Postprocessing cochlea segmentation using erosion

Author: schilling40

flamingo_tools/segmentation/postprocessing.py

@@ -1,4 +1,5 @@
 import multiprocessing as mp
+import os
 from concurrent import futures
 from typing import Callable, Tuple, Optional
 
@@ -8,8 +9,11 @@
 import pandas as pd
 
 from elf.io import open_file
-from scipy.spatial import distance
+from scipy.ndimage import binary_fill_holes
+from scipy.ndimage import distance_transform_edt
+from scipy.ndimage import label
 from scipy.sparse import csr_matrix
+from scipy.spatial import distance
 from scipy.spatial import cKDTree, ConvexHull
 from skimage import measure
 from sklearn.neighbors import NearestNeighbors
@@ -205,3 +209,237 @@ def filter_chunk(block_id):
     )
 
     return n_ids, n_ids_filtered
+
+
+# Postprocess segmentation by erosion using the above spatial statistics.
+# Currently implemented using downscaling and looking for connected components
+# TODO: Change implementation to graph connected components.
+
+
+def erode_subset(
+    table: pd.DataFrame,
+    iterations: Optional[int] = 1,
+    min_cells: Optional[int] = None,
+    threshold: Optional[int] = 35,
+    keyword: Optional[str] = "distance_nn100",
+) -> pd.DataFrame:
+    """Erode coordinates of dataframe according to a keyword and a threshold.
+    Use a copy of the dataframe as an input, if it should not be edited.
+
+    Args:
+        table: Dataframe of segmentation table.
+        iterations: Number of steps for erosion process.
+        min_cells: Minimal number of rows. The erosion is stopped before reaching this number.
+        threshold: Upper threshold for removing elements according to the given keyword.
+        keyword: Keyword of dataframe for erosion.
+
+    Returns:
+        The dataframe containing elements left after the erosion.
+    """
+    print("initial length", len(table))
+    n_neighbors = 100
+    for i in range(iterations):
+        table = table[table[keyword] < threshold]
+
+        # TODO: support other spatial statistics
+        distance_avg = nearest_neighbor_distance(table, n_neighbors=n_neighbors)
+
+        if min_cells is not None and len(distance_avg) < min_cells:
+            print(f"{i}-th iteration, length of subset {len(table)}, stopping erosion")
+            break
+
+        table.loc[:, 'distance_nn'+str(n_neighbors)] = list(distance_avg)
+
+        print(f"{i}-th iteration, length of subset {len(table)}")
+
+    return table
+
+
+def downscaled_centroids(
+    table: pd.DataFrame,
+    scale_factor: int,
+    ref_dimensions: Optional[Tuple[float,float,float]] = None,
+    capped: Optional[bool] = True,
+) -> np.typing.NDArray:
+    """Downscale centroids in dataframe.
+
+    Args:
+        table: Dataframe of segmentation table.
+        scale_factor: Factor for downscaling coordinates.
+        ref_dimensions: Reference dimensions for downscaling. Taken from centroids if not supplied.
+        capped: Flag for capping output of array at 1 for the creation of a binary mask.
+
+    Returns:
+        The downscaled array
+    """
+    centroids = list(zip(table["anchor_x"], table["anchor_y"], table["anchor_z"]))
+    centroids_scaled = [(c[0] / scale_factor, c[1] / scale_factor, c[2] / scale_factor) for c in centroids]
+
+    if ref_dimensions is None:
+        bounding_dimensions = (max(table["anchor_x"]), max(table["anchor_y"]), max(table["anchor_z"]))
+        bounding_dimensions_scaled = tuple([round(b // scale_factor + 1) for b in bounding_dimensions])
+        new_array = np.zeros(bounding_dimensions_scaled)
+
+    else:
+        bounding_dimensions_scaled = tuple([round(b // scale_factor + 1) for b in ref_dimensions])
+        new_array = np.zeros(bounding_dimensions_scaled)
+
+    for c in centroids_scaled:
+        new_array[int(c[0]), int(c[1]), int(c[2])] += 1
+
+    array_downscaled = np.round(new_array).astype(int)
+
+    if capped:
+        array_downscaled[array_downscaled >= 1] = 1
+
+    return array_downscaled
+
+
+def coordinates_in_downscaled_blocks(
+    table: pd.DataFrame,
+    down_array: np.typing.NDArray,
+    scale_factor: float,
+    distance_component: Optional[int] = 0,
+) -> list:
+    """Checking if coordinates are within the downscaled array.
+
+    Args:
+        table: Dataframe of segmentation table.
+        down_array: Downscaled array.
+        scale_factor: Factor which was used for downscaling.
+        distance_component: Distance in downscaled units to which centroids next to downscaled blocks are included.
+
+    Returns:
+        A binary list representing whether the dataframe coordinates are within the array.
+    """
+    # fill holes in down-sampled array
+    down_array[down_array > 0] = 1
+    down_array = binary_fill_holes(down_array).astype(np.uint8)
+
+    # check if input coordinates are within down-sampled blocks
+    centroids = list(zip(table["anchor_x"], table["anchor_y"], table["anchor_z"]))
+    centroids_scaled = [np.floor(np.array([c[0]/scale_factor, c[1]/scale_factor, c[2]/scale_factor])) for c in centroids]
+
+    distance_map = distance_transform_edt(down_array == 0)
+
