R/build_DeepMedic.R

In willi3by/niiMLr: Contains Functions to Create ML Models with Neuroimaging Data

#' Function to build the DeepMedic Network from scratch so it can be customized
#'
#' @param model_params all 3 spatial dimensions of input shape must be even; must
#' also specify downsamp factor for downsampled pathway and kernel size of conv layers
#' Input size is contextual input size (largest patch before downsampling)
#'
#' @return
#' @export
#'
#' @examples
build_DeepMedic <- function(model_params){
  high_res_path_image_size <- c(((model_params$input_shape[1:2]/model_params$d_factor - 16)*model_params$d_factor+16),model_params$input_shape[3],1)
  low_res_path_image_size <- c((model_params$input_shape[1:2]/3), model_params$input_shape[3],1)

  input_path_1 <- keras::layer_input(shape=high_res_path_image_size, name="input_path_1")
  # input_shape_2 <- c((model_params$input_shape[1:3]%/%2+8), model_params$input_shape[4])
  # input_shape_2 <- c((model_params$input_shape[1:2]%/%2+
  #                       (model_params$kernel_size-1)*model_params$downsamp_factor),
  #                    model_params$input_shape[3]%/%model_params$downsamp_factor,
  #                    model_params$input_shape[4])
  # for(d in 1:length(input_shape_2[1:3])){
  #   if(input_shape_2[d] %% 2 != 0){
  #     input_shape_2[d] <- input_shape_2[d] - 1
  #   }
  # }
  input_path_2 <- keras::layer_input(shape=low_res_path_image_size, name="input_path_2")

  path_1 <- input_path_1 %>%
    keras::layer_conv_3d(filters = 30, kernel_size = c(3,3,1),
                         kernel_initializer = keras::initializer_he_normal(),
                         kernel_regularizer = keras::regularizer_l1_l2(l1=0.00001, l2=0.0001),
                         name = "path_1_conv_1") %>%
    keras::layer_activation_parametric_relu() %>%
    keras::layer_batch_normalization() %>%
    keras::layer_spatial_dropout_3d(rate=0.02) %>%
    keras::layer_conv_3d(filters = 30, kernel_size = c(3,3,1),
                         kernel_initializer = keras::initializer_he_normal(),
                         kernel_regularizer = keras::regularizer_l1_l2(l1=0.00001, l2=0.0001),
                         name = "path_1_conv_2") %>%
    keras::layer_activation_parametric_relu() %>%
    keras::layer_batch_normalization() %>%
    keras::layer_spatial_dropout_3d(rate=0.02) %>%
    keras::layer_conv_3d(filters = 40, kernel_size = c(3,3,1),
                         kernel_initializer = keras::initializer_he_normal(),
                         kernel_regularizer = keras::regularizer_l1_l2(l1=0.00001, l2=0.0001),
                         name = "path_1_conv_3") %>%
    keras::layer_activation_parametric_relu() %>%
    keras::layer_batch_normalization() %>%
    keras::layer_spatial_dropout_3d(rate=0.02) %>%
    keras::layer_conv_3d(filters = 40, kernel_size = c(3,3,1),
                         kernel_initializer = keras::initializer_he_normal(),
                         kernel_regularizer = keras::regularizer_l1_l2(l1=0.00001, l2=0.0001),
                         name = "path_1_conv_4") %>%
    keras::layer_activation_parametric_relu() %>%
    keras::layer_batch_normalization() %>%
    keras::layer_spatial_dropout_3d(rate=0.02) %>%
    keras::layer_conv_3d(filters = 40, kernel_size = c(3,3,1),
                         kernel_initializer = keras::initializer_he_normal(),
                         kernel_regularizer = keras::regularizer_l1_l2(l1=0.00001, l2=0.0001),
                         name = "path_1_conv_5") %>%
    keras::layer_activation_parametric_relu() %>%
    keras::layer_batch_normalization() %>%
    keras::layer_spatial_dropout_3d(rate=0.02) %>%
    keras::layer_conv_3d(filters = 40, kernel_size = c(3,3,1),
                         kernel_initializer = keras::initializer_he_normal(),
                         kernel_regularizer = keras::regularizer_l1_l2(l1=0.00001, l2=0.0001),
                         name = "path_1_conv_6") %>%
    keras::layer_activation_parametric_relu() %>%
    keras::layer_batch_normalization() %>%
    keras::layer_spatial_dropout_3d(rate=0.02) %>%
    keras::layer_conv_3d(filters = 50, kernel_size = c(3,3,1),
                         kernel_initializer = keras::initializer_he_normal(),
                         kernel_regularizer = keras::regularizer_l1_l2(l1=0.00001, l2=0.0001),
                         name = "path_1_conv_7") %>%
    keras::layer_activation_parametric_relu() %>%
    keras::layer_batch_normalization() %>%
    keras::layer_spatial_dropout_3d(rate=0.02) %>%
    keras::layer_conv_3d(filters = 50, kernel_size = c(3,3,1),
                         kernel_initializer = keras::initializer_he_normal(),
                         kernel_regularizer = keras::regularizer_l1_l2(l1=0.00001, l2=0.0001),
                         name = "path_1_conv_8") %>%
    keras::layer_activation_parametric_relu() %>%
    keras::layer_batch_normalization() %>%
    keras::layer_spatial_dropout_3d(rate=0.02)

