Gentl

The source code for Gentl (GENeTic aLgorithm for predicting stage from medical scans of patients with cancer) [access preprint].

About

This is a repository that contains information on how to reproduce results corresponding to the bladder cancer case study reported in Paper title.

Abstract

This is a repository that contains information on how to reproduce results corresponding to the bladder cancer case study reported in Paper title.

Data

Description

• As described in our paper, the data used for our analyses comprised a total of 100 CT scans of the bladder, each from a patient with bladder cancer.

• Disease: urothelial carcinoma of the bladder

• Stages: Ta, Tis, T0, T1, T2, T3, T4

• Stage annotation technique: Performed manually by radiologists

For more details, interested readers are directed to the Dataset section of the paper.

Availability

Data will be made available under reasonable request to the corresponding author, Suryadipto Sarkar (more contact details below).

Data preprocessing

Step 1: Segmenting bladder region using ImageJ

Step 2: Segmenting cancer ROI using binary mask

Step 3: Segmenting healthy ROIs using sliding window

Feature extraction using Gray level co-occurrence matrix (GLCM)

The following five GLCM features were extracted from the cancer ROI, as well as healthy ROIs from the same patient:

• Dissimilarity
  • Measures how different neighboring pixel values are.
  • Assumptions:
    • Cancer region – high dissimilarity
    • Healthy region – low dissimilarity
• Correlation
  • Measures the linear dependency between pixel intensities in a given direction.
  • Assumptions:
    • Cancer region – low correlation
    • Healthy region – high correlation
• Energy
  • Measures texture uniformity or repetition.
  • Assumptions:
    • Cancer region – low energy
    • Healthy region – high energy
• Contrast
  • Measures the intensity variation between neighboring pixels in an image.
  • Assumptions:
    • Cancer region – high contrast
    • Healthy region – low contrast
• Homogeneity
  • Measures whether neighboring pixels in an image have similar intensity values.
  • Assumptions:
    • Cancer region – low homogeneity
    • Healthy region – high homogeneity

using $20$ configurations ($4$ angles: ${0, \frac{\pi}{4}, \frac{\pi}{2}, \frac{3\pi}{4} }$; $5$ distances: ${1, 2, 3,4, 5}$ pixels).

Feature binarization

• Performed on the cancer ROI using bimodal Gaussian mixture model (GMM) fitting using the sklearn.mixture package. All feature values that are closer in Euclidean distance to lower mean ($\mu_1$) is assigned a value of $0$, else $1$ if closer to higher mean ($\mu_2$).
• All healthy ROI feature values from the same image sample are assigned a value of $0$ if they are closer in Euclidean distance to the lower mean ($\mu_1$) obtained from the cancer ROI above, else assigned a value of $1$ if closer to $\mu_2$.

Note: Bimodal GMM fitting only done once per image, pertaining to the cancer ROI. Feature binarization of healthy ROIs performed based on mean values obtained from bimodal GMM fitting on the cancer ROI pertaining to the same image sample.

Genetic algorithm

General information about our implementation of the algorithm

• We perform the genetic algorithm on each sample image separately.

An overview of the terms gene, chromosome and population

Algorithmic workflow

Step 1: Population initialization

• The initial population ($P$) comprises binarized GLCM features extracted from the healthy ROIs.

• Reported results include $P={10, 20, 30, 40, 50}$.

Step 2: Parent selection by fitness evaluation

• Fitness metric: Euclidean or Absolute distance to target.

  • In our implementation target is binarized feature list from cancer ROI.

• Parent selection rate: 50% of the population at the end of iteration i is retained as parents for iteration i+1. Therefore, list of selected parents contains top 50% of the chromosomes closest to the target sequence.

Step 3: Crossover (initial offspring generation)

• For crossover between two parents:

  • The first parent ($p_1$) is always chosen from the top 50% of chromosomes (that is, ones having least Absolute distance to the target sequence).
  • The second parent ($p_2$) is chosen from the initial population at each iteration.

