Automating cell counting in fluorescent microscopy through deep learning with c-ResUnet.

Area of Science:

• Biomedical Imaging
• Computational Biology
• Machine Learning

Background:

• Manual cell counting in fluorescent microscopy is labor-intensive and prone to errors.
• Accurate cell quantification is crucial for assessing experimental outcomes in biological research.

Purpose of the Study:

• To develop and evaluate a Deep Learning-based automated cell counting method.
• To improve the accuracy and efficiency of cell detection and quantification in microscopy images.

Main Methods:

• A Unet-like architecture, named cell ResUnet (c-ResUnet), was employed for binary segmentation of cells.
• Performance was compared against three similar architectures, with ablation studies on artifact oversampling and weight maps.
• The model was trained and validated on an annotated dataset.

Main Results:

• The c-ResUnet achieved superior performance in both detection (IoU score = 0.81) and counting (Mean Absolute Error = 3.09) metrics.
• Weight maps enhanced performance, particularly for crowded or complex cellular environments.
• Qualitative assessment indicated human-level performance, with errors falling within acceptable interpretation ranges.

Conclusions:

• The c-ResUnet offers a robust and accurate automated solution for cell counting in fluorescent microscopy.
• The developed model and dataset can advance research in automated image analysis for biological studies.
• This approach mitigates human error and subjectivity in cell quantification tasks.