Segmentation_code

System Requirements

OS Requirements

The segmentation code has been tested on the following systems and Python version:

• macOS: Monterey (12.6.8)
• Linux: Ubuntu 16.04
• Python 3.9.16

Python dependencies and installation

keras==2.8.0
matplotlib==3.3.2
numpy==1.26.4
opencv_python==4.5.5.64
Pillow==10.0.0
scikit_learn==1.5.0
scipy==1.13.1
scikit-image==0.22.0
tensorflow==2.8.0
protobuf == 3.20.1

Install dependencies

• Create virtual environment conda create -n cell_seg python=3.9.16
• Activate virtual environment conda activate cell_seg
• Dependencies can be installed by pip install -r requirements.txt

Demo

Training and test dataset

• unzip data.zip
• dataset:

dataset	train	test
#	280	70

U-Net need fixed input size. So we rescale images and masks to 256x256.

• data/image_train_256: train images rescaled to 256x256
• data/mask_train_256: train masks rescaled to 256x256
• data/image_test_256: test images rescaled to 256x256
• data/mask_test_256: test masks rescaled to 256x256

Download pre-trained model

• Download model Model_GESU_oct16.hdf5 from https://www.dropbox.com/scl/fi/1twwtxwpksqy43x25739k/Model_GESU_oct16.hdf5?rlkey=dybz3tun4etuayv1q994gudpg&dl=0

Run train and test script

Edit and run train_test_demo.ipynb:

• set data and model path:
  • train_path, train_label: local training images and masks.
  • test_path, test_label: local test images and masks.
  • model_path : load pre-trained model path.
• training part:

>>> GESU_net = myGESUnet(img_rows = 256, img_cols= 256, train_path=train_path, train_label=train_label, test_path=test_path, test_label=test_label)
>>> GESU_net.load_data()
>>> GESU_net.train(epochs, batches, model_path)

If you use the pre-trained model, just commented out GESU_net.train(epochs, batches, model_path) and then run test part

• test part:

>>> model = myGESUnet(img_rows = 256, img_cols= 256, train_path=train_path, train_label=train_label, test_path=test_path, test_label=test_label)
imgs_train, imgs_mask_train, imgs_test = model.load_data()
>>> model.load_weights(os.path.join(model_path, "Model_GESU_oct16.hdf5"))
>>> imgs_mask_test = model.predict(imgs_test[:,:,:,0], batch_size=1, verbose=1)
>>> np.save('imgs_mask_test.npy', imgs_mask_test)

Do segmentation on the whole single images

• We extract 6821 single astrocyte images from the detection result. According to their regions ac, dm, lat, m, pc and conditions control, relapse, withdraw, we labeled the subfolder names by region + condition. Here we summarize the number of each class:

condition\region	ac	dm	lat	m	pc	total (condition)
control	326	197	197	281	735	2696
withdraw	197	908	908	908	500	2041
relapse	181	759	759	759	488	2084
total (region)	704	2699	969	969	1723	6821 (total)

• Using the inference part in train_test_demo.ipynb, and the pre-trained model Model_GESU_oct16.hdf5, we can segment the whole 6821 single astrocyte images.
• Segmented images stored in segmentation_mask_data/single_mask.zip. After resizing to original size, and place patches to the original location, we generate the entire masks, which stored in segmentation_mask_data/entire_mask.zip.

Segmentation examples

Preprocessing of dataset

• To get 6821 single astrocyte images, we need extract single astrocyte images from 211 fluorescence microscope images with the bounding box information, included in file astrocyte_json.json.
• 211 fluorescence microscope images and astrocyte_json.json stored in https://www.dropbox.com/scl/fo/g7yrbf368luf7vbtu2v8h/ADZRH1FVewFH-hlZc5Air5U?rlkey=qk4qp9nqkb2tz54m9v9b8vl1s&st=ipt8ywjp&dl=0