dbdicom 0.1.6

Reading and writing DICOM databases

Installation

Run pip install dbdicom.

Summary

The DICOM format is the universally recognised standard for medical imaging, but reading and writing DICOM data remains a challenging task for most data scientists.

The excellent python package pydicom is very commonly used and well-supported, but it is limited to reading and writing individual files, and still requires a fairly high level of understanding of DICOM to ensure compliance with the standard.

dbdicom wraps around pydicom to provide an intuitive programming interface for reading and writing data from entire DICOM databases, replacing unfamiliar DICOM-native concepts by language and notations that will be more familiar to data scientists.

The sections below list some basic uses of dbdicom. The package is currently deployed in several larger scale multicentre clinical studies led by the authors, such as the iBEAt study and the AFiRM study. The package will continue to be shaped through use in these studies and we expect it will attain a more final form when these analysis pipelines are fully operational.

Browsing a DICOM folder

Reading and opening a DICOM database

Open a DICOM database in a given folder, and print a summary of the content:

import dbdicom as db

database = db.database('C:\\Users\\MyName\\MyData\\DICOMtestData')
database.print()

The first time the database is opened this will be relatively slow because all files need to be read and summarized. If the folder is reopened again later, the table can be read directly and opening will be much faster.

Use scan() to force a rereading of the database. This may be of use when files have become corrupted, or have been removed/modified by external applications:

database.scan()

After making changes to the DICOM data, the folder should be closed properly so any changes can be either saved or rolled back as needed:

database.close()

If unsaved changes exist, close() will prompt the user to either save or restore to the last saved state.

Retrieving objects from the folder

A DICOM database has a hierarchical structure.

database/
|
|---- Patient 1/
|    |
|    |---- Study 1/
|    |     |
|    |     |---- Series 1/
|    |     |    |----Instance 1
|    |     |    |----Instance 2
|    |     |    |----Instance 3
|    |     |    
|    |     |----Series 2/
|    |    
|    |---- Study 2/
|
|---- Patient 2/  
| 

A patient can be an actual patient but can also be a healthy volunteer, an animal, a physical reference object, or a digital reference object. Typically a study consist of all the data derived in a single examination of a subject. A series usually represents and individual examination in a study, such an MR sequence. The files contain the data and are instances of real-world objects such as images or regions-of-interest.

To return a list of all patients, studies, series or instances in the folder:

instances = database.instances()
series = database.series()
studies = database.studies()
patients = database.patients()

The same functions can be used to retrieve the children of a certain parent object. For instance, to get all studies of a patient:

studies = patient.studies()

Or all series under the first of those studies:

series = studies()[0].series()

Or all instances of a study:

instances = study.instances()

And so on for all other levels in the hierarchy. These functions also work to find objects higher up in the hierarchy. For instance, to find the patient of a given series:

patient = series.patients()

In this case the function will return a single item.

Finding DICOM objects in the folder

Each DICOM file has a number of attributes describing the properties of the object. Examples are PatientName, StudyDate, etc. A convenient list of attributes for specific objects can be found here:

Each known attribute is identified most easily by a keyword, which has a capitalised notation. Objects in the folder can be can also be listed by searching on any DICOM tag:

instances = database.instances(PatientName = 'John Dory')

This will only return the instances for patient John Dory. This also works with multiple DICOM tags:

series = database.instances(
    PatientName = 'John Dory', 
    ReferringPhysicianName = 'Dr. No', 
)

In this case objects are only returned if both conditions are fullfilled. Any arbitrary number of conditions can be entered, and higher order objects can be found in the same way:

studies = database.studies(
    PatientName = 'John Dory', 
    ReferringPhysicianName = 'Dr. No', 
)

As an alternative to calling explicit object types, you can call children() and parent to move through the hierarchy:

studies = patient.children()
patient = studies[0].parent

The same convenience functions are available, such as searching by keywords:

studies = patient.children(ReferringPhysicianName = 'Dr. No')

Moving and removing objects

To remove an object from the folder, call remove() on the object:

study.remove()
instance.remove()

remove() can be called on Patient, Study, Series or Instances.

Moving an object to another parent can be done with move_to(). For instance to move a study from one patient to another:

study = folder.patients()[0].studies()[0]
new_parent = folder.patients()[1]
study.move_to(new_parent)

Copying and creating objects

Any object can be copied by calling copy():

study = folder.patients()[0].studies()[0]
new_study = study.copy()

This will create a copy of the object in the same parent object, i.e. study.copy() in the example above has created a new study in patient 0. This can be used for instance to copy-paste a study from one patient to another:

study = folder.patients()[0].studies()[0]
new_parent = folder.patients()[1]
study.copy().move_to(new_parent)

This is equivalent to using copy_to():

study.copy_to(new_parent)   

Instead of copying, and object can also be moved:

study.move_to(new_parent)   

