A Compound Loss Function With Shape Aware Weight Map for Microscopy Cell Segmentation.

Area of Science:

• Computational biology
• Image analysis
• Deep learning

Background:

• Microscopy cell segmentation is vital for biological image analysis.
• Deep learning methods show promise but face challenges like class imbalance and annotation issues.
• Existing loss functions primarily address inter-class imbalance, neglecting other critical factors.

Purpose of the Study:

• To propose a new compound loss function for microscopy cell segmentation.
• To address inter-class and intra-class imbalance, cell minutiae segmentation, and missing annotations.
• To enhance the performance of deep learning-based cell segmentation models.

Main Methods:

• Developed a compound loss function incorporating a shape-aware weight map.
• Integrated Youden's J index for inter-class imbalance and focal cross-entropy for intra-class imbalance.
• Utilized a shape-aware weight map to handle missing annotations and segment cell minutiae.

Main Results:

• The proposed compound loss function with a shape-aware weight map demonstrated superior performance.
• Evaluations on ten 2D+time datasets from the Cell Tracking Challenge confirmed its effectiveness.
• The novel loss function improved the performance of existing deep learning cell segmentation methods.

Conclusions:

• The proposed compound loss function effectively tackles multiple challenges in cell segmentation.
• The shape-aware weight map is crucial for handling annotation gaps and segmenting fine cellular details.
• This approach offers a significant advancement for deep learning-based biological image analysis.