How to document?


Here a small overview of the recommended documentation is shown. Example documentation is illustrated below.

Recommended documentation

Contact information

Mandatory for the documentation is the contact information in case users need to get in touch. Required for the contact information is:

  • Name of the model:
  • Email address of contact person:

Description

For the user to make optimal use of the developed model a description is desired. This can be done in approximately 5 lines containing the following:

  • Task of the model
  • The type of network the model is based on
  • The input the model requires
  • Target group

Mechanism

This section describes the technicalities of the in- and output of the model. Optionally a description of the model could be added:

Input:

  • Datatype:
  • File format:

Output:

  • Score of performance metric:
  • Type of image output:

Validation and performance

To get an indication of the performance in training and validation set the user can decide whether the algorithm is a fit for the particular use. Also of importance is a description of the ground truth used.

Ground truth: Description what is used as a confirmation of performance

Validation/Test set:

Description Size Source Date Performance metric Performance

Warnings

This section contains important background information to keep in mind for the user when using this algorithm.

Common errors and solutions

If your algorithm is prone to certain errors that are easily solvable with a small change, please provide the error and its solution so users don't have to contact the developer for common errors.

Error Solution

Example documentation

Contact information

  • Name of the model: Segmentation of cysts in the pancreas
  • Last edited: 2021-09-09
  • Email address: fakeemailadres@source.com
  • Link to article: journalwebsite.com/doiconnectedtoarticle

Description

This model first detects cysts based on the probability of the presence of cysts. Above a certain threshold of probability, it segments the cysts. The model requires a DICOM MR image as input and creates a probability map and segmentation mask as output. The model is based on nnUnet.

Mechanism

Input:

  • Datatype: Coronal MR image of the thorax
  • File format: DICOM image of size (512, 512, 3)
  • Target group: Patients older than 18 and younger than 95

Output:

  • Results: Probability between 0 and 1 stating the possibility of a cyst being present in the pancreas
  • Type of image output: Probability map with a heat map color-coding that indicates which regions in the image were influential to the prediction of the model: colder (blue-green) and warmer (yellow-red) colors respectively indicate low and high probability regions

Validation and performance

Ground truth: Manually segmentation by experienced radiologist at the Radboudumc, Nijmegen

Validation set:

Description Size Source Date Performance metric Performance
Coronal MR thorax images 90 images University Medical Center Utrecht, Utrecht Between 2010-2014 AUC 0.96

Test set:

Description Size Source Date Performance metric Performance
Coronal MR thorax images 100 images Leiden University Medical Center, Leiden Between 2009-2012 AUC 0.89