Co-localization of fluorescent signals using deep learning with Manders overlapping coefficient

Area of Science:

• Neuroscience
• Cell Biology
• Bioimaging

Background:

• Object-based co-localization of fluorescent signals is crucial for assessing biological interactions using spatial information.
• Accurate object identification is essential for separating fluorescent signals from background noise in microscopy.
• Current methods using convolutional neural networks (CNNs) for detecting co-localized cells face challenges due to the lack of segmented annotation datasets and the underutilization of co-localization coefficients in training.

Purpose of the Study:

• To address the limitations in co-localized cell detection by integrating a co-localization quantification coefficient into a CNN.
• To evaluate the effectiveness of the Manders overlapping coefficient (MOC) as a single-layer branch within a CNN for improving object detection.
• To enhance the accuracy of detecting and localizing fluorescent signals in biological images.

Main Methods:

• A modified Fully Convolutional One-Stage (FCOS) network with a Resnet101 backbone was employed.
• The Manders overlapping coefficient (MOC) was incorporated as a novel single-layer branch within the CNN architecture.
• The model was trained and evaluated using curated fluorescence images of neurons from various rat brain regions (hippocampus, piriform cortex, somatosensory cortex, amygdala).

Main Results:

• The modified FCOS model incorporating MOC demonstrated improved accuracy in detecting fluorescence signals compared to the original FCOS model.
• The enhanced model achieved a 1.1% increase in mean average precision (mAP) for fluorescence signal detection.
• The study successfully validated the utility of MOC in assisting bounding box prediction for co-localized objects.

Conclusions:

• Integrating the Manders overlapping coefficient (MOC) into CNN architectures significantly enhances the accuracy of object-based co-localization detection.
• The developed modified FCOS model offers a more precise tool for analyzing spatial interactions of fluorescently labeled biological entities.
• This approach provides a valuable advancement for bioimaging analysis, particularly in neuroscience research, by improving the detection of co-localized cells.