Recently I started working with Glaucoma Images preparation for AI/CV applications. Computer Vision Annotation Tool (CVAT) was deployed locally and fundus images uploaded into an organizational project within CVAT. Each Image is being annotated as follows

• Grading [Tag]
  • Glaucoma Status [SELECT RADIO BUTTON]
    • —– [DEFAULT]
    • Normal
    • Suspect
    • Glaucoma
    • Non-Gradable
    • Other Retinal
  • Grader Remarks [TEXT]
  • Consultant Remarks [TEXT]

 [
  {
    "name": "Grading",
    "id": 60,
    "color": "#97af27",
    "type": "tag",
    "attributes": [
      {
        "id": 215,
        "name": "Glaucoma Status",
        "input_type": "radio",
        "mutable": false,
        "values": [
          "—-",
          "Normal",
          "Suspect",
          "Glaucoma",
          "Not Gradable",
          "Other Retinal"
        ],
        "default_value": "—-"
      },
      {
        "id": 216,
        "name": "Grader Remarks",
        "input_type": "text",
        "mutable": false,
        "values": [
          ""
        ],
        "default_value": ""
      },
      {
        "id": 217,
        "name": "Consultant Remarks",
        "input_type": "text",
        "mutable": false,
        "values": [
          ""
        ],
        "default_value": ""
      }
    ]
  }
]Code language: JSON / JSON with Comments (json)

Exporting the Attrinbute annotations

I thought the best format would be the CVAT for Images format as that supports all attributes. Hence that was chosen. Export can be done at project level [for all tasks in that project] as well as at task level [for only that task and its jobs]

Annotations were therefore exported and saved as a zip file. Within the zip file is an ANNOTATIONS.XML that containes the metadata as well as the data.

How to convert the annotations as a simple CSV file that can be easily shared and analysed in Stata etc

Python to the rescue

It took a half day to figure the following code out

from zipfile import ZipFile
from pprint import pprint
from lxml import etree 
import csv

startFile = 'PROJECT_EXPORT.zip'
with ZipFile(startFile, 'r') as zip:
  zip.printdir()
  zip.extractall()

tree = etree.parse('annotations.xml')
root = tree.getroot()

# METADATA
set_name= root.find('meta/project/name').text
set_created= root.find('meta/project/created').text
set_updated= root.find('meta/project/updated').text
set_dumped= root.find('meta/dumped').text

# opening the csv file in 'w' mode
csvfile = open(startFile+'.csv', 'w', newline ='')
with csvfile:
  fieldnames  = ['set_name','image_name', 'Glaucoma_Status', 'Grader_Remarks', 'Consultant_Remarks',  'set_created', 'set_updated', 'set_dumped']
  writer = csv.DictWriter(csvfile, fieldnames=fieldnames)
  writer.writeheader()

  # Looping over images and getting relevant tag attributes 
  list_images = root.findall(".//image")
  for image in list_images:
    name = image.attrib.get("name")
    Glaucoma_Status = image.find('.//tag/attribute[@name="Glaucoma Status"]').text
    Grader_Remarks = image.find('.//tag/attribute[@name="Grader Remarks"]').text
    Consultant_Remarks = image.find('.//tag/attribute[@name="Consultant Remarks"]').text
    data = {'image_name': name, 'Glaucoma_Status': Glaucoma_Status, 'Grader_Remarks': Grader_Remarks, 'Consultant_Remarks': Consultant_Remarks, 'set_name':set_name, 'set_created': set_created, 'set_updated':set_updated, 'set_dumped': set_dumped}
    print(data['Glaucoma_Status'])
    writer.writerow(data)

Code language: Python (python)

Let us break it down

1. Open the exported ZIP file and Extract its contents
2. Parse the XML using LXML ElementTree and save the root element in a variable
3. Extract meta-data information from the relevant XML nodes <– Change this as you deem fit
4. Open a new CSV dile and name it the same as start file
5. Write the CSV header using dictionary writer
6. Use the findall method on root element to identify all ‘image’ elements
7. Loop over the list of all image elements and extract the three relevant attributes along with image name
8. Create a ‘data’ python dictionary mapping the image level fields as well as the meta fields into the dictionary
9. Write the ‘data’ dictionary to the CSV file

Done !

TBD: Error handling, stripping line breaks from the text fields