Skip to content

Latest commit

 

History

History
289 lines (231 loc) · 10.8 KB

dataset_schema.md

File metadata and controls

289 lines (231 loc) · 10.8 KB
Raw

SCIN Dataset Description

What follows are a description of the fields in the different CSV files.

scin_cases.csv

  • case_id (string, always present): Identifier for the case

    • This id should typically be stable across releases

  • source (string, always present): Value is "SCIN"

  • release (string, always present): Identifier for the release this case is included from.

    • The format of this string will be "major.minor.patch"

      • Major: Large number of new cases
      • Minor: Providing a small number of new categories or labels for existing images
      • Patch: Bug fixes

    • The first version will be "1.0.0"

  • year (int, always present): Year closest to the bulk of the data released

  • age_group (enum, sometimes present): The age provided by the user is an integer.

    • This value is mapped into the following age ranges:
      • AGE_18_TO_29
      • AGE_30_TO_39
      • AGE_40_TO_49
      • AGE_50_TO_59
      • AGE_60_TO_69
      • AGE_70_TO_79
      • AGE_80_OR_ABOVE

  • sex_at_birth (enum, sometimes present): User-reported sex at birth

    • Values in the dataset:

      • FEMALE
      • MALE
      • OTHER_OR_UNSPECIFIED

    • The SCIN UX also included answers for "OTHER" and "PREFER_NOT_TO_SAY". There were not enough answers in those categories to publish so this field is grouped into the "OTHER_OR_UNSPECIFIED" category for those cases.

  • fitzpatrick_skin_type (enum, sometimes present): User-reported skin type

    • Possible values: FST1-6, NONE_SELECTED
    • Note that in application, users were asked questions about how their skin responds to sun exposure and those responses were mapped to these FST values.
      • NEVER_TANS
      • TANS_MINIMALLY
      • TANS_UNIFORMLY
      • ALWAYS_TANS_WELL
      • TANS_EASILY
      • ALWAYS_TANS
      • NONE_OF_THE_ABOVE
    • Note that NONE_SELECTED means that the user selected none of the above available options matched their skin type. This may indicate it's hard for users to self-assess Fitzpatrick skin type with the limited information provided. (See scin_questions.csv for the full provided text)

  • race_ethnicity_* (bool, sometimes present): User-reported race and/or ethnicity demographic

    • Column suffixes:
      • AMERICAN_INDIAN_OR_ALASKA_NATIVE
      • ASIAN
      • BLACK_OR_AFRICAN_AMERICAN
      • HISPANIC_LATINO_OR_SPANISH_ORIGIN
      • MIDDLE_EASTERN_OR_NORTH_AFRICAN
      • NATIVE_HAWAIIAN_OR_PACIFIC_ISLANDER
      • WHITE
      • OTHER_RACE
      • PREFER_NOT_TO_ANSWER
      • TWO_OR_MORE_AFTER_MITIGATION: As part of data processing, some race/ethnicity combinations with low numbers of contributions were grouped into this category.
    • Users may have selected more than one of the above categories.

  • textures_* (bool, sometimes present): User-reported texture of the skin condition

    • Column suffixes:
      • TEXTURE_UNSPECIFIED
      • RAISED_OR_BUMPY
      • FLAT
      • ROUGH_OR_FLAKY
      • FLUID_FILLED

  • body_parts_* (bool, sometimes present): User-reported body parts affected by the skin condition

    • Column suffixes:
      • HEAD_OR_NECK
      • ARM
      • PALM
      • BACK_OF_HAND
      • TORSO_FRONT
      • TORSO_BACK
      • GENITALIA_OR_GROIN
      • BUTTOCKS
      • LEG
      • FOOT_TOP_OR_SIDE
      • FOOT_SOLE
      • OTHER

  • condition_symptoms_* (bool, sometimes present): User-reported symptoms related to the appearance and sensation of the skin condition

    • Column suffixes:
      • BOTHERSOME_APPEARANCE
      • BLEEDING
      • INCREASING_SIZE
      • DARKENING
      • ITCHING
      • BURNING
      • PAIN
      • NO_RELEVANT_EXPERIENCE

  • other_symptoms_* (bool, sometimes present): User-reported symptoms that may also be related to the skin condition

    • Column suffixes:
      • FEVER
      • CHILLS
      • FATIGUE
      • JOINT_PAIN
      • MOUTH_SORES
      • SHORTNESS_OF_BREATH
      • NO_RELEVANT_SYMPTOMS

  • related_category (enum, sometimes present): User-reported category that may be related

    • Possible values:
      • ACNE
      • GROWTH_OR_MOLE
      • HAIR_LOSS
      • OTHER_HAIR_PROBLEM
      • NAIL_PROBLEM
      • PIGMENTARY_PROBLEM
      • RASH
      • LOOKS_HEALTHY
      • OTHER_ISSUE_DESCRIPTION

  • condition_duration (enum, sometimes present): User-reported duration of the skin condition

    • Values (based on current UI question):
      • ONE_DAY
      • LESS_THAN_ONE_WEEK
      • ONE_TO_FOUR_WEEKS
      • ONE_TO_THREE_MONTHS
      • THREE_TO_TWELVE_MONTHS
      • MORE_THAN_ONE_YEAR
      • MORE_THAN_FIVE_YEARS
      • SINCE_CHILDHOOD
      • UNKNOWN

  • image_\d_path (string, sometimes present): Path to the image storage location. Up to 3 images per case

  • image_\d_shot_type (enum, sometimes present): Enum indicating under which image question the user submitted this image.

