Evolutionary design of explainable algorithms for biomedical image segmentation.

Area of Science:

• Computational Biology
• Medical Image Analysis
• Computer Vision

Background:

• Biomedical research faces challenges with the vast number and complexity of images requiring annotation and interpretation.
• Deep Learning (DL) excels in computer vision but demands large datasets and produces "black box" models lacking interpretability.

Purpose of the Study:

• To introduce Kartezio, a novel computational strategy for generating transparent and interpretable image processing pipelines.
• To evaluate Kartezio's performance against state-of-the-art Deep Learning methods in image segmentation tasks.

Main Methods:

• Development of Kartezio, a modular Cartesian Genetic Programming-based approach to evolve image processing pipelines.
• Utilizing Few-Shot Learning to drastically reduce the need for large annotated training datasets.
• Application of Kartezio to semantic and instance segmentation problems across diverse biomedical image types.

Main Results:

• Kartezio pipelines demonstrate comparable precision to DL approaches in instance segmentation.
• Kartezio requires significantly smaller training datasets, highlighting its Few-Shot Learning capabilities.
• Successful deployment across multiplexed tissue histopathology and high-resolution microscopy images.

Conclusions:

• Kartezio provides a flexible, fast, and interpretable alternative for biomedical image processing.
• Its explicable evolutionary design offers a potential advancement in the field.
• Kartezio complements and can be auxiliary to existing Deep Learning methodologies.