accurately identifying and separating individual neurons in high-resolution microscopic images of macaque brains. The approach tackles the challenge of neurons varying in size and often touching or overlapping, achieving results comparable to manual expert analysis and outperforming traditional algorithms like Watershed.

What Problem Does This Paper Address?

Analyzing microscopic images of brain tissue is crucial for understanding brain function, development, and disease. However, these images are massive and incredibly complex, making manual cell-by-cell analysis time-consuming and prone to error. A key challenge is accurately identifying individual neurons, especially when they vary significantly in size and are physically touching or appear to overlap due to the 2D projection of 3D tissue. Existing automated methods often struggle with this variability and density, leading to under- or over-segmentation. This paper aims to provide a robust automated solution for this problem.

Methods

The authors propose a multi-step processing pipeline:

  1. Tissue Segmentation: A Random Forest classifier is used to distinguish neuron tissue from background. This step leverages features from various color spaces (RGB, HSV, XYZ, Lab*) and texture descriptors (LBP).
  2. Centroid and Boundary Enhancement: A novel min-max filter is introduced. This filter, applied after Gaussian filtering, enhances neuron centroids and boundaries, making them more detectable. The key innovation here is an adaptive approach to the Gaussian filter's parameter (σ), which is determined based on the estimated neuron size.
  3. Adaptive Multiscale Procedure: To handle size-varying neurons, an adaptive multiscale strategy is employed. This involves estimating an optimal σ for each neuron, allowing the algorithm to adjust its sensitivity to different neuron sizes.
  4. Region Growing with Contour-Based Model: A region growing process, guided by a contour-based model, is used to achieve the final individualization of touching neurons. This model expands contours based on curvature and intensity information, effectively separating adjacent cells.

Datasets

The study utilized several datasets:

  • Segmentation Dataset: 100 representative images (512x512 pixels) from sections Nos. 81 and 101, manually segmented into three classes (neuronal staining, non-stained tissue, background) by an expert.
  • Individualization Dataset: 58 large images (5000x5000 pixels) from section No. 91, with each neuron's centroid manually identified by an expert. This dataset covers diverse neuron densities and anatomical regions.
  • Stereology Dataset: Eight large images from six anatomical regions, used for comparing the automated method with expert-based stereological counting.

The authors also generated "ground truth" databases from these datasets, involving multiple experts.

Key Insights and Ideas

  • Adaptive σ is Crucial: The paper highlights that a fixed Gaussian filter parameter (σ) is insufficient for accurate neuron detection due to the wide range of neuron sizes. An adaptive, neuron-specific σ estimation is key to robust performance.
  • Min-Max Filter for Enhancement: The proposed min-max filter is effective in enhancing both neuron centroids and boundaries, providing crucial information for subsequent segmentation steps.
  • Contour-Based Region Growing for Separation: Using a contour-based model within region growing allows for more precise separation of touching neurons compared to simpler region growing or watershed methods.
  • DG as a Unique Challenge: The dentate gyrus (DG) region, with its extremely dense aggregation of small neurons, presents a particularly difficult case. The study shows that while it requires specific parameter tuning, the method can still achieve good results.
  • Stereology as a Gold Standard Benchmark: The paper's validation against stereology, a well-established manual counting technique, provides strong evidence for the accuracy and reliability of their automated method.

Key Results

  • High Accuracy: The method achieved an F-score of 0.816 ± 0.062 in the hippocampus and up to 0.905 ± 0.001 in the subiculum. For neuron counting, it showed a high correlation with expert stereology (0.983 and 0.975).
  • Outperforms Watershed: Qualitative and quantitative analyses demonstrated that the proposed method outperforms the baseline Watershed algorithm, especially in complex cases with touching and size-varying neurons.
  • Neuron Size Range: The detected neuron diameters ranged from 2 to 28.6 µm, aligning with literature values.
  • Efficiency: The method is designed for large images and can be parallelized, with an average execution time of 16.1 minutes per 5000x5000 pixel image on a multi-core workstation.

Limitations and Future Work

  • Staining and Interindividual Variability: The authors plan to evaluate the impact of different staining protocols and variations between individual animals on their method's performance.
  • Stereology Bias: While the method shows high correlation with stereology, the paper acknowledges that the exclusion of "dead zones" in their stereology comparison introduces a slight bias, potentially leading to overdetection. Future work may involve precisely estimating this bias.
  • 2D vs. 3D: The current method operates on 2D images. The authors suggest that extending it to 3D (e.g., using Z-stacks) could improve accuracy but would significantly increase data volume and computational demands.
  • Deep Learning Integration: While the current segmentation uses Random Forests, preliminary tests with U-Net showed comparable results, suggesting future work could explore deep learning for segmentation and potentially individualization.

Summary updated

Abstract

In biomedical research, cell analysis is important to assess physiological and pathophysiological information. Virtual microscopy offers the unique possibility to study the compositions of tissues at a cellular scale. However, images acquired at such high spatial resolution are massive, contain complex information, and are therefore difficult to analyze automatically. In this article, we address the problem of individualization of size-varying and touching neurons in optical microscopy two-dimensional (2-D) images. Our approach is based on a series of processing steps that incorporate increasingly more information. (1) After a step of segmentation of neuron class using a Random Forest classifier, a novel min-max filter is used to enhance neurons' centroids and boundaries, enabling the use of region growing process based on a contour-based model to drive it to neuron boundary and achieve individualization of touching neurons. (2) Taking into account size-varying neurons, an adaptive multiscale procedure aiming at individualizing touching neurons is proposed. This protocol was evaluated in 17 major anatomical regions from three NeuN-stained macaque brain sections presenting diverse and comprehensive neuron densities. Qualitative and quantitative analyses demonstrate that the proposed method provides satisfactory results in most regions (e.g., caudate, cortex, subiculum, and putamen) and outperforms a baseline Watershed algorithm. Neuron counts obtained with our method show high correlation with an adapted stereology technique performed by two experts (respectively, 0.983 and 0.975 for the two experts). Neuron diameters obtained with our method ranged between 2 and 28.6 μm, matching values reported in the literature. Further works will aim to evaluate the impact of staining and interindividual variability on our protocol.
BibTeX
@inproceedings{you2019automated,
  author = {Zhenzhen You and Yaël Balbastre and Clément Bouvier and Anne‐Sophie Hérard and Pauline Gipchtein and Philippe Hantraye and Caroline Jan and Nicolas Souedet and Thierry Delzescaux},
  title = {{Automated Individualization of Size-Varying and Touching Neurons in Macaque Cerebral Microscopic Images}},
  booktitle = {Frontiers in Neuroanatomy},
  year = {2019},
  doi = {10.3389/fnana.2019.00098},
}
Figure