dicom-native 0.2.0

Feed raw DICOM X-rays into PyTorch and YOLO — preserves 6.1x more intensity data than JPG pipelines. Validated on 71 real DICOMs with Ultralytics YOLOv8.

dicom_native

Feed raw DICOM X-rays directly into PyTorch U-Net and Ultralytics YOLO — zero disk conversion, full 12-bit/16-bit fidelity.

---

Validation Results

Validated on 71 real DICOM files, Ultralytics YOLOv8, Windows 11, Python 3.13, CPU-only.

Metric	Result
DICOM files tested	71
Bit depths covered	1, 8, 12, 13, 14, 16, 32-bit
Photometric types	7 (MONO1/2, RGB, YBR, palette)
YOLO training epochs	30
read_dicom() calls (training)	333 (100% .dcm)
Training loss trend	Decreased all 3
NaN in training	None
Memory leak (100 iterations)	0.7 MB (negligible)
Native vs JPG intensity levels	1,453 vs 238 (6.1x)
Crashes	0

---

Windows / CPU Quick Start

If you are on Windows or a CPU-only machine, use these safe training defaults:

from dicom_native.integrations.yolo import build_dicom_yolo_trainer

trainer = build_dicom_yolo_trainer(
    data="dataset/data.yaml",
    model="yolov8n.pt",
    epochs=30,
    imgsz=640,
    batch=4,
    # These safe defaults are already set automatically:
    #   amp=False        (prevents CPU segfault)
    #   mosaic=0.0       (prevents IndexError on small datasets)
    #   close_mosaic=0   (prevents mosaic re-enabling)
    #   workers=0        (prevents Windows multiprocessing crashes)
)

trainer.train()

On a Linux machine with a GPU, override the defaults:

trainer = build_dicom_yolo_trainer(
    data="dataset/data.yaml",
    model="yolov8s.pt",
    epochs=100,
    imgsz=640,
    batch=16,
    extra_args={
        "amp": True,         # safe with CUDA
        "mosaic": 1.0,       # enable mosaic with large datasets
        "workers": 8,        # multi-worker loading on Linux
        "device": "0",       # GPU device
    },
)

---

Why?

Converting DICOMs to .jpg / .png before training silently destroys data:

Format	Bit depth	Unique intensity levels
JPEG	8-bit	256
PNG	8-bit	256
DICOM	12–16 bit	4 096 – 65 536

Bone cortex, soft-tissue planes, and early lesions live in those extra bits.
dicom_native reads .dcm files in memory, applies correct medical preprocessing, and hands tensors straight to your training loop.

---

Installation

# Core library
pip install dicom_native

# With PyTorch support
pip install "dicom_native[torch]"

# With YOLO support
pip install "dicom_native[yolo]"

# Everything
pip install "dicom_native[all]"

---

Core concepts

read_dicom() — the I/O engine

Every path through this library starts here. The function performs four steps in order:

1. Load pixel data via pydicom (no temp files).
2. Apply RescaleSlope / RescaleIntercept if present in the DICOM header (linearises detector response).
3. Invert if PhotometricInterpretation == MONOCHROME1 (so bright always means high density).
4. Normalise to float32 in [0, 1].

from dicom_native import read_dicom

# NumPy array (H, W) float32
array = read_dicom("chest_pa.dcm")

# PyTorch tensor (1, H, W) float32 — ready for Conv2d
tensor = read_dicom("chest_pa.dcm", output="torch")

---

PyTorch U-Net training

1. Prepare your data

data/
├── images/
│   ├── patient_001.dcm
│   ├── patient_002.dcm
│   └── ...
└── masks/
    ├── patient_001.png   ← binary segmentation mask
    ├── patient_002.png
    └── ...

2. Build dataset & dataloader

import torch
from torch.utils.data import DataLoader, random_split

from dicom_native.integrations.pytorch import NativeDicomDataset
from dicom_native.transforms import (
    Compose, Resize, PercentileClip,
    RandomHorizontalFlip, GaussianNoise,
)

# ── Augmentation pipeline (operates on float32 — no quantisation) ──────────
train_tfm = Compose([
    Resize((512, 512)),
    PercentileClip(p_low=1, p_high=99),   # suppress artefacts
    RandomHorizontalFlip(p=0.5),
    GaussianNoise(std=0.02, p=0.3),
])

val_tfm = Compose([
    Resize((512, 512)),
    PercentileClip(p_low=1, p_high=99),
])

# ── Dataset ─────────────────────────────────────────────────────────────────
full_ds = NativeDicomDataset(
    image_dir="data/images",
    mask_dir="data/masks",
    transform=train_tfm,        # swap to val_tfm for validation split
    num_classes=1,              # binary segmentation
)

n_val = int(0.2 * len(full_ds))
n_train = len(full_ds) - n_val
train_ds, val_ds = random_split(full_ds, [n_train, n_val])

train_dl = DataLoader(train_ds, batch_size=4, shuffle=True,  num_workers=4, pin_memory=True)
val_dl   = DataLoader(val_ds,   batch_size=4, shuffle=False, num_workers=2, pin_memory=True)

print(f"Train: {len(train_ds)} | Val: {len(val_ds)}")

