Abstract
An unresolved issue in contemporary biomedicine is the overwhelming number and diversity of complex images that require annotation, analysis and interpretation. Recent advances in Deep Learning have revolutionized the field of computer vision, creating algorithms that compete with human experts in image segmentation tasks. Crucially however, these frameworks require large human-annotated datasets for training and the resulting models are difficult to interpret. In this study, we introduce Kartezio, a modular Cartesian Genetic Programming based computational strategy that generates transparent and easily interpretable image processing pipelines by iteratively assembling and parameterizing computer vision functions. The pipelines thus generated exhibit comparable precision to state-of-the-art Deep Learning approaches on instance segmentation tasks, while requiring drastically smaller training datasets, a feature which confers tremendous flexibility, speed, and functionality to this approach. We also deployed Kartezio to solve semantic and instance segmentation problems in four real-world Use Cases, and showcase its utility in imaging contexts ranging from high-resolution microscopy to clinical pathology. By successfully implementing Kartezio on a portfolio of images ranging from subcellular structures to tumoral tissue, we demonstrated the flexibility, robustness and practical utility of this fully explicable evolutionary designer for semantic and instance segmentation.
Introduction
The rapid proliferation of high-resolution biomedical imaging—spanning from nanoscale particles to centimeter-scale tissue samples—has created a critical bottleneck in manual data analysis. While Deep Learning (DL) has achieved human-level performance in segmentation, it presents two major hurdles for the biomedical community: extreme "data hunger," requiring thousands of annotated images, and the "black box" problem, where model decisions lack transparency. Kartezio addresses these challenges by evolving transparent, modular pipelines that thrive on "few-shot learning."
Methodology: Cartesian Genetic Programming
Kartezio leverages Cartesian Genetic Programming (CGP) to automatically design image processing workflows. Unlike neural networks that optimize millions of weights, Kartezio evolves a "genotype"—a sequence of integers—that decodes into a "phenotype" represented as a directed acyclic graph of functions.
The system utilizes a curated library of 42 standard computer vision functions (e.g., Gaussian blurs, morphological operators). A key innovation is the use of non-evolvable nodes, such as fixed preprocessing and endpoint nodes (like the Watershed Transform). These act as intentional bottlenecks that guide the evolutionary process toward biologically meaningful outputs. The final result is a human-readable Python script, ensuring full transparency of every filtering and transformation step.
Results and Performance
Kartezio was validated across diverse benchmarks, including the Cellpose dataset and melanoma whole-slide images. The framework demonstrated remarkable efficiency and accuracy:
- Few-Shot Learning: Kartezio matched the performance of Mask R-CNN using just 6 training images and StarDist with only 3 images.
- Competitive Accuracy: On the Cellpose benchmark, Kartezio achieved an Average Precision (AP50) of 0.89, nearly matching the 0.91 achieved by specialized DL models.
- Computational Efficiency: Evolved pipelines can process a 512x512 image in approximately 0.01 seconds on a standard laptop CPU, significantly reducing the hardware requirements compared to GPU-intensive DL models.
Conclusion
By successfully implementing Kartezio on a portfolio of images ranging from subcellular structures to tumoral tissue, this study demonstrates the flexibility and robustness of an evolutionary approach to image analysis. Kartezio offers a "white-box" alternative that is both resource-frugal and highly interpretable. Future developments aim to integrate non-image data (clinical metadata) via Mixed-Type CGP and implement active learning modes to further enhance human-AI collaboration in biomedical research.
Summary updated