+    centroids_binary = []
+    for c in centroids_scaled:
+        coord = (int(c[0]), int(c[1]), int(c[2]))
+        if down_array[coord] != 0:
+            centroids_binary.append(1)
+        elif distance_map[coord] <= distance_component:
+            centroids_binary.append(1)
+        else:
+            centroids_binary.append(0)
+
+    return centroids_binary
+
+
+def erode_sgn_seg(
+    table: pd.DataFrame,
+    keyword: Optional[str] = "distance_nn100",
+    filter_small_components: Optional[int] = None,
+    scale_factor: Optional[float] = 20,
+    threshold_erode: Optional[float] = None,
+) -> Tuple[pd.DataFrame,np.typing.NDArray,np.typing.NDArray,np.typing.NDArray]:
+    """Eroding the SGN segmentation.
+
+    Args:
+        table: Dataframe of segmentation table.
+        keyword: Keyword of the dataframe column for erosion.
+        filter_small_components: Filter components smaller after n blocks after labeling.
+        scale_factor: Scaling for downsampling.
+        threshold_erode: Threshold of column value after erosion step with spatial statistics.
+
+    Returns:
+        The labeled components of the downscaled, eroded coordinates.
+        The larget connected component of the labeled components.
+    """
+
+    ref_dimensions = (max(table["anchor_x"]), max(table["anchor_y"]), max(table["anchor_z"]))
+    print("initial length", len(table))
+    distance_nn = list(table[keyword])
+    distance_nn.sort()
+
+    if len(table) < 20000:
+        iterations = 1
+        min_cells = None
+        average_dist = int(distance_nn[int(len(table) * 0.8)])
+        threshold = threshold_erode if threshold_erode is not None else average_dist
+    else:
+        iterations = 15
+        min_cells = 20000
+        threshold = threshold_erode if threshold_erode is not None else 40
+
+    print(f"Using threshold of {threshold} micrometer for eroding segmentation with keyword {keyword}.")
+
+    new_subset = erode_subset(table.copy(), iterations=iterations,
+                              threshold=threshold, min_cells=min_cells, keyword=keyword)
+    eroded_arr = downscaled_centroids(new_subset, scale_factor=scale_factor, ref_dimensions=ref_dimensions)
+    # Label connected components
+    labeled, num_features = label(eroded_arr)
+
+    # Find the largest component
+    sizes = [(labeled == i).sum() for i in range(1, num_features + 1)]
+    largest_label = np.argmax(sizes) + 1
+
+    # Extract only the largest component
+    largest_component = (labeled == largest_label).astype(np.uint8)
+    largest_component_filtered = binary_fill_holes(largest_component).astype(np.uint8)
+
+    #filter small sizes
+    if filter_small_components is not None:
+        for (size, feature) in zip(sizes, range(1, num_features + 1)):
+            if size < filter_small_components:
+                labeled[labeled == feature] = 0
+
+    return labeled, largest_component_filtered
+
+
+def get_components(table: pd.DataFrame,
+    labeled: np.typing.NDArray,
+    scale_factor: float,
+    distance_component: Optional[int] = 0,
+) -> list:
+    """Indexing coordinates according to labeled array.
+
+    Args:
+        table: Dataframe of segmentation table.
+        labeled: Array containing differently labeled components.
+        scale_factor: Scaling for downsampling.
+        distance_component: Distance in downscaled units to which centroids next to downscaled blocks are included.
+
+    Returns:
+        List of component labels.
+    """
+    unique_labels = list(np.unique(labeled))
+    component_labels = [0 for _ in range(len(table))]
+    for label_index, l in enumerate(unique_labels):
+        if l != 0:
+            label_arr = (labeled == l).astype(np.uint8)
+            centroids_binary = coordinates_in_downscaled_blocks(table, label_arr,
+                                                                scale_factor, distance_component = distance_component)
+            for num, c in enumerate(centroids_binary):
+                if c != 0:
+                    component_labels[num] = label_index
+    return component_labels
+
+
+def postprocess_sgn_seg(table: pd.DataFrame, scale_factor: Optional[float] = 20) -> pd.DataFrame:
+    """Postprocessing SGN segmentation of cochlea.
+
+    Args:
+        table: Dataframe of segmentation table.
+        scale_factor: Scaling for downsampling.
+
+    Returns:
+        Dataframe with component labels.
+    """
+    labeled, largest_component = erode_sgn_seg(table, filter_small_labels=10,
+                                               scale_factor=scale_factor, threshold_erode=None)
+
+    component_labels = get_components(table, labeled, scale_factor, distance_component = 1)
+
+    table.loc[:, "component_labels"] = component_labels
+
+    return table

scripts/prediction/expand_seg_table.py

@@ -61,7 +61,7 @@ def main(
             neighbor_counts = [n[0] for n in neighbor_counts]
             tsv_table['neighbors_in_radius'+str(r_neighbor)] = neighbor_counts
 
-    tsv_table.to_csv(out_path, sep="\t")
+    tsv_table.to_csv(out_path, sep="\t", index=False)
 
 
 if __name__ == "__main__":