Skip to content

Home

Jump to bottom
H Gazula edited this page May 1, 2023 · 7 revisions

Welcome to the Photo-SynthSeg wiki!

Project Github link: here Forked from the original SynthSeg

>>> Branches on the repo:

  • master - Original SynthSeg
  • lesion-SynthSeg - Same as above
  • fixed-spacing - Branch for SynthSeg on Photos (with fixed spacing)
  • variable-spacing - Branch for SynthSeg on Photos (with variable spacing)

NOTE: You will almost always work with the photo-{fixed|variable}-spacing branch.

Path on calico: /space/calico/1/users/Harsha/SynthSeg

Setting up the project:

Please refer to the setup script. TODO

Python Environment:

  • Option 1. Build your own using the requirements file.
  • Option 2. Source my environment- source /space/calico_001/users/Harsha/venvs/synthseg-venv/bin/activate

NOTE (important): Before running this project, please add the project directory to the PYTHONPATH environment variable as follows:

export PYTHONPATH=/path/to/Photo-SynthSeg/in/your/space

Directory Structure: (followed by a description of contents)

image

data - data for Photo-SynthSeg. For more information, please refer to instructions.txt in the same folder.

  • SynthSeg_label_maps_manual_auto_photos - Segmentation maps (whole brain) for training Photo-SynthSeg

  • SynthSeg_label_maps_manual_auto_photos_noCerebellumOrBrainstem - Segmentation maps (whole brain) without Cerebellum and Brainstem

  • SynthSeg_label_maps_manual_auto_photos_noCerebellumOrBrainstem_lh - Segmentation maps (hemi brain) without Cerebellum and Brainstem

  • SynthSeg_param_files_manual_auto_photos - (see below)

  • SynthSeg_param_files_manual_auto_photos_noCerebellumOrBrainstem -

    generation_charm_choroid_lesions_lh.npy - generation labels (i.e., the set of unique labels in the segmentation)

    generation_classes_charm_choroid_lesions_gm_lh.npy - the generation classes (which group some CSF classes, hippocampus with the amygdala, etc.)

    segmentation_new_charm_choroid_lesions_lh.npy - segmentation labels (which ignore some segmentation labels, like the CSF around the skull, which is label 24). So that you know, the number of neutral labels is 7.

    topo_classes_lh.npy - For more information, please refer to ~/Photo-SynthSeg/scripts/tutorials/2-generation-explained.py

  • SynthSeg_param_files_manual_auto_photos_noCerebellumOrBrainstem_lh

  • SynthSeg_test_data -

  • uw_photo - Please refer to photo-reconstruction wiki

ext - External modules used in the project

logs - SLURM logs for different recipes/targets run on the cluster

misc - Miscellaneous scripts written for the project

models - Collection of all Photo-SynthSeg models

  • jei-model - Eugenio trained a fixed spacing (4 mm) model. I'd safely ignore it for now.

  • SynthSeg.h5 - Model from master branch (irrelevant to photos project)

  • models-2022 - Various models trained to segment 3D reconstructed volumes from Photographs

    NOTE: S{spacing}R{run_id} means a model trained with a 'fixed-spacing' of spacing mm. VS{run-id} means 'variable-spacing'. The difference between VS and VS-accordion is- in VS, the spacing within a volume is fixed but randomized between batches, whereas in VS-accordion, the spacing is randomized both within a volume as well as between batches. If the model name ends with an n, it means the n_neutral_labels is set to some integer. no-flip means flipping has been set to False.

    NOTE (very important): Continuing the above description, please use models with an n and no no-flip.

results - Collection of all results

  • 20220328 - Ignore

  • 20220407 - Ignore

  • 20220411 - This is the latest/final set of results used in the paper. Performed automated segmentations on UW reconstructions using the models mentioned above. The script used to generate this result is ~/scripts/hg_dice_scripts/new.py.

  • mgh_inference_20230411 - Performed automated segmentations on MGH reconstructions using final select models. The script used to generate this result is ~/scripts/mgh_inference.py (in the fixed-spacing branch).

  • new-recons - Reconstructions on UW data using updated reconstruction code (aka with interleaved training)

  • old-recons - Same as above but before finalizing the reconstruction script. This folder can be safely ignored.

scripts - Collection of Python scripts used in the project

  • commands - training*, prediction and few other scripts written by Benjamin, *modified by @hvgazula.

  • hg_dice_scripts - Scripts written by @hvgazula for post-processing on reconstructions and segmentations

               1. dice_avg_plots.ipynb - Notebook to generate the boxplots of dice scores in the paper

               2. dice_avg_plots.py - Python script to generate boxplots of dice scores across all models.

               3. dice_calculations.py - script to calculate and store dice scores

               4. dice_config.py - configuration file to enable and postprocessing

               5. dice_gather.py - collect all required files (reconstructions and segmentations) as needed for postprocessing

               6. dice_mri_utils.py - enables registration or conversion of 'space' of one volume to another for dice calculations

               7. dice_plots.py - plotting (as the name suggests)

               8. dice_stats.py

               9. dice_utils.py - common utilities across all other scripts in the current directory.

               10. dice_volumes.py - calculate/store volumes

               11. new.py - postprocessing script (important)

               12. old.py - same as above but can be safely ignored.

               13. uw_config.py - some configuration parameters for postprocessing

  • tutorials - Written by Benjamin. This could be a starting point for understanding the modeling approach behind SynthSeg

submit-pipeline.sh - Please refer to ~/scripts/hg_dice_scripts/new.py or the Makefile for more information

submit-samseg.sh - shell script to run SAMSEG. This was necessitated because SAMSEG was run using Freesurfer (but in a very roundabout way)

submit.sh - SLURM submission script to run Make recipes as needed

Clone this wiki locally