Melanoma task

.gitignore

@@ -14,4 +14,4 @@ inst/doc
 /doc/
 /Meta/
 CRAN-SUBMISSION
-paper/data
+paper/data

DESCRIPTION

@@ -56,6 +56,7 @@ Imports:
 	withr
 Suggests:
     callr,
+    curl,
     future,
     ggplot2,
     igraph,
@@ -125,6 +126,7 @@ Collate:
     'PipeOpTorchReshape.R'
     'PipeOpTorchSoftmax.R'
     'TaskClassif_lazy_iris.R'
+    'TaskClassif_melanoma.R'
     'TaskClassif_mnist.R'
     'TaskClassif_tiny_imagenet.R'
     'TorchDescriptor.R'

R/TaskClassif_melanoma.R

@@ -0,0 +1,127 @@
+#' @title Melanoma Image classification
+#' @name mlr_tasks_melanoma
+#' @description
+#' Classification of melanoma tumor images.
+#'
+#' The data comes from the 2020 SIIM-ISIC challenge.
+#'
+#' @section Construction:
+#' ```
+#' tsk("melanoma")
+#' ```
+#'
+#' @template task_download
+#'
+#' @source
+#' \url{https://challenge2020.isic-archive.com/}
+#'
+#' @section Properties:
+#' `r rd_info_task_torch("melanoma", missings = FALSE)`
+#'
+#' @references
+#' `r format_bib("melanoma2021")`
+#' task = tsk("melanoma")
+#' task
+NULL
+
+# @param path (`character(1)`)\cr
+#   The cache_dir/datasets/melanoma folder
+constructor_melanoma = function(path) {
+  require_namespaces("curl")
+
+  base_url = "https://huggingface.co/datasets/carsonzhang/ISIC_2020_small/resolve/main/"
+
+  compressed_tarball_file_name = "hf_ISIC_2020_small.tar.gz"
+  compressed_tarball_path = file.path(path, compressed_tarball_file_name)
+  curl::curl_download(paste0(base_url, compressed_tarball_file_name), compressed_tarball_path)
+  utils::untar(compressed_tarball_path, exdir = path)
+  on.exit({file.remove(compressed_tarball_path)}, add = TRUE)
+
+  training_metadata_file_name = "ISIC_2020_Training_GroundTruth_v2.csv"
+  training_metadata = data.table::fread(file.path(path, training_metadata_file_name))
+
+  test_metadata_file_name = "ISIC_2020_Test_Metadata.csv"
+  test_metadata = file.path(path, test_metadata_file_name)
+
+  training_metadata = training_metadata[, split := "train"]
+  test_metadata = setnames(test_metadata,
+    old = c("image", "patient", "anatom_site_general"),
+    new = c("image_name", "patient_id", "anatom_site_general_challenge")
+  )[, split := "test"]
+  # response column needs to be filled for the test data
+  metadata = rbind(training_metadata, test_metadata, fill = TRUE)
+  metadata[, image_name := NULL]
+  metadata[, target := NULL]
+  metadata = setnames(metadata, old = "benign_malignant", new = "outcome")
+
+  melanoma_ds_generator = torch::dataset(
+    initialize = function() {
+      self$.metadata = metadata
+      self$.path = path
+    },
+    .getitem = function(idx) {
+      force(idx)
+
+      x = torchvision::base_loader(file.path(self$.path, paste0(self$.metadata[idx, ]$file_name)))
+      x = torchvision::transform_to_tensor(x)
+
+      return(list(x = x))
+    },
+    .length = function() {
+      nrow(self$.metadata)
+    }
+  )
+
+  melanoma_ds = melanoma_ds_generator()
+
+  dd = as_data_descriptor(melanoma_ds, list(x = c(NA, 3, 128, 128)))
+  lt = lazy_tensor(dd)
+
+  return(cbind(metadata, data.table(image = lt)))
+}
+
+load_task_melanoma = function(id = "melanoma") {
+  cached_constructor = function(backend) {
+    data = cached(constructor_melanoma, "datasets", "melanoma")$data
+
+    data[, outcome := factor(get(outcome), levels = c("benign", "malignant"))]
+    set(data, j = "outcome", value = factor(get(outcome), levels = c("benign", "malignant")))
+
+    char_features = c("sex", "anatom_site_general_challenge")
+    data[, (char_features) := lapply(.SD, factor), .SDcols = char_features]
+
+    dt = cbind(
+      data,
+      data.table(
+        ..row_id = seq_len(nrow(data))
+      )
+    )
+
+    DataBackendDataTable$new(data = dt, primary_key = "..row_id")
+  }
+
+  backend = DataBackendLazy$new(
+    constructor = cached_constructor,
+    rownames = seq_len(32701 + 10982),
+    col_info = load_col_info("melanoma"),
+    primary_key = "..row_id"
+  )
+
+  task = TaskClassif$new(
+    backend = backend,
+    id = "melanoma",
+    target = "outcome",
+    label = "Melanoma Classification"
+  )
+
+  task$set_col_roles("patient_id", "group")
+  task$col_roles$feature = c("sex", "anatom_site_general_challenge", "age_approx", "image")
+
+  backend$hash = task$man = "mlr3torch::mlr_tasks_melanoma"
+
+  task$filter(1:32701)
+
+  return(task)
+}
+
+register_task("melanoma", load_task_melanoma)

