Instance segmentation of cells and nuclei from multi-organ cross-protocol microscopic images

Area of Science:

• Cell Biology
• Computer Vision
• Biomedical Imaging

Background:

• Light microscopy is crucial for cell biology but faces challenges in segmenting cells and nuclei due to variations in morphology, noise, and overlapping cells.
• Accurate segmentation is essential for reliable cell analysis, necessitating advanced computer-aided methods.
• Deep learning offers powerful solutions for image processing tasks in biological research.

Purpose of the Study:

• To develop and evaluate a novel deep learning model for precise instance segmentation of cells and nuclei in microscopy images.
• To leverage cutting-edge deep learning techniques for improved accuracy and efficiency in cell image analysis.
• To provide a robust computational tool for assisting cell biology researchers.

Main Methods:

• A fine-tuned You Only Look at Once version 9 extended (YOLOv9-E) model was used for bounding box prompt generation.
• A pre-trained Segment Anything Model (SAM) was employed for initial segmentation mask generation.
• Segmentation masks were refined using non-max suppression and image processing techniques.

Main Results:

• The proposed method demonstrated strong performance on the Expert Visual Cell Annotation (EVICAN) dataset.
• Average mAP50 scores for cell segmentation reached up to 96.25% on easy test sets.
• Nucleus segmentation achieved scores up to 68.04% on easy test sets, with performance varying across difficulty levels.

Conclusions:

• The developed instance segmentation method shows favorable performance compared to existing approaches.
• The model shows potential as an assistive tool for cell culture experts.
• This approach facilitates prompt and reliable analysis of cellular structures in microscopy images.