albert-garreta / bone-age

Regressing hand bone age via interpretable features

This is an ongoing project where our goal is to extract interpretable features from hand radiographies in order to create a simple and interpretable predictive model for bone age

The data can be found in https://www.kaggle.com/kmader/rsna-bone-age

Usage

Fiting linear regressions

Running python3 scripts/data_analysis.pywill fit a linear regression using the data from data/features_df.csv, restricting to the features specified in config.FEATURES_FOR_DATA_ANALYSIS (the features id, boneage and gender are not used though) (see config.py). It also fits a linear regression for each feature, so we can see the R2 scores of each individual features. The output looks like:

Modeling with ALL FEATURES
   Std age difference (months)  Mean error  Max difference  R^2 score Num samples trained on age_bounds (years)
0                       10.364      12.244          26.581      0.695                   [22]         (0.0, 6.0)
1                        8.416      11.217          34.604      0.511                  [213]        (6.0, 12.0)
2                        7.306       8.969          20.158      0.248                  [243]       (12.0, 20.0)
3                       10.953      14.258          51.260      0.730                  [452]        (0.0, 20.0)

Modeling with single FEATURE max_purple_diameter

   Std age difference (months)  Mean error  Max difference  R^2 score Num samples trained on age_bounds (years)
0                        6.635      10.759          18.671      0.452                   [26]         (0.0, 6.0)
1                        9.827      13.693          40.005      0.298                  [216]        (6.0, 12.0)
2                        7.619       9.274          21.687      0.176                  [244]       (12.0, 20.0)
3                       14.088      19.312          62.704      0.495                  [453]        (0.0, 20.0)

[...]

Recommended features

At the moment a good selection of features for working with the whole age range (from 0 to 20 years old) is

"max_purple_diameter",
"epifisis_max_diameter_ratio",
"carp_bones_max_diameter_ratio",
"gap_ratio_5",
"gap_ratio_13",
"gap_ratio_9",

When it comes to working with narrower age ranges (0 to 6, 6 to 12, 12 to 20 years old) the following features seem to perform better

"epifisis_max_diameter_ratio",
"carp_bones_max_diameter_ratio",
"gap_ratio_5",
"gap_ratio_9",

Obtaining features_df.csv

The dataset features_df.csv is obtained by running <python3 scripts/get_features.py>. This command will create all the features listed in config.ALL_FEATURES. The code expects that, for each feature, in config.ALL_FEATURES, the class Hand has a getter method called get_<feature_name>.

New features can be added/modified by adding/modifying the corresponding method in the class Hand

The features are generated using the following information:

1. The contours of the bones, which we outlined privately.
2. Google's mediapipe library, which allows to detect hand landmarks. See landmarks_example.png for a reference. See also https://google.github.io/mediapipe/solutions/hands

NOTE At the moment due to small irregularities in how the landmarks are placed within the hand, sometimes the gap features fail to generate and thus the corresponding hand is discarded

Creating scatterplots for the features

Run python3 scripts/scatterplots_of_features.py. This will take the feature DataFrame created with get_features.py and draw scatterplots for each feature. The graphs are saved in bone-age/data/feature-scatterplots

Data

At the moment the code expects to find data in the directory bone-age/data. Hand metadata is expected to be in bone-age/data/boneage-training-dataset.csv. The json files are expected to be in bone-age/data/jsons. The raw images in bone-age/data/boneage-training. The hands with colored contours are expected to be in bone-age/data/tagged_data_contours.csv

All this can be configured within config.py