File size: 11,252 Bytes
23f06d2 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 |
import os, re, math, json, argparse, warnings
from pathlib import Path
from typing import List, Tuple, Dict, Optional
import numpy as np, h5py, pyedflib
# ---------- 10-20 electrode 2D coords (x,y) đã chuẩn hoá tương đối ----------
E1020 = {
"FP1":(-0.5, 1.0), "FP2":( 0.5, 1.0),
"F7": (-1.0, 0.6), "F3":(-0.3, 0.6), "FZ":(0.0,0.8), "F4":( 0.3, 0.6), "F8":( 1.0, 0.6),
"T3":(-1.2, 0.2), "T7":(-1.2, 0.2), "C3":(-0.4, 0.2), "CZ":(0.0,0.3), "C4":( 0.4, 0.2), "T4":( 1.2, 0.2), "T8":( 1.2, 0.2),
"T5":(-1.1, -0.2), "P7":(-1.1, -0.2), "P3":(-0.3, -0.2), "PZ":(0.0,-0.1), "P4":( 0.3, -0.2), "T6":( 1.1, -0.2), "P8":( 1.1, -0.2),
"O1":(-0.5, -0.8), "O2":( 0.5, -0.8)
}
ALIASES = {"T3":"T7","T4":"T8","T5":"P7","T6":"P8"} # đồng nhất tên
_BLOCKLIST = re.compile(r'(?:ECG|VNS|-$)', re.IGNORECASE)
def _norm_e(name:str)->Optional[str]:
name=name.upper().strip()
name=ALIASES.get(name,name)
return name if name in E1020 else None
def pair_midpoint(ch_label:str)->Optional[Tuple[float,float]]:
"""
CHB-MIT thường dùng kênh lưỡng cực 'A-B'. Lấy toạ độ là trung điểm của hai điện cực A,B.
"""
if '-' not in ch_label: return None
a,b = [x.strip().upper() for x in ch_label.split('-',1)]
a=_norm_e(a); b=_norm_e(b)
if (a is None) or (b is None): return None
ax,ay=E1020[a]; bx,by=E1020[b]
return ((ax+bx)/2.0, (ay+by)/2.0)
def build_1020_graph(channels:List[str], k:int=8, sigma:float=0.5, radius:float=1.6):
"""
Đồ thị khoảng cách tĩnh theo 10–20:
w_ij = exp(-||vi-vj||^2 / sigma^2), nếu ||vi-vj|| <= radius, ngược lại 0.
Sau đó với mỗi nút, giữ lại k láng giềng w lớn nhất (undirected).
"""
coords=[]
keep=[]
for ch in channels:
if _BLOCKLIST.search(ch): continue
m=pair_midpoint(ch)
if m is not None:
coords.append(m); keep.append(ch)
C=len(keep)
if C<2:
raise RuntimeError("Không đủ kênh ánh xạ được sang 10-20")
coords=np.array(coords, dtype=np.float32)
D=np.sqrt(((coords[None,:,:]-coords[:,None,:])**2).sum(axis=-1)) # (C,C)
W=np.exp(-(D**2)/(sigma**2))
W[D>radius]=0.0
np.fill_diagonal(W, 0.0)
edges=set()
for i in range(C):
idx=np.argsort(-W[i])[:max(1,min(k,C-1))]
for j in idx:
a,b=(i,j) if i<j else (j,i)
if W[a,b]>0: edges.add((a,b))
edges=sorted(list(edges))
ei=np.array(edges, dtype=np.int64).T
ei=np.hstack([ei, ei[::-1,:]]) if ei.size else ei
ew=np.array([W[i,j] for (i,j) in edges], dtype=np.float32)
ew=np.concatenate([ew, ew]) if ew.size else ew
return keep, ei, ew
# ---------- EDF utils ----------
def list_edf_files(patient_dir: Path):
return sorted([p for p in patient_dir.glob("*.edf") if p.is_file()])
def edf_channel_labels(edf_path: Path):
f=pyedflib.EdfReader(str(edf_path))
labels=[f.getLabel(i).strip() for i in range(f.signals_in_file)]
f._close(); del f
return [ch for ch in labels if not _BLOCKLIST.search(ch)]
def intersection_channels(edf_paths):
common=None
for p in edf_paths:
chans=set(edf_channel_labels(p))
if not chans: continue
common = chans if common is None else (common & chans)
return sorted(list(common)) if common else []
def read_edf_signals(edf_path: Path, keep_channels):
f=pyedflib.EdfReader(str(edf_path))
labels=[f.getLabel(i).strip() for i in range(f.signals_in_file)]
fs_all=[int(round(f.getSampleFrequency(i))) for i in range(f.signals_in_file)]
