btxrd-workstation / src /data /dataset.py
amitaddi
feat: deploy bone tumor segmentation workstation
1bd13a5
Raw History Blame Contribute Delete
13.9 kB
"""
Dataset classes for Bone Tumor Segmentation.
Provides:
- BTXRDSegmentationDataset: For UNet training (returns image + mask pairs)
- BTXRDSAMDataset: For SAM/MedSAM (returns image + bounding box/point prompts)
"""
import os
import glob
import numpy as np
import cv2
from PIL import Image
from pathlib import Path
from typing import Optional, List, Dict, Tuple, Callable
import torch
from torch.utils.data import Dataset
import albumentations as A
from albumentations.pytorch import ToTensorV2
from src.data.yolo_to_masks import parse_yolo_label, yolo_to_xyxy, get_bboxes_for_sam
# ──────────────────────────────────────────────────────────────────────
# Augmentation pipelines
# ──────────────────────────────────────────────────────────────────────
def get_unet_train_transforms(image_size: int = 512) -> A.Compose:
"""Training augmentations for UNet (Enhanced for Medical Segmentation)."""
return A.Compose([
A.Resize(image_size, image_size),
A.HorizontalFlip(p=0.5),
A.VerticalFlip(p=0.5),
A.ShiftScaleRotate(shift_limit=0.1, scale_limit=0.1, rotate_limit=30, border_mode=cv2.BORDER_CONSTANT, p=0.5),
A.ElasticTransform(alpha=1, sigma=50, alpha_affine=50, p=0.3),
A.GridDistortion(num_steps=5, distort_limit=0.3, p=0.3),
A.RandomBrightnessContrast(brightness_limit=0.2, contrast_limit=0.2, p=0.5),
A.GaussNoise(var_limit=(10.0, 50.0), p=0.3),
A.CLAHE(clip_limit=4.0, p=0.3),
A.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]),
ToTensorV2(),
])
def get_unet_val_transforms(image_size: int = 512) -> A.Compose:
"""Validation transforms for UNet (no augmentation)."""
return A.Compose([
A.Resize(image_size, image_size),
A.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]),
ToTensorV2(),
])
def get_sam_transforms(image_size: int = 1024) -> A.Compose:
"""Transforms for SAM input (minimal β€” SAM has its own preprocessor)."""
return A.Compose([
A.Resize(image_size, image_size),
])
# ──────────────────────────────────────────────────────────────────────
# UNet Segmentation Dataset
# ──────────────────────────────────────────────────────────────────────
class BTXRDSegmentationDataset(Dataset):
"""
Dataset for UNet-based segmentation training.
Loads X-ray images and corresponding bounding-box masks.
Each sample is (image_tensor, mask_tensor, metadata_dict).
