Database for Artificial Intelligent Surgical Instruction (DAISI)

The DAISI database consists of 14586 63×42 color images of medical procedures and examinations. Each image in the database is accompanied by a text caption describing the step of the procedure. The images from each procedure are in sequential order, from the beginning to the end of the procedure. The database contains 198 different procedures. The database is divided into training and testing sets. The training set contains 13232 images from 173 different medical procedures, and the testing set contains 1354 images from 25 different medical procedures. The procedures were randomly selected when creating the testing set. The average number of images in each training procedure is 76, with a maximum of 342 and a minimum of 5. The average number of images in each testing procedure is 54, with a maximum of 125 and a minimum of 13.

The following image exemplifies the images from 6 different procedures in the training set, each with its associated caption.

The database was created by combining datasets from Purdue University, Washington University School of Medicine, Stanford Health Care; University of California , Los Angeles; and University of Southern California.

Download

The following file includes the training and testing sets, as well as scripts to save the images automatically.

The downloaded file contains nine files:

  1. train_batch_63x42: pickle file containing the images and captions of the training set.
  2. test_batch_63x42: pickle file containing the images and captions of the testing set.
  3. install_dataset.sh: bash script to be executed after downloading the database. It will install the required python dependencies and will run the save_images.py script.
  4. save_images.py: python script that saves the images from the train and test sets into folders.
  5. createTrainJSON.py: python script that saves the training set into a JSON file.
  6. TrainProcedures.txt: text file describing the name of each procedure in the training set.
  7. TestProcedures.txt: text file describing the name of each procedure in the testing set.
  8. divisions_train.txt: support file containing the number of images in each procedure of the training set. Should not be modified.
  9. divisions_test.txt: support file containing the number of images in each procedure of the testing set. Should not be modified.

To automatically save the images, go into the folder were the .zip was downloaded and follow the steps:

  1. $ unzip DAISI.zip
  2. $ cd DAISI
  3. $ sudo chmod 774 install_dataset.sh
  4. $ ./install_dataset.sh

This will create four folders. The TrainSet and TestSet will contain all the images from the database, grouped by their respective medical procedure. The images and captions folders contains all the images and captions from the training set into a single folder, respectively. Note that this script assumes that you have pip3 installed in your working environment.