• Random portions of parents $p_1$ and $p_2$ constitute the respective offspring—with at least one gene compulsorily selected from each parent { $p_1$ , $p_2$ }.

Step 4: Mutation (final offspring generation)

• Initial offspring $\overline{o_{1,2}}$ generated from parents $p_1$ and $p_2$ in step 3 (crossover) described above, undergoes mutation to give rise to final offspring $o_{1,2}$.

Step 5: Replacement

• In this step, we replace the worst-performing individuals in the current population with new offspring, retaining the better-performing individuals.

• In the script /gentl/_ga_step5_replacement.py:

  • Input parameters:

    • population: The current population.
    • new_generation: The new generation of chromosomes.
    • goal: The target sequence.

  • Returns:

    • best_individuals: The updated population containing the best individuals.

Installation

Create a virtual environment and install the dependencies using requirements.txt

python -m venv gentlvenv
gentlvenv\Scripts\activate # Windows
source gentlvenv/bin/activate # Mac/Linux
pip install -r requirements.txt

Key metrics for cancer staging

Using Genetic Algorithm (GA), we compute three key metrics to quantify feature similarity between non-cancer and cancer regions, which are essential for cancer staging:

• average_generation:
  The average number of iterations required for GA convergence (averaged over 20 runs).

  • Lower values indicate faster convergence and higher similarity between non-cancer and cancer features.

• average_best_distance:
  The average distance between the best solution (most similar to cancer) and the target cancer feature (final generation, averaged over 20 runs).

  • Lower values suggest that some non-cancer regions closely resemble cancer regions.

• average_mean_distance:
  The average distance between the whole population and the target (final generation, averaged over 20 runs).

  • Reflects how well the entire population approximates cancer-like features, indicating diversity and convergence quality.

All processed results are stored in the folder: glcm_bladder_average_gentl_result/

This folder contains five subfolders, corresponding to experiments with different numbers of healthy ROIs.

Robustness tests

We designed tests to evaluate whether GA remains stable and reliable under different settings:

• Random initialization test:
  Run GA under 10 different random seeds, each repeated 20 times, to assess whether convergence speed and time are consistent across random initializations.

• Mutation rate test:
  Test GA performance under various mutation rates (0.001, 0.05, 0.1, 0.15, 0.9) to analyze how mutation influences diversity and convergence.

• Population capacity (Np) test:
  Examine how different offspring population sizes (25, 26, 28, 30, 40, 50, 60) affect GA's ability to avoid local optima and achieve stable solutions.

These tests ensure that GA performs consistently without being overly sensitive to random factors or hyperparameters.

Scalability tests

We analyzed how GA scales when facing larger or more complex feature sets:

• Chromosome Length Test:
  Vary chromosome lengths (10 to 70) to examine how problem size impacts convergence time and iterations.

• Max Generation Limit Test:
  Vary the maximum number of allowed generations (35 to 80) to confirm whether GA typically converges before hitting the limit.

• Population Size Test:
  Vary population sizes (20, 40, 60, 80, 100) to evaluate trade-offs between diversity and runtime efficiency.

These tests validate that GA remains feasible and efficient even as the complexity of the feature space increases.

Citing the work

MLA

Will be made available upon publication.

APA

Will be made available upon publication.

BibTex

Will be made available upon publication.

Contact

✉  suryadipto.sarkar@fau.de
✉  ssarka34@asu.edu
✉  ssarkarmanipal@gmail.com

Impressum

Suryadipto Sarkar ("Surya"), MS

PhD Candidate
Biomedical Network Science Lab
Department of Artificial Intelligence in Biomedical Engineering (AIBE)
Friedrich-Alexander University Erlangen-Nürnberg (FAU)
Werner von Siemens Strasse
91052 Erlangen

MS in CEN from Arizona State University, AZ, USA.
B.Tech in ECE from MIT Manipal, KA, India.