meghalithic / deepbryo_micro Goto Github PK

Image MHR.8679 as an example. Screenshot1 is the output with default settings when flipped 180 degrees so the scale-bar is at the top (it was upside down in Microporella.jpeg folder on desktop). DeepBryro outputs 3 ovicells.
Upside down:

DeepBryo outputs 0.98 confidence for detection of a single autozooid with two opercula, but it is actually two autozooids. It would be good to be able to program it to know that a single autozooid cannot have more than one operculum. Maybe this can be done in data processing afterwards?

Parts of zooids along the edges of images are often detected with a straight edge despite 5 padding. These might be removable using the circularity output. Eg MHR.4871

Some ascopores are overestimated because of erosion pattern. I have not moved images from omit folders when that is the case. Eg. MHR.5022.jpg ascopore masks are the size of avicularia. Below is the DeepBryo app view side by side with the original image:

Using MHR.4625 as an example. This is the output when confidence is the default 0.50, autofiltering is off (strictness is default 0.50):

following error:

NameError: name 'json' is not defined
Traceback:

File "/home/voje-lab/miniconda3/envs/deepbryo/lib/python3.7/site-packages/streamlit/runtime/scriptrunner/script_runner.py", line 552, in _run_script
    exec(code, module.__dict__)
File "/home/voje-lab/Documents/GitHub/bryozoa/DeepBryo_micro/app/app.py", line 436, in <module>
    main()
File "/home/voje-lab/Documents/GitHub/bryozoa/DeepBryo_micro/app/app.py", line 414, in main
    uploaded_img.name,
File "/home/voje-lab/Documents/GitHub/bryozoa/DeepBryo_micro/app/app.py", line 209, in summarize
    ) = mask_stats(masks[num][idx])
File "/home/voje-lab/Documents/GitHub/bryozoa/DeepBryo_micro/app/app.py", line 160, in mask_stats
    cnt_json = json.dumps(cnt_list)

• figure out padding to exclude zooids not completely in the image
  python app/app-cli.py -i ~/Desktop/microporella-jpg/IMAGE -o ~/Documents/GitHub/bryozoa/DeepBryo_micro -p [left, top, right, bottom]
• figure out scaling
  python app/app-cli.py -i ~/Desktop/microporella-jpg/IMAGE -o ~/Documents/GitHub/bryozoa/DeepBryo_micro -sc SCALE

Most irregular zooids detected by DeepBryo with an avicularia as their only feature are not correct. But image MHR.6529 zooid nr 12 (also MHR.6542, and MHR.6536) first times I've seen an avicularia alone that is correct.

Rocks are detected as ovicells in eg. MHR.5822

MHR.5822

• process non-50 mag separate
• always have same padding around scale
  • run rotated pictures separately

• Turned all images from "keep" folders to the optimal orientation for processing

• Sorted all images into folders according to orientation of scalebar for all four species. Subfolder names are in the following pattern: "M.intermedia-scalebar-left", located within their respective species folders (eg. ~/Desktop/microporella-subset-jpg/M_intermedia/M.intermedia-scalebar-left).

• Finished lists of all images where other species are present (and detected by DeepBryo), for M. discors, M. agonistes and M. speculum. These are named "Speculum_images_with_other_species.csv" etc, located in ~/Documents/GitHub/bryozoa/microporella/Data.
  In each file, column 1 is image_id. Column two is blank when image has another bryozoa species detected by DeepBryo. If the presence of other bryozoa species were not immediately obvious to the viewer I have used column two to note their position, eg. "Other species upper left". I also used column two to note images with cracks that affect measurements, or when DeepBryo detects non-bryozoa taxa as zooids with features.

• Started list of images with other species present for M. intermedia. Finished images in "M.intermedia-scalebar-left"

To do:

• Finished list of all M. intermedia images with other species present.

• CLI-csv files are still present in the M.intermedia subfolders, and should be deleted or moved

• Manually done csvs moved to their correct place on GitHub (~/Documents/GitHub/bryozoa/microporella/Data/Manual-outputs/M.intermedia)

• Redo high throughput with ext version of CLI

Holes, pebbles and other non-bryozoa are often detected as zooids with orifices, or just "disembodied" orifices. These generally have greater than 0.85 circularity (in M. speculum at least) and can be filtered out by that measure. See MHR.4858 for example: