KaryoVisio

This repository contains deep learning models developed for chromosome detection and classification as part of the KaryoVisio project.

Models

YOLO26m — Object Detection

Original Dataset

A YOLO26m model trained on the original chromosome image dataset, without specimen preprocessing.

• Model weights: best.pt and best.onnx.

Preprocessed Dataset

A YOLO26m model trained using the dataset after specimen preprocessing.

• Model checkpoints: best.pt and last.pt.
• Training configuration, metrics, and evaluation visualizations are included.

Cross-Validation — yolo26m-cv

A YOLO26m model evaluated using 5-fold cross-validation on the chromosome image dataset.

Each fold contains model checkpoints, training configuration, evaluation metrics, plots, confusion matrices, and training/validation visualizations. Separate test evaluation artifacts are also available for each fold.

Cross-validation results

Mean performance across the five folds, reported as mean ± standard deviation:

Metric	Mean ± Std. Dev.
Precision	98.928% ± 0.022%
Recall	98.869% ± 0.071%
F1-score	98.898% ± 0.029%
mAP@50	99.387% ± 0.008%
mAP@50–95	87.789% ± 0.384%

The cv_summary.csv file contains the individual fold results and aggregated statistics.

ResNet50 — Chromosome Classification

A ResNet50 model trained to classify chromosomes into their respective classes, including groups labeled A, B, C, D, and others, as well as the X and Y chromosomes.

The model was evaluated using 5-fold cross-validation. Each fold contains its own model weights, configuration, training history, and test predictions.

Cross-validation results

Mean performance across the five folds, reported as mean ± standard deviation:

Metric	Mean ± Std. Dev.
Accuracy	98.540% ± 0.027%
Macro Precision	98.489% ± 0.036%
Macro Recall	98.424% ± 0.067%
Macro F1-score	98.455% ± 0.051%
Macro AUC	99.966% ± 0.002%

The cv_summary.csv file contains the aggregated cross-validation results. Each fold also includes checkpoints, training history, configuration, test labels, and predicted probabilities.

Dataset

The models were trained using the chromosome image dataset described in the following publication:

A large-scale dataset of chromosome images for machine learning applications

Read the dataset publication — Nature Scientific Data

Repository Structure

models/
├── resnet50-paper-cv/
│   ├── cv_summary.csv
│   ├── fold0/
│   ├── fold1/
│   ├── fold2/
│   ├── fold3/
│   └── fold4/
├── yolo26m-cv/
│   ├── cv_summary.csv
│   ├── fold0/
│   ├── fold0-test/
│   ├── fold1/
│   ├── fold1-test/
│   ├── fold2/
│   ├── fold2-test/
│   ├── fold3/
│   ├── fold3-test/
│   ├── fold4/
│   └── fold4-test/
├── yolo26m-original/
│   └── model/
│       ├── best.pt
│       └── best.onnx
└── yolo26m-preprocessed/
    ├── model/
    │   ├── best.pt
    │   └── last.pt
    └── training/

The cross-validation directories preserve fold-specific model checkpoints and experimental artifacts to support reproducibility and further analysis.

Purpose

This repository provides versioned storage for trained models and their associated experimental artifacts, supporting reproducibility and the evaluation of chromosome detection and classification approaches.