Add script for measuring the number of SGNs

environment.yaml

@@ -12,6 +12,7 @@ dependencies:
   - pytorch
   - s3fs
   - torch_em
+  - trimesh
   - z5py
   # Don't install zarr v3, as we are not sure that it is compatible with MoBIE etc. yet
   - zarr <3

flamingo_tools/measurements.py

@@ -0,0 +1,146 @@
+import multiprocessing as mp
+from concurrent import futures
+from typing import Optional
+
+import numpy as np
+import pandas as pd
+import trimesh
+from skimage.measure import marching_cubes
+from tqdm import tqdm
+
+from .file_utils import read_image_data
+from .segmentation.postprocessing import compute_table_on_the_fly
+
+
+def _measure_volume_and_surface(mask, resolution):
+    # Use marching_cubes for 3D data
+    verts, faces, normals, _ = marching_cubes(mask, spacing=(resolution,) * 3)
+
+    mesh = trimesh.Trimesh(vertices=verts, faces=faces, vertex_normals=normals)
+    surface = mesh.area
+    if mesh.is_watertight:
+        volume = np.abs(mesh.volume)
+    else:
+        volume = np.nan
+
+    return volume, surface
+
+
+def compute_object_measures_impl(
+    image: np.typing.ArrayLike,
+    segmentation: np.typing.ArrayLike,
+    n_threads: Optional[int] = None,
+    resolution: float = 0.38,
+    table: Optional[pd.DataFrame] = None,
+) -> pd.DataFrame:
+    """Compute simple intensity and morphology measures for each segmented cell in a segmentation.
+
+    See `compute_object_measures` for details.
+
+    Args:
+        image: The image data.
+        segmentation: The segmentation.
+        n_threads: The number of threads to use for computation.
+        resolution: The resolution / voxel size of the data.
+        table: The segmentation table. Will be computed on the fly if it is not given.
+
+    Returns:
+        The table with per object measurements.
+    """
+    if table is None:
+        table = compute_table_on_the_fly(segmentation, resolution=resolution)
+
+    def intensity_measures(seg_id):
+        # Get the bounding box.
+        row = table[table.label_id == seg_id]
+
+        bb_min = np.array([
+            row.bb_min_z.item(), row.bb_min_y.item(), row.bb_min_x.item()
+        ]).astype("float32") / resolution
+        bb_min = np.round(bb_min, 0).astype("int32")
+
+        bb_max = np.array([
+            row.bb_max_z.item(), row.bb_max_y.item(), row.bb_max_x.item()
+        ]).astype("float32") / resolution
+        bb_max = np.round(bb_max, 0).astype("int32")
+
+        bb = tuple(
+            slice(max(bmin - 1, 0), min(bmax + 1, sh))
+            for bmin, bmax, sh in zip(bb_min, bb_max, image.shape)
+        )
+
+        local_image = image[bb]
+        mask = segmentation[bb] == seg_id
+        assert mask.sum() > 0, f"Segmentation ID {seg_id} is empty."
+        masked_intensity = local_image[mask]
+
+        # Do the base intensity measurements.
+        measures = {
+            "label_id": seg_id,
+            "mean": np.mean(masked_intensity),
+            "stdev": np.std(masked_intensity),
+            "min": np.min(masked_intensity),
+            "max": np.max(masked_intensity),
+            "median": np.median(masked_intensity),
+        }
+        for percentile in (5, 10, 25, 75, 90, 95):
+            measures[f"percentile-{percentile}"] = np.percentile(masked_intensity, percentile)
+
+        # Do the volume and surface measurement.
+        volume, surface = _measure_volume_and_surface(mask, resolution)
+        measures["volume"] = volume
+        measures["surface"] = surface
+        return measures
+
+    seg_ids = table.label_id.values
+    assert len(seg_ids) > 0, "The segmentation table is empty."
+    n_threads = mp.cpu_count() if n_threads is None else n_threads
+    with futures.ThreadPoolExecutor(n_threads) as pool:
+        measures = list(tqdm(
+            pool.map(intensity_measures, seg_ids), total=len(seg_ids), desc="Compute intensity measures"
+        ))
+
+    # Create the result table and save it.
+    keys = measures[0].keys()
+    measures = pd.DataFrame({k: [measure[k] for measure in measures] for k in keys})
+    return measures
+
+
+# Could also support s3 directly?
+def compute_object_measures(
+    image_path: str,
+    segmentation_path: str,
+    segmentation_table_path: str,
+    output_table_path: str,
+    image_key: Optional[str] = None,
+    segmentation_key: Optional[str] = None,
+    n_threads: Optional[int] = None,
+    resolution: float = 0.38,
+) -> None:
+    """Compute simple intensity and morphology measures for each segmented cell in a segmentation.
+
+    This computes the mean, standard deviation, minimum, maximum, median and
+    5th, 10th, 25th, 75th, 90th and 95th percentile of the intensity image
+    per cell, as well as the volume and surface.
+
+    Args:
+        image_path: The filepath to the image data. Either a tif or hdf5/zarr/n5 file.
+        segmentation_path: The filepath to the segmentation data. Either a tif or hdf5/zarr/n5 file.
+        segmentation_table_path: The path to the segmentation table in MoBIE format.
+        output_table_path: The path for saving the segmentation with intensity measures.
+        image_key: The key (= internal path) for the image data. Not needed fir tif.
+        segmentation_key: The key (= internal path) for the segmentation data. Not needed for tif.
+        n_threads: The number of threads to use for computation.
+        resolution: The resolution / voxel size of the data.
+    """
+    # First, we load the pre-computed segmentation table from MoBIE.
+    table = pd.read_csv(segmentation_table_path, sep="\t")
+
+    # Then, open the volumes.
+    image = read_image_data(image_path, image_key)
+    segmentation = read_image_data(segmentation_path, segmentation_key)
+
+    measures = compute_object_measures_impl(
+        image, segmentation, n_threads, resolution, table=table
+    )
+    measures.to_csv(output_table_path, sep="\t", index=False)

