Bits & Flames
bitsandflames/fyron

Dataset Preparation

This example prepares three common imaging AI dataset structures from synthetic local files:

  • nnU-Net v2 segmentation
  • YOLO classification
  • YOLO object detection

The example uses synthetic data only. Replace the manifest paths with reviewed local files from your study.

Create Synthetic Files

python
import pandas as pd
from fyron.datasets import make_nifti_pair, make_dicom_series

pair_a = make_nifti_pair("scratch/case_a")
pair_b = make_nifti_pair("scratch/case_b")
make_dicom_series("scratch/case_c_dicom", shape=(4, 8, 8))

nnU-Net v2 Dataset

python
from fyron.datasets import prepare_nnunetv2_dataset

nnunet_manifest = pd.DataFrame(
    {
        "case_id": ["case_a", "case_b", "case_c"],
        "split": ["train", "val", "test"],
        "image_path": [pair_a["image"], pair_b["image"], "scratch/case_c_dicom"],
        "label_path": [pair_a["mask"], pair_b["mask"], ""],
    }
)

nnunet_written = prepare_nnunetv2_dataset(
    nnunet_manifest,
    output_dir="datasets/nnunet",
    dataset_id=501,
    dataset_name="DemoTumor",
)

Result:

text
datasets/nnunet/Dataset501_DemoTumor/
  imagesTr/case_a_0000.nii.gz
  labelsTr/case_a.nii.gz
  imagesTs/case_c_0000.nii.gz
  dataset.json

YOLO Classification Dataset

python
from fyron.datasets import prepare_yolo_classification_dataset

classification_manifest = pd.DataFrame(
    {
        "patient_id": ["case_a", "case_b", "case_c"],
        "split": ["train", "val", "test"],
        "image_path": [pair_a["image"], pair_b["image"], "scratch/case_c_dicom"],
        "class": ["benign", "malignant", "benign"],
    }
)

yolo_cls = prepare_yolo_classification_dataset(
    classification_manifest,
    output_dir="datasets/yolo_cls",
    sample_id_col="patient_id",
    nifti_slice="middle",
)

Result:

text
datasets/yolo_cls/
  train/benign/case_a_z0002.png
  val/malignant/case_b_z0002.png
  test/benign/case_c_z0002.png

YOLO Detection Dataset

python
from fyron.datasets import prepare_yolo_detection_dataset

boxes = pd.DataFrame(
    {
        "image_id": ["case_a", "case_a", "case_b"],
        "split": ["train", "train", "val"],
        "image_path": [pair_a["image"], pair_a["image"], pair_b["image"]],
        "class": ["lesion", "device", "lesion"],
        "x_min": [1, 4, None],
        "y_min": [1, 4, None],
        "x_max": [5, 7, None],
        "y_max": [5, 7, None],
    }
)

yolo_det = prepare_yolo_detection_dataset(
    boxes,
    output_dir="datasets/yolo_det",
    sample_id_col="image_id",
)

Result:

text
datasets/yolo_det/
  images/train/case_a_z0002.png
  labels/train/case_a_z0002.txt
  images/val/case_b_z0002.png
  labels/val/case_b_z0002.txt
  data.yaml

The empty box row for case_b creates an empty label file, which YOLO uses for negative images.

Validate Outputs

python
from fyron.datasets import validate_nnunetv2_dataset, validate_yolo_dataset

nnunet_qc = validate_nnunetv2_dataset("datasets/nnunet/Dataset501_DemoTumor")
yolo_cls_qc = validate_yolo_dataset("datasets/yolo_cls", task="classify")
yolo_det_qc = validate_yolo_dataset("datasets/yolo_det", task="detect")

Keep the returned manifests and validation tables beside the training data so the exact preparation step is reproducible.

CLI Version

bash
fyron dataset-nnunetv2 \
  --manifest nnunet_manifest.csv \
  --output-dir datasets/nnunet \
  --dataset-id 501 \
  --dataset-name DemoTumor
bash
fyron dataset-yolo-classify \
  --manifest classification_manifest.csv \
  --output-dir datasets/yolo_cls \
  --sample-id-col patient_id
bash
fyron dataset-yolo-detect \
  --manifest boxes.csv \
  --output-dir datasets/yolo_det \
  --sample-id-col image_id