added training for DAPI

.gitignore

@@ -23,6 +23,9 @@
 # RStudio files
 .Rproj.user/
 
+/training_data/dapi/images/*.tif
+/training_data/dapi/masks/*.tif
+
 # produced vignettes
 vignettes/*.html
 vignettes/*.pdf

augment.py

@@ -0,0 +1,52 @@
+import os
+import cv2
+
+# Set the paths for the input folders
+mask_org_folder = "./training_data/masks_nuclei"
+images_org_folder = "./training_data/images_nuclei"
+
+# Set the paths for the output folders
+mask_folder = "./training_data/dapi/masks"
+images_folder = "./training_data/dapi/images"
+
+# Create the output folders if they don't exist
+os.makedirs(mask_folder, exist_ok=True)
+os.makedirs(images_folder, exist_ok=True)
+
+# Set the ROI size and step
+roi_size = (256, 256)
+roi_step = 128
+
+# Get the list of file names in the mask_org folder
+file_names = [file for file in os.listdir(mask_org_folder) if file.lower().endswith(('.tif', '.tiff'))]
+
+# Iterate over the file names
+for file_name in file_names:
+    # Read the mask image
+    mask_path = os.path.join(mask_org_folder, file_name)
+    mask = cv2.imread(mask_path, cv2.IMREAD_GRAYSCALE)
+
+    # Read the corresponding input image
+    image_path = os.path.join(images_org_folder, file_name)
+    image = cv2.imread(image_path, cv2.IMREAD_GRAYSCALE | cv2.IMREAD_ANYDEPTH)
+
+    # Get the dimensions of the images
+    mask_height, mask_width = mask.shape[:2]
+    image_height, image_width = image.shape[:2]
+
+    # Iterate over the ROI positions
+    for y in range(0, mask_height - roi_size[0] + 1, roi_step):
+        for x in range(0, mask_width - roi_size[1] + 1, roi_step):
+            # Extract the ROI from the mask
+            roi_mask = mask[y:y+roi_size[0], x:x+roi_size[1]]
+
+            # Extract the ROI from the image
+            roi_image = image[y:y+roi_size[0], x:x+roi_size[1]]
+
+            # Save the ROI as a new image in the mask folder
+            new_mask_path = os.path.join(mask_folder, f"{file_name}_{y}_{x}.tif")
+            cv2.imwrite(new_mask_path, roi_mask)
+
+            # Save the ROI as a new image in the images folder
+            new_image_path = os.path.join(images_folder, f"{file_name}_{y}_{x}.tif")
+            cv2.imwrite(new_image_path, roi_image)

figure01.md

@@ -2,6 +2,11 @@
 Daniel Fürth <br><br>Table of Contents:
 
 - [Emission plot Fig. 1c](#emission-plot-fig.-1c)
+- [DAPI nuclei cell segmentation in
+  2D.](#dapi-nuclei-cell-segmentation-in-2d.)
+  - [Installation](#installation)
+  - [Segmentation training](#segmentation-training)
+  - [Prediction](#prediction)
 
 ## Emission plot Fig. 1c
 
@@ -152,3 +157,110 @@ round(AZdye594, 2)
 ```
 
     [1] 2.88
+
+## DAPI nuclei cell segmentation in 2D.
+
+### Installation
+
+Make sure you have `conda` installed. Create a new conda environment:
+
+```
+conda create --name cellseg2D-env
+conda activate cellseg2D-env
+conda install -c conda-forge napari   
+conda install opencv
+```
+
+Install Tensorflow for macOS M1/M2:
+
+```
+pip install tensorflow-macos
+pip install tensorflow-metal
+```
+
+Install stardist for cell nuclei segmentation:
+
+```
+pip install gputools
+pip install stardist
+pip install csbdeep
+```
+
+### Segmentation training
+
+#### Augment training data set
+
+```
+python augment.py
+```
+
+This expands images_org and masks_org into images (input) and masks
+(ground truth). Input and ground truth are matched based on file name.
+Format is 8-bit monochrome TIF on both.
+
+If more than 255 cells needs to be segmented within a single image you
+can simply change mask format to 16-bit.
+
+#### Perform training
+
+```
+python train_nuclei.py
+```
+
+Open up tensorboard to follow the results:
+
+```
+tensorboard --logdir=.
+```
+
+![Tensorboard lets you monitor the training of the neural
+network](./repo_img/tensorboard.png)
+
+Click the **images** tab of the Tensorboard to inspect the visual output
+of the training. This is in the beginning: ![inspect the
+output](./repo_img/tensorboard_img.png)
+
+Here:
+
+- `net_input` is the input images. Notice a cell is only really present
+  in the first out of three.
+- `net_output0` is the current output from the network.
+- `net_target0` is the ground truth (what the network ideally should
+  have generated).
+
+### Prediction
+
+We have a script we can apply to any image for prediction.
+
+```
+python predict_nuclei.py 
+```
+
+<div id="fig-dapi">
+
+<table>
+<colgroup>
+<col style="width: 50%" />
+<col style="width: 50%" />
+</colgroup>
+<tbody>
+<tr class="odd">
+<td style="text-align: center;"><div width="50.0%"
+data-layout-align="center">
+<p><img src="./repo_img/example_image.jpg" id="fig-input"
+data-ref-parent="fig-dapi" data-fig.extended="false"
+alt="(a) input" /></p>
+</div></td>
+<td style="text-align: center;"><div width="50.0%"
+data-layout-align="center">
+<p><img src="./repo_img/example_labels.jpg" id="fig-output"
+data-ref-parent="fig-dapi" data-fig.extended="false"
+alt="(b) output" /></p>
+</div></td>
+</tr>
+</tbody>
+</table>
+
+Figure 1: Segmentation results.
+
+</div>

figure01.qmd

@@ -118,3 +118,84 @@ AZdye594<-quench.ratio$`max(emission)`[3]/quench.ratio$`max(emission)`[4]
 round(AZdye488, 2)
 round(AZdye594, 2)
 ```
+
+## DAPI nuclei cell segmentation in 2D.
+
+### Installation
+
+Make sure you have `conda` installed.
+Create a new conda environment:
+```{eval=FALSE}
+conda create --name cellseg2D-env
+conda activate cellseg2D-env
+conda install -c conda-forge napari   
+conda install opencv
+```
+
+Install Tensorflow for macOS M1/M2:
+```{eval=FALSE}
+pip install tensorflow-macos
+pip install tensorflow-metal
+```
+
+Install stardist for cell nuclei segmentation:
+```{eval=FALSE}
+pip install gputools
+pip install stardist
+pip install csbdeep
+```
+
+### Segmentation training
+
+#### Augment training data set
+
+```{eval=FALSE}
+python augment.py
+```
+
+This expands images_org and masks_org into images (input) and masks (ground truth). Input and ground truth are matched based on file name. Format is 8-bit monochrome TIF on both.
+
+If more than 255 cells needs to be segmented within a single image you can simply change mask format to 16-bit.
+
+#### Perform training
+
+```{eval=FALSE}
+python train_nuclei.py
+```
+
+Open up tensorboard to follow the results:
+```{eval=FALSE}
+tensorboard --logdir=.
+```
+
+
+
+![Tensorboard lets you monitor the training of the neural network](./repo_img/tensorboard.png)
+
+
+Click the **images** tab of the Tensorboard to inspect the visual output of the training.
+This is in the beginning:
+![inspect the output](./repo_img/tensorboard_img.png)
+
+Here:
+
+- `net_input` is the input images. Notice a cell is only really present in the first out of three.
+- `net_output0` is the current output from the network.
+- `net_target0` is the ground truth (what the network ideally should have generated).
+
+### Prediction
+
+We have a script we can apply to any image for prediction.
+
+```{eval=FALSE}
+python predict_nuclei.py 
+```
+
+::: {#fig-dapi layout-ncol=2}
+
+![input](./repo_img/example_image.jpg){#fig-input}
+
+![output](./repo_img/example_labels.jpg){#fig-output}
+
+Segmentation results.
+:::