"""
def __init__(self,
images_dir: str,
labels_dir: str,
masks_dir: Optional[str] = None,
transform: Optional[A.Compose] = None,
image_size: int = 512,
binary: bool = True):
"""
Args:
images_dir: Path to image directory
labels_dir: Path to YOLO label directory
masks_dir: Path to pre-generated masks (if None, creates on-the-fly)
transform: Albumentations transform pipeline
image_size: Target image size
binary: Binary segmentation (True) or multi-class (False)
"""
self.images_dir = images_dir
self.labels_dir = labels_dir
self.masks_dir = masks_dir
self.transform = transform
self.image_size = image_size
self.binary = binary
# Collect valid image-label pairs
self.samples = []
label_files = sorted(glob.glob(os.path.join(labels_dir, '*.txt')))
for label_path in label_files:
stem = Path(label_path).stem
# Find image
img_path = None
for ext in ['.jpeg', '.jpg', '.png']:
candidate = os.path.join(images_dir, stem + ext)
if os.path.exists(candidate):
img_path = candidate
break
if img_path is not None:
self.samples.append({
'image_path': img_path,
'label_path': label_path,
'stem': stem,
})
def __len__(self) -> int:
return len(self.samples)
def __getitem__(self, idx: int) -> Tuple[torch.Tensor, torch.Tensor, Dict]:
sample = self.samples[idx]
# Load image
image = cv2.imread(sample['image_path'])
image = cv2.cvtColor(image, cv2.COLOR_BGR2RGB)
img_h, img_w = image.shape[:2]
# Load or create mask
if self.masks_dir and os.path.exists(os.path.join(self.masks_dir, sample['stem'] + '.png')):
mask = cv2.imread(os.path.join(self.masks_dir, sample['stem'] + '.png'),
cv2.IMREAD_GRAYSCALE)
mask = (mask > 127).astype(np.uint8)
else:
# Create mask on-the-fly from YOLO labels
from src.data.yolo_to_masks import create_bbox_mask
annotations = parse_yolo_label(sample['label_path'])
mask = create_bbox_mask(img_w, img_h, annotations, binary=self.binary)
# Apply transforms
if self.transform:
transformed = self.transform(image=image, mask=mask)
image = transformed['image']
mask = transformed['mask']
# Ensure mask is float tensor with channel dim
if isinstance(mask, np.ndarray):
mask = torch.from_numpy(mask)
mask = mask.float().unsqueeze(0) # (1, H, W)
metadata = {
'stem': sample['stem'],
'original_size': (img_h, img_w),
}
return image, mask, metadata
# ──────────────────────────────────────────────────────────────────────
# SAM / MedSAM Prompt Dataset
# ──────────────────────────────────────────────────────────────────────
class BTXRDSAMDataset(Dataset):
"""
Dataset for SAM2/MedSAM evaluation.
Returns image as numpy array + list of prompt dicts (one per object).
Each prompt dict has 'box', 'center_point', 'class_id'.
Also returns bounding-box masks as ground truth for evaluation.
"""
def __init__(self,
images_dir: str,
labels_dir: str,
image_size: int = 1024,
expand_ratio: float = 0.0):
"""
Args:
images_dir: Path to image directory
labels_dir: Path to YOLO label directory
image_size: SAM input size (typically 1024)
expand_ratio: How much to expand bounding boxes
"""
self.images_dir = images_dir
self.labels_dir = labels_dir
self.image_size = image_size
self.expand_ratio = expand_ratio
# Collect valid samples
self.samples = []
label_files = sorted(glob.glob(os.path.join(labels_dir, '*.txt')))
for label_path in label_files:
stem = Path(label_path).stem
img_path = None
for ext in ['.jpeg', '.jpg', '.png']:
candidate = os.path.join(images_dir, stem + ext)
if os.path.exists(candidate):
img_path = candidate
break
if img_path is not None:
annotations = parse_yolo_label(label_path)
if len(annotations) > 0:
self.samples.append({
'image_path': img_path,
'label_path': label_path,
'stem': stem,
'annotations': annotations,
})
def __len__(self) -> int:
return len(self.samples)
def __getitem__(self, idx: int) -> Tuple[np.ndarray, List[Dict], np.ndarray, Dict]:
"""
Returns:
image: (H, W, 3) RGB numpy array (resized to image_size)
