DL Segmentation on FIB-SEM Datasets

This repository hosts code for performing deep-learning based segmentation in the Washington University Center for Cellular Imaging.

This new repository is prepared for WU RIS HPC.

Information

• By Michael Chien-Cheng Shih, PhD, Staff Scientist
• email: [email protected]
• Date: July 14, 2020
• GitHub Repository: https://github.com/WUCCI-WashU

Functions

• Deep Learning
  1. 01_img_prep_t2_demo.ipynb: data preprocessing
  2. 02_train_model_t2_nucleus_demo.ipynb: model training
  3. 02_train_model_t2_nucleus_demo_smUnet.ipynb: moodel training with segmentation_models (https://github.com/qubvel/segmentation_models)
  4. 03_prediction_demo.ipynb: segmentation prediction
  5. data_inference_*.ipynb: export statistics
  6. skeleton_study.ipynb: complied data from skeleton/branching analysis

DL Segmentation

• Segmentation has been done in two way:
  1. WUCCI-DL
  2. Ariadne (https://ariadne.ai/)

Methods

• Convolutional Neural Network
• U-Net

Datasets

• Data: Generated from Zeiss Crossbeam 540 FIB-SEM

  • 3D dataset
  • Voxel-size: 10 * 10 * 10 nm³

• Sample: Human hepatic cells

• Labels region:

  • cell Membrane
  • nucleus
  • nucleolus
  • lipid droplets

• Sample

  1. Day 0
     • data_d00_batch01_loc01 (TH-4891-0dayDOX-B-M3)
     • data_d00_batch02_loc02 (TH-4891 0day DOX D-K8)
     • data_d00_batch02_loc03 (TH-4891 0day DOX D A12)
  2. Day 7
     • data_d07_batch01_loc01 (TH-4891-7dayDOX)
     • data_d07_batch02_loc01 (TH-4891 7day DOX B-L8)
     • data_d07_batch02_loc02 (TH-4891 7day DOX C-R19)
  3. Day 14
     • data_d14_batch01_loc01 (TH-4891-14dayDOX-C-H16)
  4. Day 17
     • data_d17_batch01_loc01 (TH-4891-17dayDOX)
  5. Day 21
     • data_d21_batch01_loc01 (TH-4891-21dayDOX)

Directory for Deep Learning

• Data Location: Workstation-008
• Export Directory: D:\PerlmutterData\
  1. raw_input_image: original images
  2. ongoing: Amira projects for manual segmentation
  3. logs: model and record for deep learning training
  4. dl_seg_project_raw
     • data_Amira: the training label iamges and raw images
     • data_crop: small image patches create for training from data_Amira
     • pred_img: the results of prediction
  5. Archive: the archived folders

Directory for Data Export

• Data Location: Workstation-008
• Export Directory: D:\PerlmutterData\segmentation_compiled_export

  1. Amira_3D: Raw Segementation files and Amira .hx files

     • Raw files
       • raw_img: raw image files; pixel size 10 nm^3
       • cell_membrane: segmentation of area inside cell membrane (WUCCI-DL)
       • nucleus: segmentation of nucleus (WUCCI-DL)
       • nucleolus: segmentation of nucleolus (Amira)
       • lipid_droplet: segmentation of lipid droplets (WUCCI-DL)
       • mito: segmentation of mitochondria (Ariadne)
       • cristae: segmentation of cristae inside mitochondria (Ariadne)
       • ER: segmentation of endoplasmic reticulum (Ariadne)
       • inclusion: segmentation of globulus inclusion (Ariadne)
     • Amira .hx file and folders
       • data_d**_batch**_loc**_inclusion: label analysis on ER and inclusion
       • data_d**_batch**_loc**_mito: label analysis on mito and cristae
       • data_d**_batch**_loc**_skeleton: skeleton (branching) analysis on mitochondira
     • .hxtemplate: template files used for analysis

  2. data: data

     • raw: raw analysis exported from Amira
       • label analysis
         • cell_membrane
         • nucleus
         • mito
         • cristae
         • ER
         • inclusion
       • skeleton/branching analysis
         • skeleton: .xml files; includes node, points and segments
         • skeleton_ouput: .csv files; includes node, points, segments and segments_s
       • compile: complied data in single spreadsheet
         • label analysis: Volume3d, Area3d, BaryCenterX, BaryCenterY, BaryCenterZ, index, filename, day
         • skeleton/branching analysis:
           • nodes.csv: Node ID, X Coord, Y Coord, Z Coord, Coordination Number, filename, day
           • points.csv: Point ID, thickness, X Coord, Y Coord, Z Coord, filename, day
           • segments_s: Segment ID, Node ID #1, Node ID #2, Point IDs, length, thickness, filename, day
           • please visit Amira User's Guid for more information.

  3. input: all segmentation images prepare for one label image

  4. output_2: merged label images

  5. output_AND: label images inside cytoplasm

  6. surface: exported surface files in .stl

  7. video: 3D animation generated from Blender

  8. documents

     • data_foldername - filename_with_label.csv: coded filemame and origonal filname.

License

This work is licensed under a Creative Commons Attribution 4.0 International License.