  path_2 <- input_path_2 %>%
    keras::layer_conv_3d(filters = 30, kernel_size = c(3,3,1),
                         kernel_initializer = keras::initializer_he_normal(),
                         kernel_regularizer = keras::regularizer_l1_l2(l1=0.00001, l2=0.0001),
                         name = "path_2_conv_1") %>%
    keras::layer_activation_parametric_relu() %>%
    keras::layer_batch_normalization() %>%
    keras::layer_spatial_dropout_3d(rate=0.02) %>%
    keras::layer_conv_3d(filters = 30, kernel_size = c(3,3,1),
                         kernel_initializer = keras::initializer_he_normal(),
                         kernel_regularizer = keras::regularizer_l1_l2(l1=0.00001, l2=0.0001),
                         name = "path_2_conv_2") %>%
    keras::layer_activation_parametric_relu() %>%
    keras::layer_batch_normalization() %>%
    keras::layer_spatial_dropout_3d(rate=0.02) %>%
    keras::layer_conv_3d(filters = 40, kernel_size = c(3,3,1),
                         kernel_initializer = keras::initializer_he_normal(),
                         kernel_regularizer = keras::regularizer_l1_l2(l1=0.00001, l2=0.0001),
                         name = "path_2_conv_3") %>%
    keras::layer_activation_parametric_relu() %>%
    keras::layer_batch_normalization() %>%
    keras::layer_spatial_dropout_3d(rate=0.02) %>%
    keras::layer_conv_3d(filters = 40, kernel_size = c(3,3,1),
                         kernel_initializer = keras::initializer_he_normal(),
                         kernel_regularizer = keras::regularizer_l1_l2(l1=0.00001, l2=0.0001),
                         name = "path_2_conv_4") %>%
    keras::layer_activation_parametric_relu() %>%
    keras::layer_batch_normalization() %>%
    keras::layer_spatial_dropout_3d(rate=0.02) %>%
    keras::layer_conv_3d(filters = 40, kernel_size = c(3,3,1),
                         kernel_initializer = keras::initializer_he_normal(),
                         kernel_regularizer = keras::regularizer_l1_l2(l1=0.00001, l2=0.0001),
                         name = "path_2_conv_5") %>%
    keras::layer_activation_parametric_relu() %>%
    keras::layer_batch_normalization() %>%
    keras::layer_spatial_dropout_3d(rate=0.02) %>%
    keras::layer_conv_3d(filters = 40, kernel_size = c(3,3,1),
                         kernel_initializer = keras::initializer_he_normal(),
                         kernel_regularizer = keras::regularizer_l1_l2(l1=0.00001, l2=0.0001),
                         name = "path_2_conv_6") %>%
    keras::layer_activation_parametric_relu() %>%
    keras::layer_batch_normalization() %>%
    keras::layer_spatial_dropout_3d(rate=0.02) %>%
    keras::layer_conv_3d(filters = 50, kernel_size = c(3,3,1),
                         kernel_initializer = keras::initializer_he_normal(),
                         kernel_regularizer = keras::regularizer_l1_l2(l1=0.00001, l2=0.0001),
                         name = "path_2_conv_7") %>%
    keras::layer_activation_parametric_relu() %>%
    keras::layer_batch_normalization() %>%
    keras::layer_spatial_dropout_3d(rate=0.02) %>%
    keras::layer_conv_3d(filters = 50, kernel_size = c(3,3,1),
                         kernel_initializer = keras::initializer_he_normal(),
                         kernel_regularizer = keras::regularizer_l1_l2(l1=0.00001, l2=0.0001),
                         name = "path_2_conv_8") %>%
    keras::layer_activation_parametric_relu() %>%
    keras::layer_batch_normalization() %>%
    keras::layer_spatial_dropout_3d(rate=0.02) %>%
    keras::layer_upsampling_3d(size=c(3,3,1))


  # path_1_shape <- path_1$get_shape()$as_list() %>%
  #   unlist()
  # path_2_shape <- path_2$get_shape()$as_list() %>%
  #   unlist()
  # shape_diff <- path_1_shape - path_2_shape
  # if(sum(shape_diff) > 0){
  #   path_2 <- path_2 %>%
  #     keras::layer_zero_padding_3d(padding = list(
  #       list(shape_diff[1],0),
  #       list(shape_diff[2],0),
  #       list(shape_diff[3],0)
  #     ))
  # }

concat_layer <- keras::layer_add(list(path_1, path_2))

main_output <- concat_layer %>%
  keras::layer_dense(units = 150, activation="relu") %>%
  # keras::layer_alpha_dropout(rate=0.5) %>%
  keras::layer_dense(units = 150, activation="relu") %>%
  # keras::layer_alpha_dropout(rate=0.5) %>%
  keras::layer_dense(units = model_params$num_classes,
                     activation = model_params$activation, name="main_output")

model <- keras::keras_model(inputs = c(input_path_1, input_path_2), outputs = main_output)
# model <- keras::keras_model(inputs = concat_layer, outputs = main_output)
model %>% keras::compile(optimizer = model_params$optimizer,
                  loss = model_params$loss,
                  metrics = model_params$metrics)
return(model)
}