Detection of Nuclei in H&E Stained Sections Using Convolutional Neural Networks

Area of Science:

• Digital pathology
• Computational biology
• Image analysis

Background:

• Accurate nuclei detection is crucial for phenotypic profiling in histology.
• Nuclei exhibit diverse phenotypes that pose challenges for automated detection models.
• Current methods struggle to effectively model the variety of nuclear appearances.

Purpose of the Study:

• To enhance nuclei detection in histology images using convolutional neural networks (CNNs).
• To investigate the utility of feature-based image representations for improving CNN performance.
• To specifically improve the detection of challenging nuclear phenotypes like vesicular and apoptotic nuclei.

Main Methods:

• Utilized convolutional neural networks (CNNs) for nuclear detection.
• Employed a feature-based image representation using the Laplacian of Gaussian (LoG) filter.
• Evaluated different input data representations, including LoG-filtered images.
• Assessed the CNN's efficacy on specific nuclear phenotypes (vesicular, hyperchromatic, apoptotic).

Main Results:

• Feature-based representation, particularly using the LoG filter, significantly advanced nuclei detection.
• CNNs successfully learned distinct phenotypic signatures for nuclear detection.
• The frequency of detecting vesicular and apoptotic nuclei was notably increased.
• Performance was validated against manual annotations, with reported F-Scores for various representations.

Conclusions:

• Laplacian of Gaussian (LoG) feature representation enhances CNN-based nuclei detection in histology.
• CNNs are effective in learning and identifying diverse nuclear phenotypes.
• The developed system improves the detection rate of specific, challenging nuclear morphologies.