models/dapi/config.json

@@ -0,0 +1 @@
+{"n_dim": 2, "axes": "YXC", "n_channel_in": 1, "n_channel_out": 33, "train_checkpoint": "weights_best.h5", "train_checkpoint_last": "weights_last.h5", "train_checkpoint_epoch": "weights_now.h5", "n_rays": 32, "grid": [2, 2], "backbone": "unet", "n_classes": null, "unet_n_depth": 3, "unet_kernel_size": [3, 3], "unet_n_filter_base": 32, "unet_n_conv_per_depth": 2, "unet_pool": [2, 2], "unet_activation": "relu", "unet_last_activation": "relu", "unet_batch_norm": false, "unet_dropout": 0.0, "unet_prefix": "", "net_conv_after_unet": 128, "net_input_shape": [null, null, 1], "net_mask_shape": [null, null, 1], "train_shape_completion": false, "train_completion_crop": 32, "train_patch_size": [256, 256], "train_background_reg": 0.0001, "train_foreground_only": 0.9, "train_sample_cache": true, "train_dist_loss": "mae", "train_loss_weights": [1, 0.2], "train_class_weights": [1, 1], "train_epochs": 400, "train_steps_per_epoch": 100, "train_learning_rate": 0.0003, "train_batch_size": 4, "train_n_val_patches": null, "train_tensorboard": true, "train_reduce_lr": {"factor": 0.5, "patience": 40, "min_delta": 0}, "use_gpu": false}

models/dapi/thresholds.json

@@ -0,0 +1 @@
+{"prob": 0.46905075238803473, "nms": 0.3}

predict_nuclei.py

@@ -0,0 +1,35 @@
+from __future__ import print_function, unicode_literals, absolute_import, division
+import sys
+import numpy as np
+import matplotlib
+matplotlib.rcParams["image.interpolation"] = 'none'
+import matplotlib.pyplot as plt
+
+from glob import glob
+from tifffile import imread
+from csbdeep.utils import Path, normalize
+from csbdeep.io import save_tiff_imagej_compatible
+
+from stardist import random_label_cmap, _draw_polygons, export_imagej_rois
+from stardist.models import StarDist2D
+
+np.random.seed(6)
+lbl_cmap = random_label_cmap()
+
+X = sorted(glob('./training_data/images_nuclei/*.tif'))
+X = list(map(imread,X))
+
+n_channel = 1 if X[0].ndim == 2 else X[0].shape[-1]
+axis_norm = (0,1)   # normalize channels independently
+# axis_norm = (0,1,2) # normalize channels jointly
+if n_channel > 1:
+    print("Normalizing image channels %s." % ('jointly' if axis_norm is None or 2 in axis_norm else 'independently'))
+
+model = StarDist2D(None, name='dapi', basedir='models')
+
+img = normalize(X[0], 1,99.8, axis=axis_norm)
+labels, details = model.predict_instances(img)
+
+save_tiff_imagej_compatible('example_image.tif', img, axes='YX')
+save_tiff_imagej_compatible('example_labels.tif', labels, axes='YX')
+export_imagej_rois('example_rois.zip', details['coord'])