File size: 4,198 Bytes
1e315b6 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 |
import numpy as np
import torch
from torch.nn import functional as F
import os
import scanpy as sc
import json
import cv2
def annotate_with_bulk(img_features, bulk_features, normalize=True, T=1, tensor=False):
"""
Annotates tissue image with similarity scores between image features and bulk RNA-seq features.
:param img_features: Feature matrix representing histopathology image features.
:param bulk_features: Feature vector representing bulk RNA-seq features.
:param normalize: Whether to normalize similarity scores, default=True.
:param T: Temperature parameter to control the sharpness of the softmax distribution. Higher values result in a smoother distribution.
:param tensor: Feature format in torch tensor or not, default=False.
:return: An array or tensor containing the normalized similarity scores.
"""
if tensor:
# Compute similarity between image features and bulk RNA-seq features
cosine_similarity = torch.nn.CosineSimilarity(dim=1, eps=1e-6)
similarity = cosine_similarity(img_features, bulk_features.unsqueeze(0)) # Shape: [n]
# Optional normalization using the feature vector's norm
if normalize:
normalization_factor = torch.sqrt(torch.tensor([bulk_features.shape[0]], dtype=torch.float)) # sqrt(768)
similarity = similarity / normalization_factor
# Reshape and apply temperature scaling for softmax
similarity = similarity.unsqueeze(0) # Shape: [1, n]
similarity = similarity / T # Control distribution sharpness
# Convert similarity scores to probability distribution using softmax
similarity = torch.nn.functional.softmax(similarity, dim=-1) # Shape: [1, n]
else:
# Compute similarity for non-tensor mode
similarity = np.dot(img_features.T, bulk_features)
# Apply a softmax-like normalization for numerical stability
max_similarity = np.max(similarity) # Maximum value for stability
similarity = np.exp(similarity - max_similarity) / np.sum(np.exp(similarity - max_similarity))
# Normalize similarity scores to [0, 1] range for interpretation
similarity = (similarity - np.min(similarity)) / (np.max(similarity) - np.min(similarity))
return similarity
def annotate_with_marker_genes(classes, image_embeddings, all_text_embeddings):
"""
Annotates tissue image with similarity scores between image features and marker gene features.
:param classes: A list or array of tissue type labels.
:param image_embeddings: A numpy array or torch tensor of image embeddings (shape: [n_images, embedding_dim]).
:param all_text_embeddings: A numpy array or torch tensor of text embeddings of the marker genes
(shape: [n_classes, embedding_dim]).
:return:
- dot_similarity: The matrix of dot product similarities between image embeddings and text embeddings.
- pred_class: The predicted tissue type for the image based on the highest similarity score.
"""
# Calculate dot product similarity between image embeddings and text embeddings
# This results in a similarity matrix of shape [n_images, n_classes]
dot_similarity = image_embeddings @ all_text_embeddings.T
# Find the class with the highest similarity for each image
# Use argmax to identify the index of the highest similarity score
pred_class = classes[dot_similarity.argmax()]
return dot_similarity, pred_class
def load_image_annotation(image_path):
"""
Loads an image with annotation.
:param image_path: The file path to the image.
:return: The processed image, converted to BGR color space and of type uint8.
"""
# Load the image from the specified file path using OpenCV
image = cv2.imread(image_path)
# Convert the color from RGB (OpenCV loads as BGR by default) to BGR (which matches common color standards)
image = cv2.cvtColor(image, cv2.COLOR_RGB2BGR)
# Ensure the image is of type uint8 for proper handling in OpenCV and other image processing libraries
image = image.astype(np.uint8)
return image
|