Kartezio: Evolutionary design of explainable algorithms for biomedical image segmentation

Kartezio is a modular Cartesian Genetic Programming (CGP) framework that enables the automated design of fully interpretable image-processing pipelines, without the need for GPUs or extensive training datasets.
Built on top of OpenCV, Kartezio empowers researchers, engineers, and practitioners to discover novel computer vision (CV) solutions using only a handful of annotated samples and a single CPU core.

Originally developed for biomedical image segmentation, Kartezio has been successfully showcased in Nature Communications. Although it shines in medical and life science applications, Kartezio’s underlying principles are domain-agnostic.
Whether you’re working with industrial quality control, satellite imagery, embedded vision, or robotics, Kartezio helps you craft custom CV pipelines that are transparent, fast, frugal and efficient.

Why you should try Kartezio?

:nut_and_bolt: Modular and Customizable
Kartezio is built from interchangeable building blocks, called Components, that you can mix, match, or replace. Adapt the pipeline to your project’s unique requirements.

:pencil2: Few-Shot Learning
Forget the need for massive, annotated datasets. Kartezio can evolve solutions from just a few annotated examples, saving both time and computational resources.

:white_check_mark: Transparent and Certifiable
Every pipeline produced is fully transparent. Inspect the exact operations used, understand their sequence, and trust the decisions made by your model.

:earth_africa: Frugal and Local
Run everything on a single CPU, without GPUs or massive compute clusters. This makes Kartezio ideal for edge devices, embedded systems, or scenarios with limited computational resources.

:microscope: Broad Applicability
While proven in biomedical image segmentation, Kartezio’s methods readily extend to other fields—like industrial machine vision, space imaging, drone footage analysis, or any custom image-based problem.

:books: Traditional Computer Vision
Kartezio offers a straightforward, interpretable way to learn and play with traditional CV filters. This makes it an excellent resource for teaching and learning about Image Processing fundamentals.

Getting Started

1. Installation:

   pip install kartezio
   

2. 🚀 Quick Start

Here's a complete example that evolves a cell segmentation pipeline:

from kartezio.core.endpoints import EndpointThreshold
from kartezio.core.fitness import IoU  
from kartezio.evolution.base import KartezioTrainer
from kartezio.primitives.matrix import default_matrix_lib
from kartezio.utils.dataset import one_cell_dataset

# 1. Set up components
n_inputs = 1
libraries = default_matrix_lib()    # Library of image operations
endpoint = EndpointThreshold(128)   # Binary output via thresholding
fitness = IoU()                     # Intersection over Union metric

# 2. Create and configure the evolutionary trainer  
model = KartezioTrainer(
    n_inputs=n_inputs,
    n_nodes=n_inputs * 10,          # 10 processing nodes
    libraries=libraries,
    endpoint=endpoint,
    fitness=fitness,
)
model.set_mutation_rates(node_rate=0.05, out_rate=0.1)

# 3. Load your data (or use the included example dataset)
train_x, train_y = one_cell_dataset()  # Example: cell images + masks

# 4. Evolve the algorithm (100 generations)  
elite, history = model.fit(100, train_x, train_y)

# 5. Evaluate performance
score = model.evaluate(train_x, train_y)
print(f"Final IoU Score: {score:.3f}")

# 6. Export as standalone Python code
model.print_python_class("CellSegmenter")

That's it! Kartezio has evolved a complete image processing pipeline tailored to your data.

📚 Core Concepts

Architecture Overview

Kartezio uses a component-based architecture with four main types:

1. Primitives: Basic image operations (filters, morphology, arithmetic)
2. Endpoints: Output processing (thresholding, watershed, etc.)
3. Fitness Functions: Performance metrics (IoU, AP, custom metrics)
4. Libraries: Collections of primitives organized by data type

Component Registration System

Components are registered using decorators:

from kartezio.core.components import Primitive, register
from kartezio.types import ArrayData, DataList, DataType, Matrix1, Parameters

@register(Primitive)
class CustomFilter(Primitive):
    def __init__(self):
        super().__init__(inputs=Matrix1, output=DataType.MATRIX, n_parameters=1) #  -> the function will take 1 image and 1 parameter
    
    def call(self, x: DataList, args: Parameters) -> ArrayData:
        kernel_size = args[0]
        return cv2.GaussianBlur(x[0], (kernel_size, kernel_size), 0)

Evolution Process

1. Initialization: Random population of image processing graphs
2. Evaluation: Each individual processes training images
3. Selection: The best performer survive based on fitness scores
4. Mutation: Modify operations, connections, and parameters
5. Iteration: Repeat until convergence or generation limit

🛠️ Advanced Usage

Custom Fitness Functions

Define domain-specific evaluation metrics:

from kartezio.core.components import Fitness, register
from kartezio.types import DataBatch, ScoreBatch
import numpy as np

@register(Fitness)  
class CustomMetric(Fitness):
   def __init__(self, reduction="mean"):
        super().__init__(reduction)

   def evaluate(self, y_true: DataBatch, y_pred: DataBatch) -> ScoreBatch:
      n_images = len(y_true)
      mse_values = np.zeros(n_images, np.float32)

      for n in range(n_images):
         _y_true = y_true[n][0]
         _y_pred = y_pred[n][0]
         # Compute Mean Squared Error
         mse_values[n] = np.mean((_y_true - _y_pred) ** 2)
      return mse_values

Adding New Primitives

Extend Kartezio with domain-specific operations:

@register(Primitive)
class AdvancedMorphology(Primitive):
    def __init__(self):
        super().__init__(inputs=Matrix1, output=DataType.MATRIX, n_parameters=1)
    
    def call(self, x: DataList, args: Parameters) -> ArrayData:
        operation_type = args[0]  # 0: opening, 1: closing
        kernel_size = args[1]
        
        kernel = cv2.getStructuringElement(cv2.MORPH_ELLIPSE, 
                                         (kernel_size, kernel_size))
        
        if operation_type < 128:
            return cv2.morphologyEx(x[0], cv2.MORPH_OPEN, kernel)
        else:
            return cv2.morphologyEx(x[0], cv2.MORPH_CLOSE, kernel)

your_lib = Library(DataType.MATRIX)  # define the return type of the lib
your_lib.add_primitive(AdvancedMorphology())

References and Citation

If you use Kartezio in your research, please consider citing:

@article{cortacero2023evolutionary,
  title={Evolutionary design of explainable algorithms for biomedical image segmentation},
  author={Cortacero, K{\'e}vin and McKenzie, Brienne and M{\"u}ller, Sabina and Khazen, Roxana and Lafouresse, Fanny and Corsaut, Ga{\"e}lle and Van Acker, Nathalie and Frenois, Fran{\c{c}}ois-Xavier and Lamant, Laurence and Meyer, Nicolas and others},
  journal={Nature Communications},
  volume={14},
  number={1},
  pages={7112},
  year={2023},
  publisher={Nature Publishing Group UK London}
}

If you are using the multimodal version of Kartezio, please also cite:

@inproceedings{de2024multimodal,
  title={Multimodal adaptive graph evolution},
  author={De La Torre, Camilo and Cortacero, K{\'e}vin and Cussat-Blanc, Sylvain and Wilson, Dennis},
  booktitle={Proceedings of the Genetic and Evolutionary Computation Conference Companion},
  pages={499--502},
  year={2024}
}

Licensing

The Software is freely available for Non-Commercial and Academic purposes only, under the terms and conditions set out in the License file, and You may not use the Software except in compliance with the License. The Software distributed under the License is distributed on an "as is" basis, without warranties or conditions of any kind, either express or implied. See the License file for the specific language governing permissions and limitations under the License.