File size: 5,588 Bytes
f9b628d | 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 | # ── Cell 1: write dataloader.py ─────────────────────────────────────────────
import os
import cv2
import numpy as np
import torch
from pathlib import Path
from torch.utils.data import Dataset
class GlaucomaDataset(Dataset):
"""
Local-disk fundus dataset with image-driven synthetic mask generation.
Mask convention
---------------
0 = background | 1 = optic disc | 2 = optic cup
ONH centre estimated per-image from the brightest region of the image,
so every sample gets a unique mask. Training split applies augmentation.
"""
def __init__(self, root_dir: str, split: str = "train"):
self.root_dir = Path(root_dir)
self.split = split
self.augment = (split == "train")
self.image_paths: list = []
self.labels: list = []
self._load_index()
def _split_folder(self) -> str:
return {"train": "train set", "validation": "validation set",
"test": "test set"}.get(self.split, "train set")
def _load_index(self):
base = self.root_dir / self._split_folder()
for label, sub in [(0, "nrg"), (1, "rg")]:
folder = base / sub
if folder.exists():
for p in folder.rglob("*"):
if p.suffix.lower() in (".jpg", ".jpeg", ".png"):
self.image_paths.append(str(p))
self.labels.append(label)
def __len__(self):
return len(self.image_paths)
# ── ONH detection ─────────────────────────────────────────────────────────
def _onh_center(self, gray_u8: np.ndarray):
h, w = gray_u8.shape
my, mx = int(h * 0.15), int(w * 0.15)
roi = gray_u8[my: h - my, mx: w - mx]
# Large blur suppresses point reflections — a real ONH is a region, not a point
blurred = cv2.GaussianBlur(roi, (61, 61), 0)
# Centroid of top-5% brightest region is robust to specular artifacts
threshold = np.percentile(blurred, 95)
bright = (blurred >= threshold).astype(np.uint8)
n, labels, stats, centroids = cv2.connectedComponentsWithStats(bright)
if n > 1:
largest = 1 + int(np.argmax(stats[1:, cv2.CC_STAT_AREA]))
cx = int(centroids[largest][0]) + mx
cy = int(centroids[largest][1]) + my
else:
cy, cx = h // 2, w // 2
return cy, cx
def _make_mask(self, gray_f32: np.ndarray) -> np.ndarray:
h, w = gray_f32.shape
cy, cx = self._onh_center((gray_f32 * 255).astype(np.uint8))
base_r = int(min(h, w) * 0.13)
if self.augment:
r_jitter = int(base_r * 0.12)
disc_r = base_r + np.random.randint(-r_jitter, r_jitter + 1)
cup_r = int(disc_r * np.random.uniform(0.60, 0.72))
cj = int(base_r * 0.04)
cx = int(np.clip(cx + np.random.randint(-cj, cj + 1), disc_r, w - disc_r))
cy = int(np.clip(cy + np.random.randint(-cj, cj + 1), disc_r, h - disc_r))
else:
disc_r = base_r
cup_r = int(disc_r * 0.55)
cx = int(np.clip(cx, disc_r, w - disc_r))
cy = int(np.clip(cy, disc_r, h - disc_r))
y, x = np.ogrid[:h, :w]
d = np.sqrt((x - cx) ** 2 + (y - cy) ** 2)
mask = np.zeros((h, w), dtype=np.uint8)
mask[d <= disc_r] = 1
mask[d <= cup_r] = 2
return mask
# ── augmentation ──────────────────────────────────────────────────────────
def _augment(self, img: np.ndarray, mask: np.ndarray):
if np.random.random() > 0.5:
img, mask = cv2.flip(img, 1), cv2.flip(mask, 1)
if np.random.random() > 0.5:
img, mask = cv2.flip(img, 0), cv2.flip(mask, 0)
angle = np.random.uniform(-7, 7)
h, w = img.shape
M = cv2.getRotationMatrix2D((w / 2, h / 2), angle, 1.0)
kw_i = dict(flags=cv2.INTER_LINEAR, borderMode=cv2.BORDER_REFLECT_101)
kw_m = dict(flags=cv2.INTER_NEAREST, borderMode=cv2.BORDER_REFLECT_101)
img = cv2.warpAffine(img, M, (w, h), **kw_i)
mask = cv2.warpAffine(mask, M, (w, h), **kw_m)
alpha = np.random.uniform(0.92, 1.08)
beta = np.random.uniform(-0.02, 0.02)
img = np.clip(img * alpha + beta, 0.0, 1.0).astype(np.float32)
if np.random.random() > 0.5:
img = np.clip(img + np.random.normal(0, 0.005, img.shape).astype(np.float32),
0.0, 1.0)
return img, mask
# ── __getitem__ ────────────────────────────────────────────────────────────
def __getitem__(self, idx: int):
img = cv2.imread(self.image_paths[idx], cv2.IMREAD_GRAYSCALE)
if img is None:
img = np.zeros((512, 512), dtype=np.uint8)
img = cv2.resize(img, (512, 512)).astype(np.float32) / 255.0
mask = self._make_mask(img)
if self.augment:
img, mask = self._augment(img, mask)
return (torch.from_numpy(img).unsqueeze(0).float(),
torch.from_numpy(mask).long())
|