R/bibentries.R

@@ -112,6 +112,15 @@ bibentries = c(# nolint start
     booktitle = "Proceedings of the IEEE conference on computer vision and pattern recognition ",
     pages     = "2818--2826 ",
     year      = "2016 "
+  ),
+  melanoma2021 = bibentry("article",
+    title = "A patient-centric dataset of images and metadata for identifying melanomas using clinical context",
+    author = "Rotemberg, V. and Kurtansky, N. and Betz-Stablein, B. and Caffery, L. and Chousakos, E. and Codella, N. and Combalia, M. and Dusza, S. and Guitera, P. and Gutman, D. and Halpern, A. and Helba, B. and Kittler, H. and Kose, K. and Langer, S. and Lioprys, K. and Malvehy, J. and Musthaq, S. and Nanda, J. and Reiter, O. and Shih, G. and Stratigos, A. and Tschandl, P. and Weber, J. and Soyer, P.",
+    journal = "Scientific Data",
+    volume = "8",
+    pages = "34",
+    year = "2021",
+    doi = "10.1038/s41597-021-00815-z"
   )
 ) # nolint end
 

benchmarks/image_loaders/benchmark_image_loaders.R

@@ -0,0 +1,63 @@
+library(torch)
+library(torchvision)
+library(mlr3torch)
+library(here)
+
+library(data.table)
+setDTthreads(threads = 1)
+
+training_metadata = fread(here::here("cache", "ISIC_2020_Training_GroundTruth.csv"))
+
+# hard-coded cache directory that I use locally
+cache_dir = here("cache")
+
+ds_base_loader = torch::dataset(
+  initialize = function(n_images) {
+    self$.metadata = fread(here(cache_dir, "ISIC_2020_Training_GroundTruth.csv"))[1:n_images, ]
+    self$.path = file.path(here(cache_dir), "train")
+  },
+  .getitem = function(idx) {
+    force(idx)
+
+    x = torchvision::base_loader(file.path(self$.path, paste0(self$.metadata[idx, ]$image_name, ".jpg")))
+    x = torchvision::transform_to_tensor(x)
+
+    return(list(x = x))
+  },
+  .length = function() {
+    nrow(self$.metadata)
+  }
+)
+
+ds_magick_loader = torch::dataset(
+  initialize = function(n_images) {
+    self$.metadata = fread(here(cache_dir, "ISIC_2020_Training_GroundTruth.csv"))[1:n_images, ]
+    self$.path = file.path(here(cache_dir), "train")
+  },
+  .getitem = function(idx) {
+    force(idx)
+
+    image_name = self$.metadata[idx, ]$image_name
+
+    x = magick::image_read(file.path(self$.path, paste0(image_name, ".jpg")))
+    x = torchvision::transform_to_tensor(x)
+
+    return(list(x = x, image_name = image_name))
+  },
+  .length = function() {
+    nrow(self$.metadata)
+  }
+)
+
+n_images = 10
+
+ds_base = ds_base_loader(n_images)
+ds_magick = ds_magick_loader(n_images)
+
+bmr = bench::mark(
+  for (i in 1:n_images) ds_base$.getitem(i),
+  for (i in 1:n_images) ds_magick$.getitem(i),
+  memory = FALSE
+)
+
+print(bmr)