flamingo_tools/segmentation/postprocessing.py

@@ -115,19 +115,39 @@ def neighbors_in_radius(table: pd.DataFrame, radius: float = 15) -> np.ndarray:
 #
 
 
-def _compute_table(segmentation, resolution):
+def compute_table_on_the_fly(segmentation: np.typing.ArrayLike, resolution: float) -> pd.DataFrame:
+    """Compute a segmentation table compatible with MoBIE.
+
+    The table contains information about the number of pixels per object,
+    the anchor (= centroid) and the bounding box. Anchor and bounding box are given in physical coordinates.
+
+    Args:
+        segmentation: The segmentation for which to compute the table.
+        resolution: The physical voxel spacing of the data.
+
+    Returns:
+        The segmentation table.
+    """
     props = measure.regionprops(segmentation)
     label_ids = np.array([prop.label for prop in props])
-    coordinates = np.array([prop.centroid for prop in props])
+    coordinates = np.array([prop.centroid for prop in props]).astype("float32")
     # transform pixel distance to physical units
     coordinates = coordinates * resolution
+    bb_min = np.array([prop.bbox[:3] for prop in props]).astype("float32") * resolution
+    bb_max = np.array([prop.bbox[3:] for prop in props]).astype("float32") * resolution
     sizes = np.array([prop.area for prop in props])
     table = pd.DataFrame({
         "label_id": label_ids,
-        "n_pixels": sizes,
         "anchor_x": coordinates[:, 2],
         "anchor_y": coordinates[:, 1],
         "anchor_z": coordinates[:, 0],
+        "bb_min_x": bb_min[:, 2],
+        "bb_min_y": bb_min[:, 1],
+        "bb_min_z": bb_min[:, 0],
+        "bb_max_x": bb_max[:, 2],
+        "bb_max_y": bb_max[:, 1],
+        "bb_max_z": bb_max[:, 0],
+        "n_pixels": sizes,
     })
     return table
 
@@ -160,13 +180,12 @@ def filter_segmentation(
         spatial_statistics_kwargs: Arguments for spatial statistics function
 
     Returns:
-        n_ids
-        n_ids_filtered
+        The number of objects before filtering.
+        The number of objects after filtering.
     """
-    # Compute the table on the fly.
-    # NOTE: this currently doesn't work for large segmentations.
+    # Compute the table on the fly. This doesn't work for large segmentations.
     if table is None:
-        table = _compute_table(segmentation, resolution=resolution)
+        table = compute_table_on_the_fly(segmentation, resolution=resolution)
     n_ids = len(table)
 
     # First apply the size filter.

flamingo_tools/segmentation/unet_prediction.py

@@ -324,6 +324,7 @@ def run_unet_prediction(
     scale: Optional[float] = None,
     block_shape: Optional[Tuple[int, int, int]] = None,
     halo: Optional[Tuple[int, int, int]] = None,
+    use_mask: bool = True,
 ) -> None:
     """Run prediction and segmentation with a distance U-Net.
 
@@ -337,10 +338,12 @@ def run_unet_prediction(
             By default the data will not be rescaled.
         block_shape: The block-shape for running the prediction.
         halo: The halo (= block overlap) to use for prediction.
+        use_mask: Whether to use the masking heuristics to not run inference on empty blocks.
     """
     os.makedirs(output_folder, exist_ok=True)
 
-    find_mask(input_path, input_key, output_folder)
+    if use_mask:
+        find_mask(input_path, input_key, output_folder)
 
     original_shape = prediction_impl(
         input_path, input_key, output_folder, model_path, scale, block_shape, halo