To create a new object, call new_child() on the parent:

series = study.new_child()

series will now be a new (empty) series under study. This can also be written more explicitly for clarity:

series = study.new_series()

And equivalently for new_patient, new_study and new_instance. New sibling objects under the same parent can be created by:

new_series = series.new_sibling()

here new_series will be a series under the same study as series. Objects higher up the hierarchy can be created using new_pibling (i.e. sibling of the parent):

new_study = series.new_pibling()

This is shorthand for:

new_study = series.parent().new_sibling()

When new objects are created, they can be assigned properties up front, for instance:

new_study = series.new_pibling(
    StudyDescription='Parametric maps',
    StudyDate = '12.12.22')

This will ensure that all data that appear under the new study will have these attributes.

Export and import

To import DICOM files from an external folder, call import_dicom() on a database with a list of files:

database.import_dicom(files)

To export dicom datasets out of the folder to an external folder, call export_as_dicom() on any dicom object with the export path as argument:

series.export_as_dicom(path)

Exporting in other formats is similar:

study.export_as_csv(path)
study.export_as_nifti(path)
study.export_as_png(path)

The pixel data from a series can also be exported in numpy format:

series.export_as_npy(path)

This exports the array in dimensions (n,x,y) where n enumerates the images and x,y are the pixels. To export in different dimensions use the sortby keyword with one or more DICOM tags:

series.export_as_npy(path, sortby=['SliceLocation','AcquisitionTime'])

This exports an array with dimensions (z,t,n,x,y) sorted by slice location and acquisition time.

Creating and modifying DICOM files

Reading DICOM attributes

An object's DICOM attributes can be read by using the DICOM keyword of the attribute:

nr_of_rows = instance.Rows

All attributes can also be accessed at series, study, patient or folder level. In this case they will return a list of unique values. For instance to return a list with all distinct series descriptions in a study:

desc = study.SeriesDescription

DICOM attributes can also be accessed using the list notation, using either the keyword as a string or a (group, element) pair:

columns = instance['Columns']
columns = instance[(0x0028, 0x0010)]

The tags can also be accessed as a list, for instance:

dimensions = ['Rows', (0x0028, 0x0010)]
dimensions = instance[dimensions] 

This will return a list with two items. As shown in the example, the items in the list can be either KeyWord strings or (group, element) pairs. This also works on higher-level objects:

dimensions = ['Rows', (0x0028, 0x0010)]
dimensions = patient[dimensions] 

Editing attributes

DICOM tags can be modified using the same notations:

instance.EchoTime = 23.0

or also:

instance['EchoTime'] = 23.0

or also:

instance[(0x0018, 0x0081)] = 23.0

Multiple tags can be inserted in the same line:

shape = ['Rows', 'Columns']
instance[shape] = [128, 192]

When setting values in a series, study or patient, all the instances in the object will be modified. For instance, to set all the Rows in all instances of a series to 128:

series.Rows = 128

Custom attributes

Apart from the predefined public and private DICOM keywords, dbdicom also provides a number of custom attributes for more convenient access to higher level properties. In order to distinguish these from existing DICOM attributes which are defined in CamelCase, the custom attributes follow the lower_case notation.

For instance, to set one of the standard matplotlib color maps, you can do:

image.colormap = 'YlGnBu'
series.colormap = 'Oranges'

and so on.. The colormaps can be retrieved the same way:

cm_image = image.colormap
cm_series = series.colormap

As for standard DICOM attributes this returns a list if unique values for the series.

Custom attributes can easily be added to any DICOM dataset type and the number of available attributes is set to grow as the need arises.

Read and write

By default all changes to a database are made on disk. For instance if a DICOM attribute is changed

instance.Rows = 128

The data are read from disk, the change is made, the data are written to disk again and memory is cleared. Equally, if a series is copied to another study, all its instances will be read, any necessary changes made, and then written to disk and cleared from memory.

For many applications reading and writing from disk is too slow. For faster access at the cost of some memory usage, the data can be read into memory before performing any manipulations:

series.read()

After this all changes are made in memory. To clear the data from memory and continue working from disk, use clear():

series.clear()

These operations can be called on the entire database, on patients, studies, series or instances.

Save and restore

All changes made in a DICOM folder are reversible until they are saved. To save all changes, use save():

database.save()

This will permanently burn all changes that are made on disk. In order to reverse any changes made, use restore() to revert back to the last saved state:

database.restore()

This will roll back all changes on disk to the last changed state. save() and restore() can also be called at the level of individual objects:

series.restore()

will reverse all changes made since the last save, but only for this series. Equivalently:

series.save()

will save all changes made in the series (but not other objects in the database) permanently.

Working with series

A DICOM series typically represents images that are acquired together, such as 3D volumes or time series. Some dedicated functionality exists for series that is not relevant for objects elsewhere in the hierarchy.

