Documentation for dcm2reg
dcm2reg is a module which is useful for preprocessing medical image data,
whenever the end goal is to prepare the data for development of e.g. deep
learning models which can automatically delineate organs and/or tumours.
The module is available in python and is known to work on Windows, however
other common operating systems (OS X, Linus) likely work as well, as long as
dependencies can be properly installed.
The preprocessing which dcm2reg is useful for broadly consists of three
steps that propagate the data from a file format known as DICOM to the
NIfTI-format while also performing registration on the files (thus the name
dcm2reg). The three steps are:
Conversion between file formats.
Registration of images.
Evaluation of the performed registration.
Refer to Background for more information on file formats, the registration task etc.
Motivation
Developing deep learning models for delineation is of high interest in e.g. the planning phase of radiation therapy for patients with cancer. Delineations are crucial in this application because doctors want to deliver an appropriate radiation dosage to cancer tumours while affecting organs at risk with as small dosages as possible. While deep learning models will hardly replace doctors entirely in delineation tasks at first, simply developing tools that can give suggestions which the doctors can validate upon will cut down manual labour on an otherwise time-consuming task. Furthermore, data-based models will hopefully be able to reduce some of the individual error-prone biases that doctors inevitably harbour and thus improve standardisation of treatments.
Contributions
dcm2reg was created by three data science students at Aarhus University
during a semester-long data project: Andreas Busch, Philip Leth and Jonathan
Nielsen. However, it is also built upon the dcm2nii module with scripts
modified by several other contributors, including Jintao Ren, Zixiang-Wei
(Alan) and Jasper Nijkamp, the latter two of which also supervised the project
as a whole.
The overall goal of the project was to implement a pipeline to preprocess medical scan images of different types to enable supervised deep learning for segmentation on those images. This package comprises the resulting pipeline.
The contributions of the data science students can be generally summarised as follows:
Splitting the conversion script written by Jintao into several restructured ones and exporting scan selections to an Excel-sheet to allow for manual checking and control.
Flagging issues with scans for conversion.
Handling errors and changing decisions for scan selection.
Writing functions to evaluate registration of images based on a metric score.
Exporting metric scores and flagging ones below appropriate threshold.
Easing the overall usage of the different scripts by restructuring them, e.g. with function arguments.
Writing documentation both in and out of code.