Predicting skin lesions
Dr Juan H Klopper
Introduction
While it is a simple task to download images that are part of a predesigned dataset, the ultimate goal is to use your own images. In this file, we take a look at how to classify images of skin lesions as either benign or malignant. These images are available on Kaggle.
Since images take up a lot of memory, we will also use an image data generator that will load images from a local drive in batches, as and when needed.
Libraries
We will use the Keras library. Note that this markdown file makes use of Keras as of October 2019 (TensorFlow 2.0).
setwd(getwd())
library(keras)Path to data
It is important that the data is saved in a particular structure on your hard drive. You should have separate directories (folders) for the training, validation, and test sets. In each of these, you should have separate folders for the different classes. Since we have two classes, benign and malignant, each of our three directories holds both of these as sub directories. Make sure that the names of the sub directories are spelled similarly in each of the three directories.
Below, we set a file path to each of the three directories.
train_dir <- file.path("d:", "Kaggle", "R", "skin", "train")
validation_dir <- file.path("d:", "Kaggle", "R", "skin", "validation")
test_dir <- file.path("d:", "Kaggle", "R", "skin", "test")Will will use the number of images in each of the sub directories to counts the steps in our model.
train_benign_dir <- file.path(train_dir, "benign")
train_malignant_dir <- file.path(train_dir, "malignant")
validation_benign_dir <- file.path(validation_dir, "benign")
validation_malignant_dir <- file.path(validation_dir, "malignant")
test_benign_dir <- file.path(test_dir, "benign")
test_malignant_dir <- file.path(test_dir, "malignant")Below, we find the file counts.
num_train_benign <- length(list.files(train_benign_dir))
num_train_malignant <- length(list.files(train_malignant_dir))
num_validation_benign <- length(list.files(validation_benign_dir))
num_validation_malignant <- length(list.files(validation_malignant_dir))
num_test_benign <- length(list.files(test_benign_dir))
num_test_malignant <- length(list.files(test_malignant_dir))
total_train <- num_train_benign + num_train_malignant
total_validation <- num_validation_benign + num_validation_malignant
total_test <- num_test_benign + num_test_malignantData generators
To take the image in batches from a local drive, we set up a generator with the image_data_generator() function. We make use of image augmentation to improve training.
IMG_HEIGHT <- 112 # Small image sizes for demo purposes only
IMG_WIDTH <- 112
batch_size <- 4 # Samll batch size for demo purposes onlyWe set up these generators for the training, validation, and test images.
image_gen_train <- keras::image_data_generator(rescale = 1./255,
rotation_range = 10,
width_shift_range = 0.15,
height_shift_range = 0.15,
horizontal_flip = TRUE,
zoom_range = 0.05)
image_gen_validation <- keras::image_data_generator(rescale = 1./255) # The validation and test images are not augmented
image_gen_test <- keras::image_data_generator(rescale = 1./255)Now we use the flow_from_directory() function. It will pull the images in batches as needed during training.
train_data_gen <- keras::flow_images_from_directory(train_dir,
generator = image_gen_train,
batch_size = batch_size,
target_size = c(IMG_HEIGHT,
IMG_WIDTH),
class_mode = "binary")
validation_data_gen <- keras::flow_images_from_directory(validation_dir,
generator = image_gen_validation,
batch_size = batch_size,
target_size = c(IMG_HEIGHT,
IMG_WIDTH),
class_mode = "binary")
test_data_gen <- keras::flow_images_from_directory(test_dir,
generator = image_gen_test,
batch_size = batch_size,
target_size = c(IMG_HEIGHT,
IMG_WIDTH),
class_mode = "binary")Creating a model
Our model is a simple two convolutional layer model.