To extract the images in a series as a numpy array, use array():

array, _ = series.array()

This will return an array with dimensions (n,x,y) where n enumerates the images in the series. The array can also be returned with other dimensions:

array, _ = series.array(['SliceLocation', 'FlipAngle'])

This returns an array with dimensions (z,t,n,x,y) where z corresponds to slice locations and t to flip angles. The 3d dimension n enumerates images at the same slice location and flip angle. Any number of dimensions can be added in this way. If an application requires the pixels to be listed first, use the pixels_first keyword:

array, _ = series.array(['SliceLocation', 'FlipAngle'], pixels_first=True)

In this case the array has dimensions (x,y,z,t,n). Replacing the images of a series with a given numpy array works the same way:

series.array(array)

The function array() also returns the header information for each slice in a second return value:

array, header = series.array(['SliceLocation', 'FlipAngle'])

The header is a numpy array of instances with the same dimensions as the array - except for the pixel coordinates: in this case (z,t,n). This can be used to access any additional data in a transparent way. For instance, to list the flip angles of the first slice z=0, n=0:

FA = [hdr.FlipAngle for hdr in header[0,:,0]]

The header array is also useful when a calculation is performed on the array and the results need to be saved in the DICOM database again. In this case header can be used to carry over the metadata.

As an example, let's calculate a maximum intensity projection (MIP) of a 4D time series and write the result out in the same series:

array, header = series.array(['SliceLocation', 'AcquisitionTime'])
mip = np.amax(array, axis=0)
series.set_array(mip, header[0,:,:])

In this case the header information of the MIP is taken from the first image of the time series. Provding header information is not required - if the header argument is not specified then a template header is used.

Another useful tool on series level is extracting a subseries. Let's say we have an MRI series with phase and magnitude data mixed, and we want to split it up into separate series:

phase = series.subseries(image_type='PHASE')
magn = series.subseries(image_type='MAGNITUDE')

This will create two new series in the same study. The image_type keyword is defined in dbdicom for MR images to simplify access to phase or magnitude data, but the method also works for any standard DICOM keyword, or combinations thereof. For instance, to extract a subseries of all images with a flip angle of 20 and a TR of 5:

sub = series.subseries(FlipAngle=20, RepetitionTime=5)

Another useful feature at series level is to overlay one series on another.

overlay = series.map_to(target)

If series is a binary mask (or can be interpreted as one), a similar function can be used to overlay the mask on another series:

overlay = series.map_mask_to(target)

Creating DICOM data from scratch

To create a DICOM series from a numpy array, use dbdicom.series():

import numpy as np
import dbdicom as db

array = np.random.normal(size=(10, 128, 192))
series = db.series(array)

After this you can save it to a folder in DICOM, or set some header elements before saving:

series.PatientName = 'Random noise'
series.StudyDate = '19112022'
series.AcquisitionTime = '120000'
series.save(path)

You can build an entire database explicitly as well. For instance, the following code builds a database with two patients (James Bond and Scarface) who each underwent and MRI and an XRay study:

database = db.database()

james_bond = database.new_patient(PatientName='James Bond')
james_bond_mri = james_bond.new_study(StudyDescription='MRI')
james_bond_mri_localizer = james_bond_mri.new_series(SeriesDescription='Localizer')
james_bond_mri_T2w = james_bond_mri.new_series(SeriesDescription='T2w')
james_bond_xray = james_bond.new_study(StudyDescription='Xray')
james_bond_xray_chest = james_bond_xray.new_series(SeriesDescription='Chest')
james_bond_xray_head = james_bond_xray.new_series(SeriesDescription='Head')

scarface = database.new_patient(PatientName='Scarface')
scarface_mri = scarface.new_study(StudyDescription='MRI')
scarface_mri_localizer = scarface_mri.new_series(SeriesDescription='Localizer')
scarface_mri_T2w = scarface_mri.new_series(SeriesDescription='T2w')
scarface_xray = scarface.new_study(StudyDescription='Xray')
scarface_xray_chest = scarface_xray.new_series(SeriesDescription='Chest')
scarface_xray_head = scarface_xray.new_series(SeriesDescription='Head')

Work in progress: a numpy-like interface

We are currently building a numpy-type interface for creating new DICOM objects. For instance to create a new series with given dimensions in a study you can do:

img = study.zeros((10, 128, 192), dtype='mri')

This will create a DICOM series of type 'MRImage' (shorthand 'mri') with 10 slices of 128 columns and 192 rows each. This can also be done from scratch:

import dbdicom as db

series = db.series((10, 128, 192))

Currently, writing in data types other than 'MRImage' is not supported, so the data type argument is not necessary.

User interactions

dbdicom can be used in standalone scripts or interactively. To streamline integration in a GUI, communication with the user is performed via two dedicated attributes status and dialog. dialog and status attributes are available to any DICOM object. The status attribute is used to send messages to the user, or update on progress of a calculation:

series.message("Starting calculation...")

