KaryoFlow

Rectified-Flow Refinement and Localization-Quality Ranking for Chromosome Detection

KaryoFlow advances generative detection into a competitive approach to chromosome localization. It adapts DiffusionDet through rectified-flow box refinement and improves candidate selection with localization-quality ranking. This repository provides the trained checkpoints and configurations used in the reported experiments.

Code and demonstrations · Checkpoint directory · Configuration and result index

Contents

• Models and results
• Choose a checkpoint
• Download and evaluate
• Data and intended use
• License and attribution

Models and results

Held-out chromosome detection, measured by COCO box mAP on the 0–1 scale. Each dataset is trained and evaluated separately.

Model	Dataset 1 mAP	Dataset 2 mAP
YOLOX-S	0.5748	0.7951
Cascade R-CNN	0.6864	0.8527
DiffusionDet	0.6908	0.8033
RTMDet-L	0.6957	0.8609
KaryoFlow	0.7082	0.8597
KaryoFlow + LQCR	0.7156	0.8659
KaryoFlow H3	0.6913	0.8592

Dataset 1 results average three independently trained models. On Dataset 2, KaryoFlow variants average three independently trained models; comparison detectors average three inference runs using one trained checkpoint per detector. The result index records both training and inference seeds. The manuscript tables provide the accompanying quantitative results.

Choose a checkpoint

Model	Dataset 1 configurations	Dataset 2 configurations	Checkpoint filename
KaryoFlow	KaryoFlow-D1	KaryoFlow-D2	d1_karyoflow_seed42.pth / d2_karyoflow_seed42.pth
KaryoFlow + LQCR	LQCR-D1	LQCR-D2	d1_karyoflow_lqcr_seed42.pth / d2_karyoflow_lqcr_seed42.pth
KaryoFlow H3	H3-D1	H3-D2	d1_karyoflow_h3_seed42.pth / d2_karyoflow_h3_seed42.pth

The filenames above illustrate seed 42. Seeds 123 and 789 are also included for each KaryoFlow variant on both datasets. H3 checkpoints contain student parameters only. Euler and Heun configurations reuse the corresponding KaryoFlow checkpoints.

The checkpoint directory also includes Cascade R-CNN, DiffusionDet, RTMDet-L and YOLOX-S comparison models. In total, 34 distinct checkpoint files support 54 recorded evaluations. Use models.json to select the exact configuration–checkpoint pair rather than inferring it from the filename alone.

Download and evaluate

First obtain the code package and follow its installation and data-preparation instructions. Run the following commands from the code repository root.

Download one model

hf download procaffenator/KaryoFlow \
  weights/d2_karyoflow_lqcr_seed42.pth --local-dir .

Evaluate the corresponding configuration:

python tools/run.py test configs/d2/karyoflow_lqcr/seed_42.py \
  --weights weights/d2_karyoflow_lqcr_seed42.pth \
  --data-root /path/to/dataset2/coco \
  --output outputs/d2_lqcr_seed42

Download all checkpoints

hf download procaffenator/KaryoFlow --include "weights/*.pth" --local-dir .
python tools/check_release.py --weights

These checkpoints use the KaryoFlow/MMDetection implementation. Downloading them does not require converting them to Transformers models. A single checkpoint evaluates one training run; reproducing a three-training-run mean requires all three corresponding checkpoints.

Data and intended use

The models are intended for chromosome-localization research and reproduction of the reported experiments. They have not been validated as a standalone clinical diagnostic system.

Dataset 1 images are not distributed. Downloading its checkpoints does not provide access to the underlying cohort. Dataset 2 must be obtained from its original source and prepared using the provided conversion instructions. See dataset sources and preparation.

License and attribution

Original KaryoFlow code and documentation, together with the contributors' rights in the 34 trained checkpoints in this repository, are licensed under the Apache License, Version 2.0. The checkpoint license statement lists every covered file and its upstream initialization source. A license copy is also provided in the checkpoint directory.

Upstream pretrained parameters retain their applicable terms and notices. Code adapted from DiffusionDet remains subject to CC BY-NC 4.0; the checkpoint grant does not remove the noncommercial conditions on that code. Datasets and example images retain their source terms. See License scope, NOTICE and third-party attribution.