    • Possible values: CLOSE_UP, AT_AN_ANGLE, AT_DISTANCE

scin_labels.csv

Labeling example

In the condition labeling process, each labeller provides up to 3 conditions with a confidence value of 1-5 for each condition. A case may be labeled up by up to 3 labelers. This final list of condition names and confidence values are combined into a list of condition names. For each condition, there is a rank and weight. These values are combined in the weighted_skin_condition_label column.

Example (fictional):

Dermatologist labels:

  • Dermatologist 1's label: Nummular eczema with confidence 3, Acute and chronic dermatitis with confidence 2, Psoriasis vulgaris with confidence 1

  • Dermatologist 2's label: Eczematous dermatitis with confidence 3

  • Dermatologist 3's label: Eczematous dermatitis with confidence 4, Post-inflammatory hyperpigmentation with confidence 3

The labels would be as follows:

  • dermatologist_skin_condition_label_name: Nummular eczema, Acute and chronic dermatitis, Psoriasis vulgaris, Eczematous dermatitis, Eczematous dermatitis, Post-inflammatory hyperpigmentation

  • dermatologist_skin_condition_confidence: [3, 2, 1, 3, 4, 3]

  • weighted_skin_condition_label: {'Eczema': 0.69, 'Acute and chronic dermatitis': 0.11, 'Post-Inflammatory hyperpigmentation: 0.11, 'Psoriasis': 0.08} (Note that the condition names change. In particular, both "Nummular eczema" and "Eczematous dermatitis" are mapped to "Eczema". See the weighted_skin_condition_label description below for more details.)

CSV schema:

  • case_id: Same id as the scin_cases.csv

  • dermatologist_gradable_for_skin_condition_\d (bool, sometimes present): Label from a dermatologist indicating if the skin condition can be determined from the case's images.

    • The last digit is a value in range [1, 3](inclusive), indicating one of 3 dermatologists.
    • This may be false for a number of reasons, including: there's no discernible pathology (e.g. typical skin), there's insufficient information, there are multiple conditions present, or the image quality is an issue.

  • dermatologist_skin_condition_label_name (List[string], sometimes present): List of condition names derived from dermatologist-provided labels of what condition(s) may be present in the case.

    • Note that these condition names are the original values used by the labeling.
    • There may be duplicate entries in this list. A duplicate entry means more than 1 dermatologist labeller included that condition in their response.
    • Example value: [Nummular eczema, Psoriasis vulgaris, ...]

  • dermatologist_skin_condition_confidence (List[int], sometimes present) A list of confidence scores. Each confidence score is in range [1,5](inclusive). 1 means lowest confidence and 5 means highest confidence

    • The entries in this list represent the confidence value for the corresponding entry in the previous dermatologist_skin_condition_label_name column.
    • Example value: [1, 3, …]
      • Combined with the dermatologist_skin_condition_label_name column, this would mean a dermatologist labeled the case with "Nummular eczema" with confidence 1 and "Psoriasis vulgaris" with confidence 3.

  • weighted_skin_condition_label (Dict[string, float], sometimes present): This column represents a final differential label that is generated as follows:

    • These are generated as follows:
      • For each dermatologist label (list of condition name and confidence value), the label is reweighted to the inverse of the rank within each labeler. For example, a label with values 'Condition A': 4, 'Condition B': 3, 'Condition C': 1 would have a reweighted value of 'Condition A': 1, 'Condition B': 0.5, 'Condition C': 0.3
      • These weights are then summed by condition across all dermatologist labels.
      • Then the results are normalized so the weights sum to 1.

  • dermatologist_gradable_for_fitzpatrick_skin_type_\d (string, sometimes present): Label from a dermatologist indicating if the Fitzpatrick skin type can be estimated from the case

    • Values: "YES", "NO"
    • There may be up to 3 of these labels for a case.

  • dermatologist_fitzpatrick_skin_type_label_\d (enum, sometimes present): Label from a dermatologist estimating the Fitzpatrick skin type

    • There may be up to 3 of these labels for a case.

  • gradable_for_monk_skin_tone_india (bool, sometimes present): Label from a trained lay person from a pool of graders in India indicating if the Monk skin tone label can be determined from the case's images.

  • gradable_for_monk_skin_tone_us (bool, sometimes present): Label from a trained lay person from a pool of graders in the United States indicating if the Monk skin tone label can be determined from the case's images.

  • monk_skin_tone_label_india (enum, sometimes present): Label from a trained lay person from a pool of graders in India indicating the Monk skin tone label value.

    • Possible values: 1-10

  • monk_skin_tone_label_us (enum, sometimes present): Label from a trained lay person from a pool of graders in United States indicating the Monk skin tone label value.

    • Possible values: 1-10