Datasets

The framework provides PyTorch-compatible dataset classes for common mammography datasets. All datasets follow the lazy loading pattern — only metadata is scanned at initialization; actual image loading happens in __getitem__.

Available Datasets

Class	Dataset	Features
INbreastDataset	INbreast	DICOM + XML/COCO annotations, on-the-fly mask generation
CustomINbreastDataset	Custom INbreast	Extends INbreast with TIF mask support
CBISDDSMDataset	CBIS-DDSM	Full-image and patch modes, automatic DICOM pairing

Output Format

All datasets return a standardized dictionary:

sample = dataset[0]
# {
#     "image":    torch.Tensor  [1, H, W],
#     "mask":     torch.Tensor  [1, H, W],
#     "metadata": {"case_id": "...", ...}
# }

INbreast Dataset

from medical_image.datasets.inbreast import INbreastDataset

dataset = INbreastDataset(
    root_dir="/path/to/INbreast Release 1.0",
    target_size=(512, 512),  # optional resize
)

sample = dataset[0]
image = sample["image"]     # [1, 512, 512]
mask = sample["mask"]       # [1, 512, 512] binary
meta = sample["metadata"]   # case_id, file info

Supports two directory layouts:

1. COCO JSON — annotations.json + images/ directory

2. XML (legacy) — AllDICOMs/, AllXML/, optional AllROI/

CBIS-DDSM Dataset

from medical_image.datasets.cbis_ddsm import CBISDDSMDataset

# Full-image mode
dataset = CBISDDSMDataset(
    root_dir="/path/to/CBIS-DDSM",
    percentage=0.5,   # use 50% of data
)

# Patch-based mode (sliding window)
dataset = CBISDDSMDataset(
    root_dir="/path/to/CBIS-DDSM",
    mode="patch",
    patch_size=256,
    stride=128,
)

PyTorch DataLoader Integration

from torch.utils.data import DataLoader

dataset = INbreastDataset("/path/to/data", target_size=(256, 256))
loader = DataLoader(
    dataset,
    batch_size=8,
    shuffle=True,
    collate_fn=dataset.collate_fn,
    num_workers=4,
)

for batch in loader:
    images = batch["image"]   # [B, 1, H, W]
    masks = batch["mask"]     # [B, 1, H, W]

Creating a Custom Dataset

Subclass BaseDataset and implement two methods:

from medical_image.datasets.base_dataset import BaseDataset

class MyDataset(BaseDataset):
    def _build_sample_list(self):
        # Scan root_dir and populate self._samples
        for path in self.root_dir.glob("*.dcm"):
            self._samples.append({"path": path})

    def _load_sample(self, idx):
        info = self._samples[idx]
        image = DicomImage(str(info["path"]))
        image.load()
        return {
            "image": image.pixel_data.unsqueeze(0),
            "mask": torch.zeros_like(image.pixel_data).unsqueeze(0),
            "metadata": {"path": str(info["path"])},
        }