KYAGABA/MIC_EXT_SELF_VERSION_PRIVATE

README

Files and Structure

Directory Structure

├── metadata (2.3 GB)
│    ├── patient_metadata.csv.gz
│    ├── study_metadata.csv.gz
│    ├── image_metadata.csv.gz
│    ├── question_metadata.csv.gz
│    ├── question_image_metadata.csv.gz
│    ├── answer_metadata.csv.gz
│    ├── answer_image_metadata.csv.gz
│    └── dataset_info.json
├── stats (4.5 GB)
│    └── ...
├── scene_data.zip (1.3 GB)
├── qa.zip (7.5 GB)
├── exports (12.4 GB)
│    └── ...
└── quality_mappings.csv (5 KB)

Metadata (”metadata” dir)

We provide metadata for all scene graphs and question-answer pairs in the metadata directory. The metadata is provided on different levels (patient, study, image, question, question-image, answer, and answer-image) with according number of rows. Each of the metadata files is provided in two redundant versions:

• .csv.gz (compressed csv): for easy interpretation and
• .parquet: for fast reading

These metadata files can be used to filter the dataset on different levels (patient, study, question, …) by different criteria. Therefore, each file comes with unique IDs and additional metadata that may be relevant for that level. An overview is provided below:

Metadata file	1 row per	Index columns	Example metadata	Total # Rows
patient_metadata	patient	patient_id	total_studies (of for this patient), total_study_timespan (interval of study timestamps)	65,317
study_metadata	study	patient_id, study_id	quality (ratings), num_observations (in the scene graph), procedure (from DICOM metadata), timestamp_since_first (relative to first study of patient), timespan_since_prev (relative to previous study)	227,239
image_metadata	image	patient_id, study_id, image_id	view_position (from DICOM metadata), patient_orientation (from DICOM metadata), size, localization_quality	376,175
question_metadata	question	patient_id, study_id, question_id	quality (ratings), question_type, question_strategy, contains_report_answers (are any answers derived from report sentences), contains_template_answers (are any answers based on templates but not directly from sentences), num_answers	42,172,827
question_image_metadata	question-image pair	patient_id, study_id, question_id, image_id	localization_quality	70,045,778
answer_metadata	answer	patient_id, study_id, question_id, answer_id	quality (ratings), answer_type (main answer, details or related information), answer_level (hierarchy level of sub-answers), from_report (whether it was derived from a report sentence)	90,915,200
answer_image_metadata	answer-image pair	patient_id, study_id, question_id, answer_id, image_id	num_regions, localization_quality	151,539,450

Additionally, the dataset_info.json describes the sets of possible values for different tags of answers/observations, i.e. possible finding entity names, region names, finding categories and subcategories, answer types, modifiers, etc.

Statistics (”stats” dir)

We provide additional information and statistics about scene graphs and question-answer pairs in the stats directory. This include aggregate statistics as well as observation-level (for scene graphs) or answer-level (for questions) information. It may for example be used for more advanced data filtering or to compute dataset characteristics without having to load individual scene-graphs or qa-samples (which would be much more expansive).

Aggregate statistics about scene graphs are named as study*.csv and include (among others) the percentages of positive/negative observations for different regions, entities, and categories.

Observation-level information for scene-graphs are named as all_obs*.csv and include (among others) information about positive/negative observations, entities, regions, categories.

(Sub-)answer-level information for qa-samples are named as all_ans*.csv and include (among others) information about positive/negative answers, entities, regions, categories.

Scene Graph Format (”scene_data.zip”)

All scene graphs (and related metadata) can be found in the scene_data.zip file, which contains a folder structure in the following format:

p1x/p1xxxxxxx/sxxxxxxxx.scene_graph.json
p1x/p1xxxxxxx/sxxxxxxxx.metadata.json

where p1x refers to the first 2 digits of the subject_id, p1xxxxxxx to the full subject_id, and sxxxxxxxx to the full study_id.

The sxxxxxxxx.metadata.json file contains study metadata as also provided in the study_metadata.csv.gz file.