model <- keras::keras_model_sequential() %>%
layer_conv_2d(filters = 16,
kernel_size = 3,
padding = "same",
activation = "relu",
input_shape = c(IMG_HEIGHT,
IMG_WIDTH,
3)) %>%
layer_max_pooling_2d() %>%
layer_dropout(0.2) %>%
layer_conv_2d(filters = 32,
kernel_size = 3,
padding = "same",
activation = "relu") %>%
layer_max_pooling_2d() %>%
layer_dropout(0.2) %>%
layer_flatten() %>%
layer_dense(512,
activation = "relu") %>%
layer_dense(1,
activation = "sigmoid")model %>% summary()## Model: "sequential"
## ___________________________________________________________________________
## Layer (type) Output Shape Param #
## ===========================================================================
## conv2d (Conv2D) (None, 112, 112, 16) 448
## ___________________________________________________________________________
## max_pooling2d (MaxPooling2D) (None, 56, 56, 16) 0
## ___________________________________________________________________________
## dropout (Dropout) (None, 56, 56, 16) 0
## ___________________________________________________________________________
## conv2d_1 (Conv2D) (None, 56, 56, 32) 4640
## ___________________________________________________________________________
## max_pooling2d_1 (MaxPooling2D) (None, 28, 28, 32) 0
## ___________________________________________________________________________
## dropout_1 (Dropout) (None, 28, 28, 32) 0
## ___________________________________________________________________________
## flatten (Flatten) (None, 25088) 0
## ___________________________________________________________________________
## dense (Dense) (None, 512) 12845568
## ___________________________________________________________________________
## dense_1 (Dense) (None, 1) 513
## ===========================================================================
## Total params: 12,851,169
## Trainable params: 12,851,169
## Non-trainable params: 0
## ___________________________________________________________________________model %>% compile(loss = "binary_crossentropy",
optimizer = optimizer_adam(),
metrics = c("accuracy"))Training
We now train the model over \(10\) epochs, with early stopping.
history <- keras::fit_generator(model,
train_data_gen,
steps_per_epoch = floor(total_train / batch_size),
epochs = 10,
validation_data = validation_data_gen,
validation_steps = floor(total_validation / batch_size),
callbacks = callback_early_stopping(monitor = "val_loss",
min_delta = 0.01,
patience = 4))Evaluating the model
We can now use the test image generator to check on our model.
score <- model %>% evaluate_generator(test_data_gen,
steps = floor(total_test / batch_size))Let’s have a look at the accuracy.
cat('Test accuracy: ', score$acc, "\n")## Test accuracy: 0.6989051Saving and reloading the model
We can save this model as in HDF5 format.
model %>% save_model_hdf5("skin.h5")Reloading is simple.
load_model <- load_model_hdf5("skin.h5")
load_model %>% summary()## Model: "sequential"
## ___________________________________________________________________________
## Layer (type) Output Shape Param #
## ===========================================================================
## conv2d (Conv2D) (None, 112, 112, 16) 448
## ___________________________________________________________________________
## max_pooling2d (MaxPooling2D) (None, 56, 56, 16) 0
## ___________________________________________________________________________
## dropout (Dropout) (None, 56, 56, 16) 0
## ___________________________________________________________________________
## conv2d_1 (Conv2D) (None, 56, 56, 32) 4640
## ___________________________________________________________________________
## max_pooling2d_1 (MaxPooling2D) (None, 28, 28, 32) 0
## ___________________________________________________________________________
## dropout_1 (Dropout) (None, 28, 28, 32) 0
## ___________________________________________________________________________
## flatten (Flatten) (None, 25088) 0
## ___________________________________________________________________________
## dense (Dense) (None, 512) 12845568
## ___________________________________________________________________________
## dense_1 (Dense) (None, 1) 513
## ===========================================================================
## Total params: 12,851,169
## Trainable params: 12,851,169
## Non-trainable params: 0
## ___________________________________________________________________________We can use the test set as before.
new_score <- load_model %>% evaluate_generator(test_data_gen,
steps = floor(total_test / batch_size))cat('Test accuracy: ', new_score$acc, "\n")## Test accuracy: 0.6989051