This code processes annotations for ISIC 2017 lesion images from previous years, and assigns unannotated images to the new groups. annotation_processing_script.py calls all the functions that are neccesary to assign the images to the new groups. The following variables need to be set by the user:
pandas>=1.0.5
numpy>=1.18.5
AnnotationPath: path to folder containing folders for each year. these folders containin the annotations as csv files.
SavePath: Path to new folder where 2 csv's for each group will be saved.
DataTypeFilename: Filename of csv where data/annotation types are assigned.
TrainPath: Path to ISIC-2017 part 3 training set ground truth
ValidationPath: Path to ISIC-2017 part 3 validation set ground truth
TestPath: Path to ISIC-2017 part 3 test set ground truth
NumGroups: Number of groups that take the course.
NumImages: Number of images that need to be assigned to a group.
AnnotationBalance: Distribution Malignant/Benign for images.
RequiredAnnotations: List of which annotation types an image has to be annotated before the program considers the image annotated.
RequiredAmounts: List of same length as RequiredAnnotations containing the amount of annotations for each catergory in RequiredAnnotations before the program considers the image annotated.
The following data/annotation types can be chosen in RequiredAnnotations. Only Asymmetry, Border and Color contain a significant amount of annotations (around 10000 each). The output of the function annotation_stats is shown first:
| data_type | ann_count | mean_ann_img | min_ann_img |
|---|---|---|---|
| Asymmetry | 10471 | 6.4 | 3 |
| Black | 1099 | 5.5 | 2 |
| Blood | 250 | 2.5 | 2 |
| Blue | 1197 | 4.0 | 3 |
| Border | 10762 | 6.6 | 3 |
| Brown | 800 | 8.0 | 8 |
| Color | 10175 | 6.2 | 3 |
| Color_Categorised | 1599 | 3.4 | 3 |
| Color_Fade | 1000 | 5.0 | 3 |
| Compactness | 1764 | 5.3 | 3 |
| Dermo | 1485 | 3.8 | 3 |
| Diameter | 295 | 3.0 | 1 |
| Erythema | 664 | 5.0 | 4 |
| Flaking | 300 | 3.0 | 3 |
| Red | 1100 | 5.5 | 3 |
| Rough_Surface | 300 | 3.0 | 3 |
| Skin_Color | 299 | 3.0 | 2 |
| White | 300 | 3.0 | 3 |
| Yellow | 800 | 8.0 | 8 |
Asymmetry: How asymmetric a lesion is scored on a scale or boolean
Border: How irregular the border of a lesion is scored on a scale or boolean
Color: How uneven the color of a lesion is scored on a scale or boolean
Color_Categorised: Colors that are found in the lesion are noted with comma's between them. (different groups use different methods)
Dermo: Dermascopic/Differential structures. for instance pigment dots or globules.
Blood: Blood vessels can be found in the lesion
Blue: The lesions have a blue glow.
Color_Fade: Lesions have a faded border.
Compactness: uses formula c = circumference^2/(4π*Area)
Erythema: redness surrounding the lesion.
Red: Red is present in the lesion.
Black: Black is present in the lesion.
Skin_Color: Skin_Color is present in the lesion.
Flaking: Lesions shows flaking skin
Rough_Surface: surface of lesion is rough
White: White is present in the lesion.
Diameter: score based on the diameter of the lesion
Yellow: Yellow is present in the lesion.
Brown: Brown is present in the lesion.
create_annotation_df: loops through all folders from each year and applies select_data to generate one large dataframe containing features. Has columns ['ID', 'group_number', 'year', 'annotator', 'orig_column', 'data_type', 'data']. This dataframe can be used to select data for more specific applications.
annotation_stats: Uses annotations_df to print the ammount of annotations (ann_count), average amount of annotations per image(mean_ann_img) and minimum amount of annotations per image(min_ann_img) for each annotation type
get_annotations: returns pandas series of all annotations as lists using annotations_df and an ISIC_ID as a string.
drop_annotation_count_categories: Returns series of all images that are considered annotated using annotations_df, RequiredAnnotations, and RequiredAmounts
categorise_annotations: returns 4 dataframes contaning image ID and groundtruth: combinations of: unannotated, annotated, benign and malignant. This is done using a series of images id's that are considered annotated and paths to the groundtruths.
create_group_sets: Returns a dict containing dataframes for each group with assigned ISIC ID's and groundtruths. This is based on the 4 dataframes explained in the previous function and NumGroups NumImages AnnotationBalance.
save_group_sets: saves image id's (with and without ground truth) as csv to the folder SavePath, using the dict containing dataframes with image id's and groundtruth.