denoise.data

class denoise.data.PatchIndex(d_idx: int, top: int, left: int)

Bases: object

One patch address within the volume. d_idx: center slice index for 2.5D stack top,left: spatial location within padded H/W coordinates

d_idx: int

left: int

top: int

class denoise.data.TilingMeta(D: int, H_in: int, W_in: int, H_pad: int, W_pad: int, ph: int, pw: int, stride_h: int, stride_w: int, n_rows: int, n_cols: int, P_per_slice: int, pad_bottom: int, pad_right: int, neighbors: int, edge_mode: str, pad_mode: str)

Bases: object

Metadata needed for stitching (later) and sanity checks.

D: int

H_in: int

H_pad: int

P_per_slice: int

W_in: int

W_pad: int

edge_mode: str

n_cols: int

n_rows: int

neighbors: int

pad_bottom: int

pad_mode: str

pad_right: int

ph: int

pw: int

stride_h: int

stride_w: int

class denoise.data.TomoDatasetInfer(*args: Any, **kwargs: Any)

Bases: Dataset

Dataset that yields overlapping patches from a CT volume (NumPy) with 2.5D channels.

Input volume: vol [D, H, W] (NumPy array)

Each item returns:

• x_patch: torch.FloatTensor [C, ph, pw]

• info: dict with patch coordinates and indices (for stitching later)

stitch_predictions(pred_patches: numpy.ndarray, *, window: Literal['uniform', 'hann', 'cosine'] = 'hann', output_size: Tuple[int, int] | None = None, keep_k_dim: bool = True, eps: float = 1e-06) → numpy.ndarray

Stitch patch predictions back into a (sub)volume using overlap-add blending.

Parameters:

• pred_patches –

  Patch predictions in dataset order (same order as self.index):

  • [T, ph, pw] regression (implicit K=1)

  • [T, 1, ph, pw] regression (explicit K=1)

  • [T, K, ph, pw] segmentation logits/probs

  where T must equal len(self.index) == total_patches.

• window – Blending window: ‘uniform’ (avg) or ‘hann’/’cosine’ (recommended).

• output_size – (H, W) to crop final result. If None, uses (H_in, W_in).

• keep_k_dim –

  If pred_patches was [T, ph, pw], output can be:

  • keep_k_dim=True -> [D_sel, 1, H, W]

  • keep_k_dim=False -> [D_sel, H, W]

  If pred_patches already has K dim, output keeps it.

• eps – Small constant to avoid divide-by-zero in normalization.

Returns:

out – - [D_sel, K, H, W] (typical)

• or [D_sel, H, W] if K==1 and keep_k_dim=False

where D_sel = (d_end - d_start), i.e., only the slices this dataset processed. These correspond to the center slice predictions in your 2.5D setup.

class denoise.data.TomoDatasetTrain(*args: Any, **kwargs: Any)

Bases: Dataset

Training class for 2.5D N2I.

This class loads in two lists corresponding to the two sub reconstructions (saved as .tiffs) and normalizes them.

params:

• params (obj) yaml object, essentially a dictionary

• config_file (str) location of the configuration file

denoise.data.save_normalization_value(config_file, mean, std)

This functin saves the mean and standard deviation back to the yaml file which is then used during inferencing

params:

• config_file (str) location of the config file

• mean (float) mean used for normalization

• std (float) standard deviation used for normalization