The sxxxxxxxx.scene_graph.json file contains the scene graph in the following format:

{
  "patient_id": "p1xxxxxxx",  // see metadata
  "study_id": "sxxxxxxxx",  // see metadata
  // original report sentences (= sentence nodes of scene graph)
  "sentences": {
    "S01": {
      "sent_id": "S01",
      "section": "FINDINGS", 
      "section_type": "FINDINGS",
      "sentence": "No new focal consolidation."
    }, 
    ...  // more sentences
  }, 
  // keys for "observations"
  "top_level_obs_ids": ["O01", "O02", ...],  
  // observations in the study (= observation nodes of scene graph)
  "observations": {
    "O01": {
      "obs_id": "O01",
      "name": "no focal consolidation",
      "summary_sentence": "There is no focal consolidation.", 
      "child_type": null, 
      "child_level": 0, 
      "regions": [{"region": "lungs", "distances": []}], 
      "non_resolved_regions": [], 
      "laterality": "bilateral", 
      "default_regions": ["lungs"],
      "obs_entities": ["consolidation"],
      "obs_entities_parents": [], 
      "non_resolved_obs_entities": [], 
      "obs_categories": ["ANATOMICAL_FINDING", "DISEASE"],
      "obs_subcategories": ["LUNG_FIELD", "PULMONARY_DISEASES", "INFECTION"],
      "probability": "negative", 
      "certainty": "certain", 
      "positiveness": "neg",
      "modifiers": {"temporal": [],"severity": [], "texture": [], "spread": ["focal"]},
      "changes": ["no new"],
      "change_sentence": "No new focal consolidation is visible.", 
      "from_report": true,  // derived from report sentences or template-based?
      "obs_quality": {...},
      "localization": {
          // one item for each image
          "[image_id]": {
              "image_id": "[image_id]",
              "localization_reference_ids": ["lungs"],
              // list of bboxes in (x_1, y_1, x_2, y_2) format in pixel cooridnates
              "bboxes": [[888.0, 370.0, 1610.0, 1642.0 ],   
                         [136.0, 402.0, 898.0, 1678.0  ]],
              "instance_mask_ids": ["lungs"],
              "missing_localization": [],
              "is_fallback": false,  
              "localization_quality": ... 
            },
          },
    },
    ...  // more observations
    },
    // information related to indication section of report
    "indication": {
       "indication_summary": "Female with HIV, experiencing chest pain and dyspnea; should be evaluated for infiltrate and effusion.", 
       "patient_info": "Female, HIV-positive, with chest pain and dyspnea.",
       "indication": "Chest pain and dyspnea.", 
       "evaluation": "Evaluate for infiltrate and effusion.",
       "associated_sentence_ids": ["S05", ...],
       "associated_obs_ids": ["O03", ...],
       "answer_for_indication": {
         // this has the same form as an observation node in "observations"
           "obs_id": "OIND",  // this ID is always the same
           "name": "...",
           ...
         }
    },
    // regions relevant for the study (= region nodes of scene graph)
    "regions": {
    "left lung": {
      "region": "left lung",
      "laterality": "left",
      "localization": { ... }  // same format as for observation nodes
      "region_localization_quality": ...
    },
    ...  // more regions
  },
  // relations between observation and region nodes
  "located_at_relations": [
    {"region": "lungs", "observation_id": "O01", "distances": [], "where_specified": "direct"},
    ... // more relations
  ],
  // relations between observation node pairs
  "obs_relations": [
    {"parent_observation_id": "O02", "child_observation_id":"O02.01", "child_type":"associated_with"},
    ... // more relations
  ],
  // relations between observation and sentence nodes
  "obs_sent_relations": [
    {"observation_id": "O01", "sentence_id": "S01"},
    ... // more relations
  ]
  // relations between region node pairs
    "region_region_relations": [
      {"region": "lungs", "related_region": "left lung", "relation_type": "sub_region"},
      {"region": "left lung", "related_region": "right lung", "relation_type": "right"},
      ... // more relations
    ]
    // quality levels for different aspects (larger = better)
    "study_quality": {
    ...
  },
  // localization quality level per imge-id (larger = better)
  "study_img_localization_quality": {
    ...
  }
}

Question-Answer Format (”qa.zip”)

All question-answer data can be found in the qa.zip file, which contains a folder structure in the following format:

p1x/p1xxxxxxx/sxxxxxxxx.qa.json

where p1x refers to the first 2 digits of the subject_id, p1xxxxxxx to the full subject_id, and sxxxxxxxx to the full study_id.

