From 8614df0f62a503fea1bc938cf0c66903239d9d16 Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?Al=C3=A1n=20Mu=C3=B1oz?= <alan.munoz@ed.ac.uk>
Date: Mon, 26 Sep 2022 18:57:50 +0100
Subject: [PATCH] refactor(extraction): improve cell functions speed
---
src/extraction/core/functions/cell.py | 161 +++++++-------------------
1 file changed, 44 insertions(+), 117 deletions(-)
diff --git a/src/extraction/core/functions/cell.py b/src/extraction/core/functions/cell.py
index 8c4ea97e..5b95541a 100644
--- a/src/extraction/core/functions/cell.py
+++ b/src/extraction/core/functions/cell.py
@@ -3,12 +3,16 @@ Base functions to extract information from a single cell
These functions are automatically read by extractor.py, and so can only have the cell_mask and trap_image as inputs and must return only one value.
"""
+import math
+import typing as t
+
+import bottleneck as bn
+import faiss
import numpy as np
from scipy import ndimage
-from sklearn.cluster import KMeans
-def area(cell_mask):
+def area(cell_mask) -> int:
"""
Find the area of a cell mask
@@ -17,10 +21,10 @@ def area(cell_mask):
cell_mask: 2d array
Segmentation mask for the cell
"""
- return np.sum(cell_mask, dtype=int)
+ return bn.nansum(cell_mask)
-def eccentricity(cell_mask):
+def eccentricity(cell_mask) -> float:
"""
Find the eccentricity using the approximate major and minor axes
@@ -33,7 +37,7 @@ def eccentricity(cell_mask):
return np.sqrt(maj_ax**2 - min_ax**2) / maj_ax
-def mean(cell_mask, trap_image):
+def mean(cell_mask, trap_image) -> float:
"""
Finds the mean of the pixels in the cell.
@@ -43,10 +47,10 @@ def mean(cell_mask, trap_image):
Segmentation mask for the cell
trap_image: 2d array
"""
- return np.mean(trap_image[np.where(cell_mask)], dtype=float)
+ return bn.nanmean(trap_image[cell_mask])
-def median(cell_mask, trap_image):
+def median(cell_mask, trap_image) -> int:
"""
Finds the median of the pixels in the cell.
@@ -56,10 +60,10 @@ def median(cell_mask, trap_image):
Segmentation mask for the cell
trap_image: 2d array
"""
- return np.median(trap_image[np.where(cell_mask)])
+ return bn.nanmedian(trap_image[cell_mask])
-def max2p5pc(cell_mask, trap_image):
+def max2p5pc(cell_mask, trap_image) -> float:
"""
Finds the mean of the brightest 2.5% of pixels in the cell.
@@ -70,18 +74,15 @@ def max2p5pc(cell_mask, trap_image):
trap_image: 2d array
"""
# number of pixels in mask
- npixels = cell_mask.sum()
+ npixels = bn.nansum(cell_mask)
top_pixels = int(np.ceil(npixels * 0.025))
- # sort pixels in cell
- sorted_vals = np.sort(trap_image[np.where(cell_mask)], axis=None)
- # find highest 2.5%
- top_vals = sorted_vals[-top_pixels:]
+ # sort pixels in cell and find highest 2.5%
+ top_values = trap_image[bn.rankdata(trap_image[cell_mask])[:top_pixels]]
# find mean of these highest pixels
- max2p5pc = np.mean(top_vals, dtype=float)
- return max2p5pc
+ return bn.nanmean(top_values)
-def max5px(cell_mask, trap_image):
+def max5px(cell_mask, trap_image) -> float:
"""
Finds the mean of the five brightest pixels in the cell.
@@ -92,63 +93,13 @@ def max5px(cell_mask, trap_image):
trap_image: 2d array
"""
# sort pixels in cell
- sorted_vals = np.sort(trap_image[np.where(cell_mask)], axis=None)
- top_vals = sorted_vals[-5:]
+ pixels = trap_image[cell_mask]
+ top_values = pixels[bn.rankdata(pixels)[:5].astype(int) - 1]
# find mean of five brightest pixels
- max5px = np.mean(top_vals, dtype=float)
+ max5px = bn.nanmean(top_values)
return max5px
-def max5px_med(cell_mask, trap_image):
- """
- Finds the mean of the five brightest pixels in the cell divided by the median pixel value.
-
- Parameters
- ----------
- cell_mask: 2d array
- Segmentation mask for the cell
- trap_image: 2d array
- """
- # sort pixels in cell
- sorted_vals = np.sort(trap_image[np.where(cell_mask)], axis=None)
- top_vals = sorted_vals[-5:]
- # find mean of five brightest pixels
- max5px = np.mean(top_vals, dtype=float)
- # find the median
- med = np.median(sorted_vals)
- if med == 0:
- return np.nan
- else:
- return max5px / med
-
-
-def max2p5pc_med(cell_mask, trap_image):
- """
- Finds the mean of the brightest 2.5% of pixels in the cell
- divided by the median pixel value.
