Melanoma Classifier

This project aims to classify a new melanoma picture as malignant or benign.

(Be careful when interpreting the results, it will never replace a doctor's opinion.)

0) Requirements

Don't forget to modify the CONFIG.txt file (default configuration works)

!pip install -U efficientnet
from API import *

CFG = get_config()

1) Extracting tabular features

1. First of all, before extracting features, we need to segment all images from the dataset. Here is the API call:

# CFG : the config dictionnary
# dataframe: a pandas.DataFrame with at least 1 column: "filename" containing the name(including the extension)
#            for each image, and another column "target"(OPTIONNAL) with the labels "benign" or "malignant"
# images_path: the path where all inputs images are located
# segmentations_path: the path where all outputs segmentations will be saved
get_segmentations(CFG, dataframe, images_path, segmentations_path)

2. Finally, we can compute tabular features:

# CFG : the config dictionnary
# dataframe: a pandas.DataFrame with at least 1 column: "filename" containing the name(including the extension)
#            for each image, and another column "target"(OPTIONNAL) with the labels "benign" or "malignant"
# images_path: the path where all inputs images are located
# segmentations_path: the path where all outputs segmentations are located
df = get_tabular_dataframe(CFG, dataframe, images_path, segmentations_path)

The returned DataFrame is normalized (MinMax scaler)

Features are calculated with region properties. I also tried to reproduce the famous ABCD rule (Assymetry, Border irregularity & Colors Descriptors). (In the following tab, A = Area, P = Perimeter, d = minor axis length, D = major axis length. Others variables are explained in references)

Morphological formulas	Assymetry ref	Border irregularity ref	Colors ref
extent			F4, F5, F6
solidity			F10, F11, F12
			F13, F14, F15

2) Writing & reading TFRecords

1. Writing a TFRecord

# CFG : the config dictionnary
# df: a pandas.DataFrame containing columns in this order: "filename" containing the name(including the extension)
#            for each image, "target"(OPTIONNAL) with the labels "benign" or "malignant", and all other columns are the features.
#            In our exemple, we will have 21 columns ("filename","target",+19 features)
# images_path: the path where all inputs images are located
# output_path: path where TFRecord files will be stored)
# nb: (OPTIONNAL, default=1) the number of TFRecord files to create
# preprocess_function: (OPTIONNAL) a function to preprocess images. None = no preprocessing. Default: whit-balancing + hair removal.
write_tfrecord(CFG, df, images_path, output_path, preprocess_function)

2. Reading a TFRecord

# CFG : the config dictionnary
# tfrecord_train(test): the path containing tfrecord files built with training(test) data
# labeled: if images are labeled (True if training, False if testing)
# augment (OPTIONNAL): if Images should be augmented (only if labeled=True for training)
dataset_train = read_tfrecord(CFG, tfrecord_train, labeled=True)
dataset_test  = read_tfrecord(CFG, tfrecord_test, labeled=False)
# dataset_train(test) (output): a dataset to give to our model.

3) Create Model

Here we have pre-trained weights for different configurations:

• B0 (net_count=1, fine_tune=False, preprocess_function=None): Download
• B0 (net_count=1, fine_tune=True, preprocess_function=None): Download
• B0-2 (net_count=3, fine_tune=True, preprocess_function=None): Download
• B0-4 (net_count=5, fine_tune=False, preprocess_function=None): Download

# CFG : the config dictionnary
# fine_tune(OPTIONNAL, default:False): True:Fine-tuning, False:Transfer-learning
# model_weights(OPTIONNAL): the path containing valid weights for the model
model = get_model(CFG, fine_tune=False, model_weights="???")
# model(output): a tensorflow model

4) Fit & Predict

1. Fit

callbacks = [tf.keras.callbacks.ReduceLROnPlateau(),]
             #tf.keras.callbacks.EarlyStopping(monitor='val_loss', patience=5),
             #tf.keras.callbacks.ModelCheckpoint("models/best.h5", save_best_only=True, monitor='val_auc', mode='max', save_weights_only=True)]

history = model.fit(
    dataset_train,
    steps_per_epoch = len(df)/(CFG['batch_size'] * CFG['REPLICAS']),
    epochs = CFG['epochs'],
    callbacks = callbacks,
)

2. Predict

preds = model.predict(
        dataset_test,
        steps = len(df)/(CFG['batch_size'] * CFG['REPLICAS']),
)

To-Do List

• Easier way to implement his own preprocessing function (when writing a TFRecord file)
• TPU support (It needs to be tested, not sure at all)
• Improve the segmentation part to improve features' quality

License

CC 4.0 Attribution-NonCommercial International

The software is for educational and academic research purposes only.

Ce(tte) œuvre est mise à disposition selon les termes de la Licence Creative Commons Attribution - Pas d’Utilisation Commerciale 4.0 International.