Each of the sxxxxxxxx.qa.json files contains all question-answer pairs (and additional tags) for a single study in the following format:

{
    "patient_id": "p1xxxxxxx",  // see metadata
  "study_id": "sxxxxxxxx",  // see metadata
    "questions": [
    // -> one object per question-answer pair
    {
        "question_id": "xxxxxxxxxxxx",  // see metadata
        "question_type": "describe_all",  // template used for generation
        "question_strategy": "abnormal",  // strategy used for generation
        "variables": { ... },  // template variables used for generation
        "obs_ids":["O01", ...],  // observations (from scene graph) used in answer
        "contains_report_answers": true/false,  // any answers derived from report sentences?
        "contains_template_answers": true/false,  // any answers based on templates but not directly from sentences?
        "extraction_quality": { ... },
        "question_img_localization_quality": { ... },
        "question": "Describe the given study.",
        // list of sub-answers (top-level answers with their sub-answers)
        "answers":[
          {
            "answer_id": "xxxxxxxxxxxx",  // see metadata
            "answer_type":"main_answer",  // main_answer, details, or related_information
            "answer_level": 0,  // 0 for top-level, >0 for each child-level
            "text": "There is no focal consolidation.",  // this is the answer text
            "name_tag": "No focal consolidation",  // summary name of this sub-answer
            "laterality": "bilateral",
            "regions": ["lungs"],
            "obs_entities": ["consolidation"],
            "obs_entities_parents": [],
            "obs_categories": ["ANATOMICAL_FINDING", "DISEASE"], 
            "obs_subcategories": ["LUNG_FIELD", "PULMONARY_DISEASES", "INFECTION"],
            "certainty": "certain",
            "positiveness": "neg",
            // list of modifiers (tuples of modifier type and value)
            "modifiers": [["spread", "focal"]], 
            "localization": {
              // one item for each image
              "[image_id]": {
                  "image_id": "[image_id]",
                  "localization_reference_ids": ["lungs"],
                  // list of bboxes in (x_1, y_1, x_2, y_2) format in pixel cooridnates
                  "bboxes": [[888.0, 370.0, 1610.0, 1642.0 ],   
                             [136.0, 402.0, 898.0, 1678.0  ]],
                  "instance_mask_ids": ["lungs"],
                  "missing_localization": [],
                  "is_fallback": false,  
                  "localization_quality": ... 
                },
              },
              // contain child-answers if there are any (same object format as top-level answer)
            "sub_answers": [],
            "from_report": true/false,  // derived from report sentences?
            "extraction_quality": {...},
            "answer_quality": {...},
          },
          ... // more top-level answers
        ],
        "question_quality": {...}
  },
  ... // more questions
]} 
            

Exports (”exports” dir)

Here we provide subsets of the full dataset.

We have two full copies of the dataset:

• A_frontal(Fine-tuning grade): Only questions with a quality rating of A, A+, or A++, and only frontal images (7,532,281 QA-pairs). We recommend this dataset for fine-tuning / instruction-tuning purposes.
• B_frontal (Pre-training grade): Only questions with a quality rating of B, A, A+, or A++, and only frontal images (31,230,906 QA-pairs). We recommend this dataset for pre-training purposes. (this is a superset of A_frontal)

Each of these datasets contains the metadata dir, scene_graph.zip and qa.zip.

Additionally, we provide filtered metadata files for further subsets of these. They are provided as sub-folders in the metadata dirs. We provide the following:

• A_frontal/metadata/Ap: Quality rating of A+ or A++ (2,389,739 QA-pairs)
• A_frontal/metadata/App: Quality rating of A++ (1,318,885 QA-pairs)
• A_frontal/metadata/q1M: random 1M question subset with A, A+, or A++ ratings (1M QA-pairs)
• A_frontal/metadata/Ap_q1M: random 1M question subset with A+ or A++ ratings (1M QA-pairs)
• A_frontal/metadata/App_q1M: random 1M question subset with A++ ratings (1M QA-pairs)
• B_frontal/metadata/q1M: random 1M question subset with B, A, A+, or A++ ratings (1M QA-pairs)

Quality mappings (”quality_mappings.csv”)

This file provides mappings from the raw encodings of quality values (fields in the JSON-files or columns in the metadata files) to their respective fields.