data-raw/melanoma.R

@@ -0,0 +1,80 @@
+devtools::load_all()
+
+# manually construct the task once
+# library(here)
+# library(data.table)
+library(data.table)
+withr::local_options(mlr3torch.cache = TRUE)
+
+constructor_melanoma = function(path) {
+  require_namespaces("curl")
+
+  base_url = "https://huggingface.co/datasets/carsonzhang/ISIC_2020_small/resolve/main/"
+
+  compressed_tarball_file_name = "hf_ISIC_2020_small.tar.gz"
+  compressed_tarball_path = file.path(path, compressed_tarball_file_name)
+  curl::curl_download(paste0(base_url, compressed_tarball_file_name), compressed_tarball_path)
+  utils::untar(compressed_tarball_path, exdir = path)
+  on.exit({file.remove(compressed_tarball_path)}, add = TRUE)
+
+  training_metadata_file_name = "ISIC_2020_Training_GroundTruth_v2.csv"
+  training_metadata = fread(file.path(path, training_metadata_file_name))
+
+  test_metadata_file_name = "ISIC_2020_Test_Metadata.csv"
+  test_metadata = fread(file.path(path, test_metadata_file_name))
+
+  training_metadata = training_metadata[, split := "train"]
+  test_metadata = setnames(test_metadata,
+    old = c("image", "patient", "anatom_site_general"),
+    new = c("image_name", "patient_id", "anatom_site_general_challenge")
+  )[, split := "test"]
+  # response column needs to be filled for the test data
+  metadata = rbind(training_metadata, test_metadata, fill = TRUE)
+  metadata[, image_name := NULL]
+  metadata[, target := NULL]
+  metadata = setnames(metadata, old = "benign_malignant", new = "outcome")
+
+  melanoma_ds_generator = torch::dataset(
+    initialize = function() {
+      self$.metadata = metadata
+      self$.path = path
+    },
+    .getitem = function(idx) {
+      force(idx)
+
+      x = torchvision::base_loader(file.path(self$.path, paste0(self$.metadata[idx, ]$file_name)))
+      x = torchvision::transform_to_tensor(x)
+
+      return(list(x = x))
+    },
+    .length = function() {
+      nrow(self$.metadata)
+    }
+  )
+
+  melanoma_ds = melanoma_ds_generator()
+
+  dd = as_data_descriptor(melanoma_ds, list(x = c(NA, 3, 128, 128)))
+  lt = lazy_tensor(dd)
+
+  return(cbind(metadata, data.table(image = lt)))
+}
+
+bench::system_time(melanoma_dt <- constructor_melanoma(file.path(get_cache_dir(), "datasets", "melanoma")))
+# melanoma_dt = constructor_melanoma(file.path(get_cache_dir(), "datasets", "melanoma"))
+
+# change the encodings of variables: diagnosis, outcome
+melanoma_dt[, outcome := factor(get(outcome), levels = c("benign", "malignant"))]
+
+char_features = c("sex", "anatom_site_general_challenge")
+melanoma_dt[, (char_features) := lapply(.SD, factor), .SDcols = char_features]
+
+tsk_melanoma = as_task_classif(melanoma_dt, target = "outcome", id = "melanoma")
+tsk_melanoma$set_col_roles("patient_id", "group")
+tsk_melanoma$col_roles$feature = c(char_features, "age_approx", "image")
+
+tsk_melanoma$label = "Melanoma Classification"
+
+ci = col_info(tsk_melanoma$backend)
+
+saveRDS(ci, here::here("inst/col_info/melanoma.rds"))

data-raw/tiny_imagenet.R

@@ -2,12 +2,4 @@ devtools::load_all()
 
 ci = col_info(get_private(tsk("tiny_imagenet")$backend)$.constructor())
 
-saveRDS(ci, here::here("inst/col_info/tiny_imagenet.rds"))
-
-mlr3:::DataBackendCbind$new(c)
-
-
-split = factor(rep(c("train", "valid", "test"), times = c(100000, 10000, 10000)))
-
-ci = rbind(ci, data.table(id = "split", type = "factor", levels = levels(split)))
-setkeyv(ci)
+saveRDS(ci, here::here("inst/col_info/tiny_imagenet.rds"))