fs=int(round(np.median(fs_all))) if fs_all else 256
idxs=[]; out_labels=[]
for ch in keep_channels:
try:
i=labels.index(ch)
if _BLOCKLIST.search(ch): continue
idxs.append(i); out_labels.append(ch)
except ValueError:
f._close(); del f
raise RuntimeError(f"Channel {ch} not found in {edf_path.name}")
sigs=np.vstack([f.readSignal(i) for i in idxs]).astype(np.float32) # (C,N)
f._close(); del f
return sigs, fs, out_labels
# ---------- nhãn co giật ----------
def parse_summary(summary_path: Path) -> Dict[str, List[Tuple[float,float]]]:
mapping={}
if not (summary_path and summary_path.exists()):
return mapping
curr=None; buf=[]
with summary_path.open("r", errors="ignore") as f:
for line in f:
line=line.strip()
mfile=re.search(r'File Name:\s*(\S+\.edf)', line, re.IGNORECASE)
if mfile:
if curr and buf:
pairs=[(buf[i], buf[i+1]) for i in range(0,len(buf)-1,2)]
mapping.setdefault(curr, []).extend(pairs)
curr=mfile.group(1); buf=[]; continue
if re.search(r'Seizure (Start|End) Time', line, re.IGNORECASE):
nums=[float(x) for x in re.findall(r'[\d.]+', line)]
if nums: buf.extend(nums)
if curr and buf:
pairs=[(buf[i], buf[i+1]) for i in range(0,len(buf)-1,2)]
mapping.setdefault(curr, []).extend(pairs)
return mapping
def parse_seizures_file(seiz_file: Path) -> List[Tuple[float,float]]:
intervals=[]
if not seiz_file.exists(): return intervals
with seiz_file.open("r", errors="ignore") as f:
for line in f:
nums=[float(x) for x in re.findall(r'[-+]?\d*\.?\d+', line)]
if len(nums)>=2: intervals.append((nums[0], nums[1]))
return intervals
def slice_starts(N, fs, clip_sec, hop_sec):
T=int(fs*clip_sec); hop=int(fs*hop_sec)
if N<T: return np.zeros((0,), dtype=np.int64)
return np.arange(0, N-T+1, hop, dtype=np.int64)
def zscore_perclip(clip):
mu=clip.mean(axis=1, keepdims=True); sd=clip.std(axis=1, keepdims=True)+1e-8
return (clip-mu)/sd
def label_for_window(start, T, fs, intervals, min_overlap_sec):
a,b=start, start+T
thr=int(round(min_overlap_sec*fs)); overlap=0
for (u,v) in intervals:
u_s=int(round(u*fs)); v_s=int(round(v*fs))
overlap += max(0, min(b, v_s) - max(a, u_s))
if overlap>=thr: return 1
return 0
class H5Appender:
def __init__(self, out_path: Path, C: int, T: int, fs: int, channels, edge_index, edge_weight, gzip_level=4):
self.f=h5py.File(str(out_path),"w")
self.ds_clips=self.f.create_dataset("clips", shape=(0,C,T,1), maxshape=(None,C,T,1),
dtype="float32", chunks=(16,C,T,1), compression="gzip", compression_opts=gzip_level)
self.ds_labels=self.f.create_dataset("labels", shape=(0,), maxshape=(None,), dtype="i8",
chunks=True, compression="gzip", compression_opts=gzip_level)
self.ds_fileids=self.f.create_dataset("file_ids", shape=(0,), maxshape=(None,), dtype="i4",
chunks=True, compression="gzip", compression_opts=gzip_level)
self.f.attrs["fs"]=fs; self.f.attrs["T"]=T; self.f.attrs["patient"]=out_path.stem
self.f.create_dataset("channels", data=np.array([c.encode() for c in channels]))
self.f.create_dataset("edge_index", data=edge_index.astype(np.int64))
self.f.create_dataset("edge_weight", data=edge_weight.astype(np.float32))
self.n=0
def append(self, clips_CT, labels, file_id:int):
if clips_CT.size==0: return
M,C,T=clips_CT.shape
clips=clips_CT[...,None].astype(np.float32)
self.ds_clips.resize(self.n+M, axis=0)
self.ds_labels.resize(self.n+M, axis=0)
