Maximum Matching Accuracy: An Instance Segmentation Evaluation Metric Utilizing Globally Optimal Matching

Area of Science:

• * Computational biology
• * Image analysis
• * Machine learning

Background:

• * Accurate instance segmentation is crucial for biological image analysis.
• * Current metrics like Intersection-over-Union (IoU) have mathematical limitations, leading to unreliable model evaluations.
• * These limitations include discontinuous scoring, sensitivity to object size, and suboptimal object matching, especially with common cell imaging failures like merged or split cells.

Purpose of the Study:

• * To develop a more robust and reliable metric for evaluating instance segmentation models in biological imaging.
• * To address the inherent mathematical weaknesses of existing metrics.
• * To provide a principled foundation for fair and consistent benchmarking of cell segmentation algorithms.

Main Methods:

• * Proposed Maximum Matching Accuracy (MMA), a threshold-free, continuous metric.
• * MMA establishes a globally optimal one-to-one matching between predicted and ground truth objects.
• * Utilizes per-pixel normalization for aggregating total overlap, avoiding per-object normalization issues.

Main Results:

• * MMA demonstrated more stable, sensitive, and interpretable scores compared to AP@50, PQ, SEG, and AJI.
• * Evaluation across synthetic failures, corruption tests, and model ranking comparisons confirmed MMA's superiority.
• * MMA provides more intuitive and reliable model rankings, especially under common cell segmentation failure modes.

Conclusions:

• * Maximum Matching Accuracy (MMA) offers a significant improvement over existing metrics for instance segmentation in biological imaging.
• * The proposed metric provides a more principled and reliable foundation for benchmarking cell segmentation models.
• * MMA's threshold-free and per-pixel normalization approach enhances evaluation accuracy and interpretability.