2nd/10_Wiki/Topics/AI_and_ML/Prenatal-Neurology.md

---
id: wiki-2026-0508-prenatal-neurology
title: Prenatal Neurology
category: 10_Wiki/Topics
status: verified
canonical_id: self
aliases: [Fetal Neurology, Prenatal Neuroscience]
duplicate_of: none
source_trust_level: A
confidence_score: 0.85
verification_status: applied
tags: [neurology, fetal-medicine, neurodevelopment, medical-imaging]
raw_sources: []
last_reinforced: 2026-05-10
github_commit: pending
tech_stack:
  language: Python
  framework: MONAI / nnU-Net / 3D Slicer
---

# Prenatal Neurology

## 매 한 줄
> **"매 fetal nervous system 의 development, imaging, anomaly detection — neural tube 부터 birth 까지"**. 1980s ultrasound 의 advent 로 시작, 2010s fetal MRI 로 detail 폭증, 2020s deep learning 으로 automated segmentation/screening. 2026 currently SVRTK + diffusion priors 로 motion-corrected fetal MRI volumes 를 minutes 안에.

## 매 핵심

### 매 developmental milestones
- **Week 3-4**: neural plate → neural tube closure. Failure → spina bifida, anencephaly.
- **Week 5-7**: 3 primary vesicles → 5 secondary (telencephalon, diencephalon, mesenc, metenc, myelenc).
- **Week 8-16**: neuronal proliferation in ventricular zone.
- **Week 12-22**: neuronal migration along radial glia. Failure → lissencephaly, heterotopia.
- **Week 22-40**: gyrification, cortical organization, myelination begins.

### 매 imaging modalities
- **Ultrasound (US)**: routine 18-22 wk anatomy scan; transvaginal early.
- **Fetal MRI**: T2-HASTE / SSFSE; problem-solving when US ambiguous.
- **Doppler**: middle cerebral artery flow (anemia, hypoxia).
- **Fetal MEG / EEG**: research only.

### 매 common anomalies
1. **Neural tube defects** (NTDs): spina bifida, anencephaly. Folate-preventable.
2. **Ventriculomegaly**: atrial width >10mm.
3. **Corpus callosum agenesis**: 1:4000.
4. **Posterior fossa**: Dandy-Walker, Blake's pouch cyst.
5. **Cortical malformations**: lissencephaly, polymicrogyria.
6. **TORCH infections**: CMV, Zika → microcephaly, calcifications.

### 매 AI in fetal imaging (2024-2026)
- **SVRTK / NiftyMIC**: slice-to-volume reconstruction from motion-corrupted MRI.
- **nnU-Net fetal**: automatic brain extraction + tissue segmentation.
- **dHCP atlas**: developing Human Connectome Project — gestational-age-specific atlas.
- **Diffusion priors**: latent diffusion models for fetal MRI super-resolution (2024-2025).
- **Automated biometry**: BPD, HC, AC, FL from US in real time (e.g., Caption Health-style).

## 💻 패턴

### Fetal brain extraction (nnU-Net)
```python
# Train on FeTA Challenge dataset (gestational ages 20-35 wk)
# nnU-Net handles preprocessing, augmentation, ensemble
import subprocess
subprocess.run([
    "nnUNetv2_train", "Dataset080_FetalBrain", "3d_fullres", "0",
    "--npz",
])
# Inference
subprocess.run([
    "nnUNetv2_predict",
    "-i", "input_dir", "-o", "output_dir",
    "-d", "080", "-c", "3d_fullres", "-f", "0",
])
```

### Slice-to-volume reconstruction (SVRTK)
```bash
# Motion-corrupted T2 stacks → isotropic 3D volume
mirtk reconstruct recon.nii.gz \
  4 stack_axi.nii.gz stack_cor.nii.gz stack_sag.nii.gz stack_obl.nii.gz \
  -mask brain_mask.nii.gz \
  -resolution 0.5 \
  -iterations 3 \
  -thickness 3.0 3.0 3.0 3.0
```

### Tissue segmentation w/ MONAI
```python
import torch
from monai.networks.nets import SwinUNETR
from monai.transforms import Compose, LoadImaged, NormalizeIntensityd, EnsureChannelFirstd

model = SwinUNETR(img_size=(96, 96, 96), in_channels=1, out_channels=8,
                  feature_size=48, use_checkpoint=True)
model.load_state_dict(torch.load("feta_swinunetr.pt"))
# Outputs: CSF, GM, WM, ventricles, cerebellum, brainstem, deep GM, hippocampus
```

### Gestational-age-specific atlas registration
```python
# dHCP: 36 atlases from 28-44 weeks PMA
import ants
fixed = ants.image_read(f"dhcp_atlas/week_{ga_weeks}.nii.gz")
moving = ants.image_read("fetal_brain_recon.nii.gz")
reg = ants.registration(fixed, moving, type_of_transform="SyN")
warped = reg["warpedmovout"]
```

### Automated US biometry (real-time)
```python
# YOLOv8 finds standard plane → keypoint regression for BPD/HC/AC/FL
from ultralytics import YOLO
plane_model = YOLO("us_plane_classifier.pt")
biometry = YOLO("us_keypoints.pt")
res = plane_model(frame)
if res[0].names[res[0].probs.top1] == "axial_thalami":
    pts = biometry(frame)[0].keypoints
    bpd_mm = euclidean(pts[0], pts[1]) * pixel_spacing
```

### Cortical folding metric (gyrification index)
```python
# GI = total surface area / convex hull area (per hemisphere)
import nibabel as nib, numpy as np
from skimage.measure import marching_cubes, mesh_surface_area
seg = nib.load("cortex.nii.gz").get_fdata() > 0
verts, faces, _, _ = marching_cubes(seg, level=0.5)
surf = mesh_surface_area(verts, faces)
# Convex hull surface
from scipy.spatial import ConvexHull
hull = ConvexHull(verts)
gi = surf / hull.area
```

## 매 결정 기준
| 상황 | Approach |
|---|---|
| Routine screening 18-22 wk | Ultrasound (anatomy scan) |
| Suspected CNS anomaly on US | Fetal MRI (32-34 wk optimal) |
| Motion-corrupted MRI | SVRTK reconstruction |
| Quantitative volumetry | dHCP atlas + nnU-Net |
| Suspected NTD | High-resolution US + AFP + acetylcholinesterase |

**기본값**: US first; MRI for problem-solving; AI segmentation for research/quantitative endpoints.

## 🔗 Graph

## 🤖 LLM 활용
**언제**: fetal imaging analysis, neurodevelopmental research, congenital anomaly screening pipelines.
**언제 X**: clinical diagnosis without licensed clinician — AI augments, never replaces.

## ❌ 안티패턴
- **Adult MRI tools on fetal data**: gestational-age-specific contrast / atlas required.
- **Ignoring motion artifact**: fetal motion → reconstruct first.
- **No GA stratification**: 24wk vs 36wk brain are different organs.
- **Single-modality conclusion**: combine US + MRI + genetics.
- **Overcalling ventriculomegaly**: 10-12mm often resolves; counsel carefully.

## 🧪 검증 / 중복
- Verified (FeTA Challenge MICCAI, dHCP, ISUOG guidelines, AIUM practice parameters).
- 신뢰도 A.

## 🕓 Changelog
| 날짜 | 변경 |
|---|---|
| 2026-05-08 | Phase 1 |
| 2026-05-10 | Manual cleanup — fetal neurodevelopment + AI imaging stack |