Detect COVID-19 from X-Ray images using Keras.
Nothing in this repository should ever be considered fit for clinical use. This project is intended to be a proof-of-concept demonstrating binary classification of images.
git clone https://github.com/brianlechthaler/covid_xray_classification
cd covid_xray_classification
python -m pip install -r requirements.txt
python setup.py build ; python setup.py install
# Don't forget to have your kaggle.json api key in the right place!
python example.py
Available from example.py
.
# Import everything we need
from covid_xray_classification.models.xception import Small
from covid_xray_classification.data import Downloader, Reshaper
from pandas import read_csv
from os.path import join
from tensorflow.keras.callbacks import EarlyStopping, ModelCheckpoint, TensorBoard
from tensorflow.keras.optimizers import Adam
from tensorflow.keras.preprocessing import image_dataset_from_directory
from tensorflow.keras.metrics import BinaryAccuracy,Precision,Recall
# Download default dataset to default location
Downloader().download()
# Specify a few runtime variables
columns = ['patientid',
'filename',
'classification',
'datasource']
input_folder_prefix = 'dataset'
reshaped_dataset_folder = 'reshaped'
batch_size = 16
image_size = (512,512)
rng_seed = 127001
validation_split = 0.1
epochs = 50
learning_rate = 1e-3
model_name = 'COVID_Chest_X-Ray_BinaryClassification_512x512'
# Make folders based on labels corresponding to the images the folders contain.
# Here we put both train and test images in the same place, as we will split up the images ourselves later on.
for file_prefix in ['train', 'test']:
# Extract metadata from the file containing it.
metadata_table = read_csv(join(input_folder_prefix,
f"{file_prefix}.txt"),
sep=' ')
# Define custom columns for the data we're importing.
metadata_table.columns = columns
# Create a Reshaper to reshape our data according to the parameters we specify
reshape_task = Reshaper(table=metadata_table,
input_folder=join(input_folder_prefix,
file_prefix),
output_folder=join(input_folder_prefix,
'reshaped'))
# Reshape our data according to specified parameters.
reshape_task.reshape()
# Create a small Xception model designed to work with the size of images in our dataset.
net = Small(image_size=image_size)
model = net.model
# Define callbacks.
# Here we use the EarlyStopping callback to stop training if the validation accuracy stops increasing.
# This both saves a significant amount of power and usually decreases the total number of epochs to a fraction of what most will end up specifying
# We also make sure that this callback will automatically pick the best epoch at the end of training.
callbacks = [EarlyStopping("loss",
patience=5,
mode='min',
restore_best_weights=True),
# Save the model at the end of each epoch.
ModelCheckpoint("epoch_{epoch}.h5"),
# Log data to directory logs for TensorBoard
TensorBoard(
log_dir="logs")]
# Specify the metrics we wish to evaluate at the end of each epoch.
metrics = ["accuracy",
BinaryAccuracy(),
Precision(name='precision'),
Recall(name='recall')]
# Compile the model.
# Here we use Adam as our optimizer, and binary_crossentropy as our loss as we are only doing binary classification:
# in other words, if all we need is 0: negative, 1: positive
model.compile(optimizer=Adam(learning_rate),
loss="binary_crossentropy",
metrics=metrics)
# Create a training dataset from 90% of the images in the dataset.
train_dataset = image_dataset_from_directory(
directory=join(input_folder_prefix,
'reshaped'),
labels='inferred',
label_mode='binary',
batch_size=batch_size,
image_size=image_size,
validation_split=validation_split,
subset='training',
seed=rng_seed
)
# Create a validation dataset from 10% of the images in the dataset.
validation_dataset = image_dataset_from_directory(
directory=join(input_folder_prefix,
'reshaped'),
labels='inferred',
label_mode='binary',
batch_size=batch_size,
image_size=image_size,
validation_split= validation_split,
subset='validation',
seed=rng_seed
)
# Cache our datasets in memory to speed up training
train_ds = train_dataset.prefetch(buffer_size=32)
val_ds = validation_dataset.prefetch(buffer_size=32)
# Train our model.
model.fit(
train_ds,
epochs=epochs,
callbacks=callbacks,
validation_data=val_ds)
# Save our model for later use.
model.save(join('dataset',
f"{model_name}.h5"))