Feature-Based Representation Improves Color Decomposition and Nuclear Detection Using a Convolutional Neural Network.

Area of Science:

• Computational Biology
• Biomedical Imaging
• Histopathology

Background:

• Accurate nuclei detection is crucial for phenotypic profiling in histology and 3D cell cultures.
• Existing methods face challenges with diverse imaging modalities and cell phenotypes.

Purpose of the Study:

• To develop an improved nuclei detection method applicable to both bright-field histology and confocal microscopy of 3D cell cultures.
• To enhance the accuracy of detecting nuclei with varying phenotypes, including vesicular, necrotic, or poorly stained cells.

Main Methods:

• Utilized a feature-based image representation employing the Laplacian of Gaussian (LoG) filter to accentuate blob-like structures.
• Integrated LoG response with improved color decomposition (CD) for enhanced nuclear detection using convolutional neural networks (CNNs).
• Evaluated multiple input data representations and CNN architectures, including non-negative matrix factorization for CD in bright-field imaging.

Main Results:

• The feature-based representation significantly improved color decomposition and subsequent nuclei detection across different imaging modalities.
• The combined approach advanced the detection of nuclei, particularly those with challenging phenotypes like vesicular or necrotic cells.
• The system demonstrated improved performance, with F-scores reported against manually annotated images.

Conclusions:

• The proposed feature-based representation coupled with improved CD and CNNs offers a robust solution for nuclei detection in diverse biological samples.
• This advancement facilitates more accurate phenotypic profiling by improving the detection of subtle or difficult-to-identify nuclear features.