""" 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 — 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']}")