3. Training loop (U-Net example)

import torch
import torch.nn as nn

# Replace with your U-Net implementation, e.g. segmentation_models_pytorch
# pip install segmentation-models-pytorch
import segmentation_models_pytorch as smp

device = torch.device("cuda" if torch.cuda.is_available() else "cpu")

model = smp.Unet(
    encoder_name="resnet34",
    encoder_weights="imagenet",
    in_channels=1,          # grayscale DICOM
    classes=1,
).to(device)

optimizer = torch.optim.AdamW(model.parameters(), lr=1e-4)
criterion = smp.losses.DiceLoss(mode="binary")

for epoch in range(1, 51):
    # ── Train ──────────────────────────────────────────────────────────────
    model.train()
    for images, masks in train_dl:
        # images: (B, 1, H, W) float32  |  masks: (B, 1, H, W) float32
        images, masks = images.to(device), masks.to(device)

        preds = model(images)
        loss  = criterion(preds, masks)

        optimizer.zero_grad()
        loss.backward()
        optimizer.step()

    # ── Validate ───────────────────────────────────────────────────────────
    model.eval()
    dice_scores = []
    with torch.no_grad():
        for images, masks in val_dl:
            images, masks = images.to(device), masks.to(device)
            preds = torch.sigmoid(model(images)) > 0.5
            # Compute Dice per batch
            intersection = (preds * masks).sum()
            dice = (2 * intersection) / (preds.sum() + masks.sum() + 1e-6)
            dice_scores.append(dice.item())

    print(f"Epoch {epoch:3d} | Val Dice: {sum(dice_scores)/len(dice_scores):.4f}")

---

Ultralytics YOLO training

1. Prepare your data

dataset/
├── images/
│   ├── train/
│   │   ├── patient_001.dcm
│   │   └── patient_002.dcm
│   └── val/
│       └── patient_003.dcm
├── labels/
│   ├── train/
│   │   ├── patient_001.txt   ← YOLO format bounding boxes
│   │   └── patient_002.txt
│   └── val/
│       └── patient_003.txt
└── data.yaml

data.yaml:

path: /absolute/path/to/dataset
train: images/train
val:   images/val

nc: 2
names: ["fracture", "lesion"]

Label format (patient_001.txt):

0  0.512  0.388  0.124  0.096   # class x_c y_c w h  (all normalised to [0,1])
1  0.310  0.701  0.088  0.072

2. Train

from dicom_native.integrations.yolo import build_dicom_yolo_trainer

trainer = build_dicom_yolo_trainer(
    data="dataset/data.yaml",
    model="yolov8s.pt",          # or "yolo11s.pt"
    epochs=100,
    imgsz=640,
    batch=16,
)

trainer.train()
# Weights saved to runs/detect/train/weights/best.pt

3. Inference on a new DICOM

from dicom_native.integrations.yolo import dicom_predict

results = dicom_predict(
    model_path="runs/detect/train/weights/best.pt",
    dicom_path="new_patient.dcm",
    conf=0.4,
)

for r in results:
    print(r.boxes.xyxy)   # bounding box coordinates
    r.show()              # display with matplotlib
    r.save("output.jpg")  # save annotated image

---

Transform reference

Class	Purpose
Compose([...])	Chain multiple transforms
Resize((H, W))	Bilinear resize for images, nearest for masks
MinMaxNorm()	Re-normalise to [0, 1] after augmentation
PercentileClip(p_low, p_high)	Clip + normalise; suppresses implant artefacts
WindowLevel(ww, wc)	Radiological window/level contrast
RandomHorizontalFlip(p)	Random L-R flip
RandomVerticalFlip(p)	Random U-D flip
RandomRotate90(p)	Random 90°/180°/270° rotation
GaussianNoise(std, p)	Additive Gaussian noise

All transforms accept (image, mask=None) and return (image, mask).

---

API reference

read_dicom(path, *, output, force_single_channel, dtype)

Parameter	Default	Description
path	required	Path to .dcm file
output	"numpy"	"numpy" → (H,W) ndarray · "torch" → (1,H,W) tensor
force_single_channel	True	Average RGB channels if present
dtype	np.float32	Output numeric type

read_dicom_metadata(path)

Returns a dict of key DICOM tags without loading pixel data (fast).

NativeDicomDataset(...)

Parameter	Default	Description
image_dir	required	Directory of .dcm files
mask_dir	required	Directory of mask files
transform	None	Compose pipeline
mask_suffix	""	Extra suffix in mask filename
num_classes	1	>1 returns one-hot mask (C,H,W)
skip_missing_masks	False	Skip unpaired images silently

build_dicom_yolo_trainer(data, model, epochs, imgsz, batch, target_size, extra_args)

Returns a patched DetectionTrainer. Call .train() to start.

---

Design decisions

Why not subclass YOLODataset statically?
Ultralytics' internal class hierarchy changes frequently across minor versions.
Monkey-patching load_image on a constructed instance is more robust.

Why convert back to uint8 for YOLO?
The YOLO backbone, Albumentations augmentations, and letterbox all assume uint8 BGR. Passing float32 would silently corrupt augmentations like RandomBrightnessContrast. The conversion preserves relative contrast; absolute 16-bit precision is lost, but the backbone never saw those raw values anyway.

Why PercentileClip instead of plain MinMax?
Metal implants and collimator edges create extreme outliers. A global min-max would compress 95% of diagnostically relevant tissue into a narrow band.

---

License

MIT — see LICENSE.