When operating in command line mode this will print the message to the terminal. If dbdicom is used in a compatible GUI, this will print the same message to the status bar. Equivalently, the user can be updated on the progress of a calculation via:

for i in range(length):
    series.progress(i, length, 'Calculating..)

This will print the message with a percentage progress at each iteration. When used in a GUI, this will update the progress bar of the GUI.

By default a dbdicom record will always update the user on progress of any calculation. When this beaviour is undersired, the record can be muted as in via series.mute(). After this the user will no longer recieve updates. In order to turn messages back on, unmute the record via series.unmute().

Dialogs can be used to send messages to the user or prompt for input. In some cases a dialog may halt the operation of te program until the user has performed the appropriate action, such as hitting enter or entering a value. In command line operator or scripts the user will be prompted for input at the terminal. When using in a GUI, the user will be prompted via a pop-up:

series.dialog.question("Do you wish to proceed?", cancel=True)

When used in a script, this will ask the user to enter either "y" (for yes), "n" (for no) or "c" (for cancel) and the program execution will depend on the answer. When the scame script is deployed in a GUI, the question will be asked via a pop-up window and a button push to answer. A number of different dialogs are available via the dialog attribute (see reference guide).

About dbdicom

Why DICOM?

``[...] after 2 hours of reading, I still cannot figure out how to determine the 3D orientation of a multi-slice (Supplement 49) DICOM file. I'm sure it is in there somewhere, but if this minor factoid can't be deciphered in 2 hours, then the format and its documentation is too intricate.''. Robert W. Cox, PhD, Director, Scientific and Statistical Computing Core, National Institute of Mental Health link.

This echoes a common frustration for anyone who has ever had a closer to look at DICOM. DICOM seems to make simple things very difficult, and the language often feels outdated to modern data scientists.

But there are good reasons for that. DICOM not only retains imaging data, but also all other relevant data about the subject and context in which the data are taken. Detailing provenance of the data and linkage to other data is critical in radiology, but the nature of these meta data is very broad, complex and constantly changing. Storing them in some consistent and standardised way that is future proof therefore requires a systematic approach and some necessary level of abstraction.

DICOM does this well and has for that reason grown to be the single accepted standard in medical imaging. This also explains the outdated look and feel. DICOM standardises not only the format, but also the language of medical imaging. And successful standards, by definition, don't change.

Why dbdicom?

Reading and especially writing DICOM data remains a challenging enterprise for the practicing data scientist. A typical image processing pipeline might use the excellent python package pydicom for extracting image arrays and any required header information from DICOM data, but will then write out the results in more manageable format such as nifty. In the process the majority of header information will have to be discarded, including detailed imaging parameters and linkage between original and derived images, follow-up studies, etc.

The practice of converting outputs in a lossy image format may be sufficient in the early stages of method development, but forms a major barrier to research or deployment of these processing methods in a real-world context. This requires results in DICOM format so they can be linked to other data of the same patients, integrated in the radiological workflow, and reviewed and edited through integrated radiological viewers. Integration of datasets ensures that all derived data are properly traceable to the source, and can be compared between subjects and within a subject over time. It also allows to test for instance whether a new (expensive) imaging method provides an additive benefit over and above (cheap) data from medical history, clinical exams or blood tests.

DICOM integration of processing outputs is typically performed by DICOM specialists in the private sector, for new products that have proven clinical utility. However, this requires a major separate investment, delays the point of real-world validation until after commercialisation and massively increases the risk of costly late-stage failures.

What is dbdicom?

dbdicom is a programming interface that makes reading and writing DICOM data intuitive for the practicing medical imaging scientist working in Python. DICOM-native language and terminology is hidden and replaced by concepts that are more natural for those developing in Python. The documentation therefore does not reference confusing DICOM concepts such as composite information object definitions, application entities, service-object pairs, unique identifiers, etc.

dbdicom wraps around DICOM using a language and code structure that is native to the 2020's. This should allow DICOM integration from the very beginning of development of new image processing methods, which means they can be deployed in clinical workflows from the very beginning. It also means that any result you generate can easily be integrated in open access DICOM databases and can be visualised along with any other images of the same subject with a standard DICOM viewer such as OHIF.

dbdicom is developed by through the UKRIN-MAPS project of the UK renal imaging network, which aims to provide clinical translation of quantitative renal MRI on a multi-vendor platform. UKRIN-MAPS is funded by the UK's Medical Research Council.

Acknowledgements

dbdicom relies heavily on pydicom for read/write of individual DICOM files, with some additional features provided by nibabel and dcm4che. Basic array manipulation is provided by numpy, and sorting and tabulating of data by pandas. Export to other formats is provided by matplotlib.