prompts: List of prompt dicts for SAM
gt_mask: (H, W) binary mask from bounding boxes
metadata: Dict with original size, stem, class info
"""
sample = self.samples[idx]
# Load image
image = cv2.imread(sample['image_path'])
image = cv2.cvtColor(image, cv2.COLOR_BGR2RGB)
orig_h, orig_w = image.shape[:2]
# Get prompts at original resolution
prompts = get_bboxes_for_sam(
orig_w, orig_h, sample['annotations'], self.expand_ratio
)
# Create GT mask at original resolution
from src.data.yolo_to_masks import create_bbox_mask
gt_mask = create_bbox_mask(orig_w, orig_h, sample['annotations'], binary=True)
# Resize image for SAM
scale_x = self.image_size / orig_w
scale_y = self.image_size / orig_h
image_resized = cv2.resize(image, (self.image_size, self.image_size))
gt_mask_resized = cv2.resize(gt_mask, (self.image_size, self.image_size),
interpolation=cv2.INTER_NEAREST)
# Scale prompts to resized coordinates
scaled_prompts = []
for p in prompts:
box = p['box'].copy().astype(float)
box[0] *= scale_x
box[1] *= scale_y
box[2] *= scale_x
box[3] *= scale_y
center = p['center_point'].copy().astype(float)
center[0] *= scale_x
center[1] *= scale_y
scaled_prompts.append({
'box': box.astype(int),
'class_id': p['class_id'],
'center_point': center.astype(int),
'box_original': p['box'],
})
metadata = {
'stem': sample['stem'],
'original_size': (orig_h, orig_w),
'resized_size': (self.image_size, self.image_size),
'scale': (scale_x, scale_y),
'num_objects': len(prompts),
'class_ids': [p['class_id'] for p in prompts],
}
return image_resized, scaled_prompts, gt_mask_resized, metadata
# ──────────────────────────────────────────────────────────────────────
# Utility: DataLoader creation
# ──────────────────────────────────────────────────────────────────────
def create_unet_dataloaders(project_root: str,
batch_size: int = 8,
image_size: int = 512,
num_workers: int = 4):
"""Create train and val DataLoaders for UNet."""
from torch.utils.data import DataLoader
train_ds = BTXRDSegmentationDataset(
images_dir=os.path.join(project_root, 'BTXRD dataset', 'images_512', 'train'),
labels_dir=os.path.join(project_root, 'BTXRD dataset', 'labels', 'train'),
masks_dir=os.path.join(project_root, 'BTXRD dataset', 'masks_512', 'train'),
transform=get_unet_train_transforms(image_size),
image_size=image_size,
)
val_ds = BTXRDSegmentationDataset(
images_dir=os.path.join(project_root, 'BTXRD dataset', 'images_512', 'val'),
labels_dir=os.path.join(project_root, 'BTXRD dataset', 'labels', 'val'),
masks_dir=os.path.join(project_root, 'BTXRD dataset', 'masks_512', 'val'),
transform=get_unet_val_transforms(image_size),
image_size=image_size,
)
train_loader = DataLoader(
train_ds, batch_size=batch_size, shuffle=True,
num_workers=num_workers, pin_memory=True, drop_last=True
)
val_loader = DataLoader(
val_ds, batch_size=batch_size, shuffle=False,
num_workers=num_workers, pin_memory=True
)
return train_loader, val_loader
if __name__ == '__main__':
# Quick test
project_root = '.'
print("Testing BTXRDSegmentationDataset...")
ds = BTXRDSegmentationDataset(
images_dir=os.path.join(project_root, 'BTXRD dataset', 'images', 'val'),
labels_dir=os.path.join(project_root, 'BTXRD dataset', 'labels', 'val'),
transform=get_unet_val_transforms(512),
)
print(f" Val dataset size: {len(ds)}")
img, mask, meta = ds[0]
print(f" Image shape: {img.shape}, Mask shape: {mask.shape}")
print(f" Mask unique values: {mask.unique().tolist()}")
print("\nTesting BTXRDSAMDataset...")
sam_ds = BTXRDSAMDataset(
images_dir=os.path.join(project_root, 'BTXRD dataset', 'images', 'val'),
labels_dir=os.path.join(project_root, 'BTXRD dataset', 'labels', 'val'),
)
print(f" SAM dataset size: {len(sam_ds)}")
img, prompts, gt_mask, meta = sam_ds[0]
print(f" Image shape: {img.shape}")
print(f" Num prompts: {len(prompts)}")
print(f" GT mask shape: {gt_mask.shape}")
print(f" First prompt box: {prompts[0]['box']}")