self.ds_fileids.resize(self.n+M, axis=0)
self.ds_clips[self.n:self.n+M]=clips
self.ds_labels[self.n:self.n+M]=labels.astype(np.int64)
self.ds_fileids[self.n:self.n+M]=np.full((M,), file_id, dtype=np.int32)
self.n+=M
def close(self): self.f.close()
def process_patient_1020(root: Path, patient: str, out_path: Path,
clip_sec: float=4.0, hop_sec: float=2.0,
fs_target: int=256, min_overlap_sec: float=0.25,
graph_k: int=8, sigma: float=0.5, radius: float=1.6):
pat_dir=root/patient
assert pat_dir.exists(), f"Not found: {pat_dir}"
edfs=list_edf_files(pat_dir); assert edfs, f"No EDF in {pat_dir}"
# kênh giao nhau trong bệnh nhân
keep_channels=intersection_channels(edfs)
# chỉ giữ kênh map được sang 10-20 (bỏ kênh lạ)
keep_channels=[ch for ch in keep_channels if pair_midpoint(ch) is not None]
if len(keep_channels)<8:
warnings.warn(f"[{patient}] chỉ còn {len(keep_channels)} kênh sau khi map 10-20")
# đồ thị 10-20 tĩnh
keep_channels, ei, ew = build_1020_graph(keep_channels, k=graph_k, sigma=sigma, radius=radius)
summ = next(iter(list(pat_dir.glob("*summary*.txt"))), None)
summ_map=parse_summary(summ) if summ else {}
seiz_map={}
for p in edfs:
iv=parse_seizures_file(p.with_suffix(p.suffix+".seizures"))
if iv: seiz_map[p.name]=iv
def intervals_for(name): return seiz_map.get(name, summ_map.get(name, []))
app=None; file_id=0
total_pos=0; total=0
for edf in edfs:
sigs, fs, chans=read_edf_signals(edf, keep_channels)
# resample
if fs!=fs_target:
ratio=fs_target/fs
N_new=int(round(sigs.shape[1]*ratio))
t_old=np.linspace(0, sigs.shape[1]-1, sigs.shape[1], dtype=np.float32)
t_new=np.linspace(0, sigs.shape[1]-1, N_new, dtype=np.float32)
sigs=np.stack([np.interp(t_new, t_old, ch) for ch in sigs], axis=0).astype(np.float32)
C,N=sigs.shape; T=int(fs_target*clip_sec)
starts=slice_starts(N, fs_target, clip_sec, hop_sec)
if app is None:
app=H5Appender(out_path, C=C, T=T, fs=fs_target, channels=chans, edge_index=ei, edge_weight=ew)
ivals=intervals_for(edf.name)
labels=np.array([label_for_window(int(s), T, fs_target, ivals, min_overlap_sec) for s in starts], dtype=np.int64)
M=len(starts); clips=np.empty((M,C,T), dtype=np.float32)
for i,s in enumerate(starts): clips[i]=zscore_perclip(sigs[:, s:s+T])
app.append(clips, labels, file_id=file_id)
total += M; total_pos += int(labels.sum()); file_id+=1
app.close()
print(f"[Done] {patient} -> {out_path} | clips={total} (pos={total_pos}, neg={total-total_pos}), C={len(keep_channels)}, T={int(fs_target*clip_sec)}")
if __name__=="__main__":
import argparse
ap=argparse.ArgumentParser()
ap.add_argument("--root", required=True)
ap.add_argument("--patient", required=True)
ap.add_argument("--out", required=True)
ap.add_argument("--clip-sec", type=float, default=4.0)
ap.add_argument("--hop-sec", type=float, default=2.0)
ap.add_argument("--fs", type=int, default=256)
ap.add_argument("--min-overlap", type=float, default=0.25)
ap.add_argument("--k", type=int, default=8)
ap.add_argument("--sigma", type=float, default=0.5)
ap.add_argument("--radius", type=float, default=1.6)
args=ap.parse_args()
process_patient_1020(Path(args.root), args.patient, Path(args.out),
clip_sec=args.clip_sec, hop_sec=args.hop_sec, fs_target=args.fs,
min_overlap_sec=args.min_overlap, graph_k=args.k, sigma=args.sigma, radius=args.radius)
|