A Real-Time Cell Image Segmentation Method Based on Multi-Scale Feature Fusion

Area of Science:

• Biomedical image analysis
• Computational biology
• Artificial intelligence in medicine

Background:

• Cellular growth assessment is crucial for disease diagnosis and therapy development.
• Accurate cell segmentation is vital for quantifying cellular growth indicators.
• Existing methods struggle with multi-scale heterogeneity, unclear boundaries, and efficiency-accuracy trade-offs.

Purpose of the Study:

• To develop an innovative network architecture for efficient and accurate cell segmentation.
• To address challenges in multi-scale heterogeneity and boundary delineation.
• To provide a reliable automated tool for tumor microenvironment research.

Main Methods:

• A preprocessing pipeline using CLAHE and Gaussian blur for image enhancement.
• A bidirectional feature pyramid network (BiFPN) for enhanced multi-scale feature recognition.
• Adaptive kernel convolution (AKConv) for capturing heterogeneous cell distributions and improving boundary segmentation.
• Probability density-guided non-maximum suppression (Soft-NMS) to reduce under-detection.

Main Results:

• Achieved 95.7% mAP50 (box) and 95% mAP50 (mask) on the GSCs dataset.
• Demonstrated an inference speed of 38 frames per second.
• Successfully supported dual-modality output for cell confluence and precise counting.

Conclusions:

• The proposed model offers a significant advancement in automated cell segmentation.
• It provides a reliable and efficient tool for quantitative analysis of cellular growth in tumor microenvironments.
• The method effectively addresses limitations of current segmentation techniques.