-
- Parameters
- ----------
- cell_mask: 2d array
- Segmentation mask for the cell
- trap_image: 2d array
- """
- # number of pixels in mask
- npixels = cell_mask.sum()
- top_pixels = int(np.ceil(npixels * 0.025))
- # sort pixels in cell
- sorted_vals = np.sort(trap_image[np.where(cell_mask)], axis=None)
- # find highest 2.5%
- top_vals = sorted_vals[-top_pixels:]
- # find mean of these highest pixels
- max2p5pc = np.mean(top_vals, dtype=float)
- med = np.median(sorted_vals)
- if med == 0:
- return np.nan
- else:
- return max2p5pc / med
-
-
def std(cell_mask, trap_image):
"""
Finds the standard deviation of the values of the pixels in the cell.
@@ -159,7 +110,7 @@ def std(cell_mask, trap_image):
Segmentation mask for the cell
trap_image: 2d array
"""
- return np.std(trap_image[np.where(cell_mask)], dtype=float)
+ return bn.std(trap_image[cell_mask])
def k2_major_median(cell_mask, trap_image):
@@ -177,47 +128,23 @@ def k2_major_median(cell_mask, trap_image):
median: float
The median of the major cluster of two clusters
"""
- if np.any(cell_mask):
- X = trap_image[np.where(cell_mask)].reshape(-1, 1)
- # cluster pixels in cell into two clusters
- kmeans = KMeans(n_clusters=2, random_state=0).fit(X)
- high_clust_id = kmeans.cluster_centers_.argmax()
- # find the median of pixels in the largest cluster
- major_cluster = X[kmeans.predict(X) == high_clust_id]
- major_median = np.median(major_cluster, axis=None)
- return major_median
- else:
- return np.nan
-
-def k2_minor_median(cell_mask, trap_image):
- """
- Finds the median of the minor cluster after clustering the pixels in the cell into two clusters.
+ if bn.anynan(trap_image):
+ cell_mask[np.isnan(trap_image)] = False
+ X = trap_image[cell_mask].reshape(-1, 1).astype(np.float32)
+ # cluster pixels in cell into two clusters
+ indices = faiss.IndexFlatL2(X.shape[1])
+ # (n_clusters=2, random_state=0).fit(X)
+ _, indices = indices.search(X, k=2)
+ high_indices = bn.nanargmax(indices, axis=1).astype(bool)
+ # find the median of pixels in the largest cluster
+ # high_masks = np.logical_xor( # Use casting to obtain masks
+ # high_indices.reshape(-1, 1), np.tile((0, 1), X.shape[0]).reshape(-1, 2)
+ # )
+ major_median = bn.nanmedian(X[high_indices])
+ return major_median
- Parameters
- ----------
- cell_mask: 2d array
- Segmentation mask for the cell
- trap_image: 2d array
- Returns
- -------
- median: float
- The median of the minor cluster.
- """
- if np.any(cell_mask):
- X = trap_image[np.where(cell_mask)].reshape(-1, 1)
- # cluster pixels in cell into two clusters
- kmeans = KMeans(n_clusters=2, random_state=0).fit(X)
- low_clust_id = kmeans.cluster_centers_.argmin()
- # find the median of pixels in the smallest cluster
- minor_cluster = X[kmeans.predict(X) == low_clust_id]
- minor_median = np.median(minor_cluster, axis=None)
- return minor_median
- else:
- return np.nan
-
-
-def volume(cell_mask):
+def volume(cell_mask) -> float:
"""
Estimates the volume of the cell assuming it is an ellipsoid with the mask providing a cross-section through the median plane of the ellipsoid.
@@ -247,20 +174,20 @@ def conical_volume(cell_mask):
def spherical_volume(cell_mask):
- '''
+ """
Estimates the volume of the cell assuming it is a sphere with the mask providing a cross-section through the median plane of the sphere.
Parameters
----------
cell_mask: 2d array
Segmentation mask for the cell
- '''
- area = cell_mask.sum()
- r = np.sqrt(area / np.pi)
+ """
+ total_area = area(cell_mask)
+ r = math.sqrt(total_area / np.pi)
return (4 * np.pi * r**3) / 3
-def min_maj_approximation(cell_mask):
+def min_maj_approximation(cell_mask) -> t.Tuple[int]:
"""
Finds the lengths of the minor and major axes of an ellipse from a cell mask.
@@ -278,8 +205,8 @@ def min_maj_approximation(cell_mask):
# get the size of the top of the cone (points that are equally maximal)
cone_top = ndimage.morphology.distance_transform_edt(dn == 0) * padded
# minor axis = largest distance from the edge of the ellipse
- min_ax = np.round(nn.max())
+ min_ax = np.round(bn.nanmax(nn))
# major axis = largest distance from the cone top
# + distance from the center of cone top to edge of cone top
- maj_ax = np.round(dn.max() + cone_top.sum() / 2)
+ maj_ax = np.round(bn.nanmax(dn) + bn.nansum(cone_top) / 2)
return min_ax, maj_ax
--
GitLab