flamingo_tools/test_data.py

@@ -1,7 +1,75 @@
 import os
+from typing import Tuple
 
 import imageio.v3 as imageio
-from skimage.data import binary_blobs
+import requests
+from skimage.data import binary_blobs, cells3d
+from skimage.measure import label
+
+from .segmentation.postprocessing import compute_table_on_the_fly
+
+SEGMENTATION_URL = "https://owncloud.gwdg.de/index.php/s/kwoGRYiJRRrswgw/download"
+
+
+def get_test_volume_and_segmentation(folder: str) -> Tuple[str, str, str]:
+    """Download a small volume with nuclei and corresponding segmentation.
+
+    Args:
+        folder: The test data folder. The data will be downloaded to this folder.
+
+    Returns:
+        The path to the image, stored as tif.
+        The path to the segmentation, stored as tif.
+        The path to the segmentation table, stored as tsv.
+    """
+    os.makedirs(folder, exist_ok=True)
+
+    segmentation_path = os.path.join(folder, "segmentation.tif")
+    resp = requests.get(SEGMENTATION_URL)
+    resp.raise_for_status()
+
+    with open(segmentation_path, "wb") as f:
+        f.write(resp.content)
+
+    nuclei = cells3d()[20:40, 1]
+    segmentation = imageio.imread(segmentation_path)
+    assert nuclei.shape == segmentation.shape
+
+    image_path = os.path.join(folder, "image.tif")
+    imageio.imwrite(image_path, nuclei)
+
+    table_path = os.path.join(folder, "default.tsv")
+    table = compute_table_on_the_fly(segmentation, resolution=0.38)
+    table.to_csv(table_path, sep="\t", index=False)
+
+    return image_path, segmentation_path, table_path
+
+
+def create_image_data_and_segmentation(folder: str, size: int = 256) -> Tuple[str, str, str]:
+    """Create test data containing an image, a corresponding segmentation and segmentation table.
+
+    Args:
+        folder: The test data folder. The data will be written to this folder.
+
+    Returns:
+        The path to the image, stored as tif.
+        The path to the segmentation, stored as tif.
+        The path to the segmentation table, stored as tsv.
+    """
+    os.makedirs(folder, exist_ok=True)
+    data = binary_blobs(size, n_dim=3).astype("uint8") * 255
+    seg = label(data)
+
+    image_path = os.path.join(folder, "image.tif")
+    segmentation_path = os.path.join(folder, "segmentation.tif")
+    imageio.imwrite(image_path, data)
+    imageio.imwrite(segmentation_path, seg)
+
+    table_path = os.path.join(folder, "default.tsv")
+    table = compute_table_on_the_fly(seg, resolution=0.38)
+    table.to_csv(table_path, sep="\t", index=False)
+
+    return image_path, segmentation_path, table_path
 
 
 # TODO add metadata

scripts/measurements/measure_sgns.py

@@ -0,0 +1,48 @@
+import json
+import os
+
+import numpy as np
+import pandas as pd
+from flamingo_tools.s3_utils import create_s3_target, BUCKET_NAME
+
+
+def open_json(fs, path):
+    s3_path = os.path.join(BUCKET_NAME, path)
+    with fs.open(s3_path, "r") as f:
+        content = json.load(f)
+    return content
+
+
+def open_tsv(fs, path):
+    s3_path = os.path.join(BUCKET_NAME, path)
+    with fs.open(s3_path, "r") as f:
+        table = pd.read_csv(f, sep="\t")
+    return table
+
+
+def main():
+    fs = create_s3_target()
+    project_info = open_json(fs, "project.json")
+    for dataset in project_info["datasets"]:
+        if dataset == "fens":
+            continue
+        print(dataset)
+        dataset_info = open_json(fs, os.path.join(dataset, "dataset.json"))
+        sources = dataset_info["sources"]
+        for source, source_info in sources.items():
+            if not source.startswith("SGN"):
+                continue
+            assert "segmentation" in source_info
+            source_info = source_info["segmentation"]
+            table_path = source_info["tableData"]["tsv"]["relativePath"]
+            table = open_tsv(fs, os.path.join(dataset, table_path, "default.tsv"))
+            component_labels = table.component_labels.values
+            remaining_sgns = component_labels[component_labels != 0]
+            print(source)
+            print("Number of SGNs (all components)   :", len(remaining_sgns))
+            _, n_per_component = np.unique(remaining_sgns, return_counts=True)
+            print("Number of SGNs (largest component):", max(n_per_component))
+
+
+if __name__ == "__main__":
+    main()

scripts/sgn_stain_predictions/check_segmentation.py

@@ -0,0 +1,35 @@
+import os
+from glob import glob
+
+import imageio.v3 as imageio
+import napari
+
+
+ROOT = "/mnt/vast-nhr/projects/nim00007/data/moser/cochlea-lightsheet/LS_sampleprepcomparison_crops"
+SAVE_ROOT = "/mnt/vast-nhr/projects/nim00007/data/moser/cochlea-lightsheet/LS_sampleprepcomparison_crops/segmentations"
+
+
+def main():
+    files = sorted(glob(os.path.join(ROOT, "**/*.tif")))
+    for ff in files:
+        if "segmentations" in ff:
+            return
+        print("Visualizing", ff)
+        rel_path = os.path.relpath(ff, ROOT)
+        seg_path = os.path.join(SAVE_ROOT, rel_path)
+
+        image = imageio.imread(ff)
+        if os.path.exists(seg_path):
+            seg = imageio.imread(seg_path)
+        else:
+            seg = None
+
+        v = napari.Viewer()
+        v.add_image(image)
+        if seg is not None:
+            v.add_labels(seg)
+        napari.run()
+
+
+if __name__ == "__main__":
+    main()