2nd/10_Wiki/Topics/AI_and_ML/Medical-Imaging-Data-Augmentation.md

id, title, category, status, canonical_id, aliases, duplicate_of, source_trust_level, confidence_score, verification_status, tags, raw_sources, last_reinforced, github_commit, tech_stack

id	title	category	status	canonical_id	aliases	duplicate_of	source_trust_level	confidence_score	verification_status	tags	raw_sources	last_reinforced	github_commit	tech_stack
wiki-2026-0508-medical-imaging-data-augmentation	Medical Imaging Data Augmentation	10_Wiki/Topics	verified	self	Medical Augmentation MONAI Augmentation 의료영상 증강	none	A	0.9	applied	medical-imaging data-augmentation monai deep-learning segmentation		2026-05-10	pending	language framework python monai-pytorch

Medical Imaging Data Augmentation

매 한 줄

"매 환자 데이터는 매 적고, 매 anatomy 는 망가뜨릴 수 없다". 의료영상 augmentation 은 일반 이미지 대비 (1) 데이터가 매우 적고 (2) 라벨이 픽셀 단위 정확해야 하며 (3) 비현실적 변형이 진단을 망친다는 제약 안에서 기하·강도·합성 변환을 신중히 적용해야 한다.

매 핵심

매 도메인 특성

• 3D volume (CT/MRI), DICOM/NIfTI 포맷, voxel spacing 다양.
• 라벨이 segmentation mask / bbox / 환자 단위 진단 — affine 변환 시 동기화.
• HU scale (CT), bias field (MRI) 등 강도 분포가 modality 마다 다름.

매 변환 카테고리

1. Spatial / 기하: flip, rotate (소각), translate, scale, elastic deformation, B-spline.
2. Intensity: brightness/contrast, gamma, Gaussian noise, Rician noise (MRI), bias field, MR motion artifact.
3. Spacing / Resolution: random resample, low-res sim.
4. Topology-preserving: mixup/cutmix 의 의료 variant — 단, lesion mask 가 깨지지 않도록 patch-aware.
5. Synthesis: GAN/diffusion 으로 lesion 합성, healthy↔lesion translation.

매 라이브러리

1. MONAI: PyTorch 기반 의료영상 표준, Compose, dictionary transform.
2. TorchIO: 3D 친화, MRI artifact 풍부.
3. Albumentations: 2D slice/엑스레이.
4. NVIDIA DALI: GPU augmentation 파이프라인.

💻 패턴

1. MONAI dict transform pipeline

from monai.transforms import (Compose, LoadImaged, EnsureChannelFirstd,
    Spacingd, Orientationd, ScaleIntensityRanged, RandCropByPosNegLabeld,
    RandAffined, RandGaussianNoised, RandBiasFieldd, ToTensord)

train_t = Compose([
    LoadImaged(keys=["img","seg"]),
    EnsureChannelFirstd(keys=["img","seg"]),
    Orientationd(keys=["img","seg"], axcodes="RAS"),
    Spacingd(keys=["img","seg"], pixdim=(1,1,1), mode=("bilinear","nearest")),
    ScaleIntensityRanged(keys="img", a_min=-200, a_max=300, b_min=0, b_max=1, clip=True),
    RandCropByPosNegLabeld(keys=["img","seg"], label_key="seg",
                           spatial_size=(96,96,96), pos=1, neg=1, num_samples=4),
    RandAffined(keys=["img","seg"], rotate_range=0.1, scale_range=0.1,
                mode=("bilinear","nearest"), prob=0.5),
    RandGaussianNoised(keys="img", std=0.01, prob=0.2),
    RandBiasFieldd(keys="img", coeff_range=(0.0,0.1), prob=0.2),
    ToTensord(keys=["img","seg"]),
])

2. Elastic deformation

from monai.transforms import Rand3DElasticd
Rand3DElasticd(keys=["img","seg"], sigma_range=(5,7), magnitude_range=(50,150),
               mode=("bilinear","nearest"), prob=0.3)

Anatomy 자연스러운 변형 — sigma 너무 작으면 비현실, 크면 underfit.

3. TorchIO MRI artifact

import torchio as tio
tx = tio.Compose([
    tio.RandomMotion(degrees=5, translation=5, p=0.2),
    tio.RandomGhosting(num_ghosts=(2,5), p=0.2),
    tio.RandomBiasField(coefficients=0.3, p=0.3),
    tio.RandomNoise(std=(0,0.05), p=0.3),
])

4. CT HU window robust

def hu_window(vol, center, width):
    lo, hi = center - width/2, center + width/2
    return np.clip((vol - lo) / (hi - lo), 0, 1)
# 학습 중 center/width 를 약하게 jitter

5. Lesion-aware MixUp (segmentation)

def lesion_mixup(x1, y1, x2, y2, alpha=0.2):
    lam = np.random.beta(alpha, alpha)
    x = lam*x1 + (1-lam)*x2
    y = (y1 + y2).clip(0,1)        # union mask
    return x, y

6. Diffusion synthetic lesion

# Stable Diffusion fine-tuned on chest X-ray with lesion mask conditioning
img = pipe(prompt="pneumonia consolidation right lower lobe", mask=mask).images[0]
# 합성 데이터는 별도 split, 평가는 real only

7. Test-Time Augmentation (TTA)

preds = []
for tta in [identity, flip_x, flip_y, rot90]:
    preds.append(undo(model(tta(x))))
final = torch.stack(preds).mean(0)

매 결정 기준

상황	Augmentation
작은 segmentation 데이터	strong elastic + intensity + cropbypos/neg
분류 (X-ray)	mild affine + cutout, lesion 보호
MRI multi-site	bias field + intensity histogram match
CT multi-protocol	HU window jitter + spacing resample
매우 적은 라벨	+ synthetic (GAN/diffusion) + self-supervised pretrain

기본값: MONAI Compose + spacing/orient 정규화 → mild affine + intensity + (3D 면) crop-by-label.

🔗 Graph

• 부모: Data-Augmentation
• Adjacent: Diffusion-Model

🤖 LLM 활용

언제: pipeline boilerplate, modality 별 적절 변환 추천, 코드 review. 언제 X: 임상적 plausibility 판단 (radiologist 검증 필요).

❌ 안티패턴

• 90° rotate / 큰 scale → CT 좌표계/anatomy 깨짐.
• segmentation mask 에 bilinear interpolation — 라벨 손상.
• intensity normalize 를 augmentation 후 적용 → 분포 불일치.
• synthetic 데이터를 real 평가셋과 섞기 — leakage.
• 모든 patient slice 를 독립 sample 로 — patient-level leakage, split 단위는 patient.

🧪 검증 / 중복

• Verified. 신뢰도 A.

🕓 Changelog

날짜	변경
2026-05-08	Phase 1
2026-05-10	Manual cleanup