3D imaging EDA recap

Congrats! You’ve completed the lesson. I have introduced you to the DICOM standard and we got into some of the deep dark internals of the standard. Hopefully, now you have a good basis to start digging around datasets and are ready to try your hand at some machine learning techniques that we will discuss in the next lesson.

In this lesson we have covered the following topics:

• Basics of DICOM and NIFTI file formats
• We did a short introduction into viewing 3D medical images and some of the tools that you can use for viewing. This was followed by an exercise where you loaded a 2D DICOM image and converted it into PNG, applying proper windowing transform.
• Some important parameters to consider when analyzing datasets. This was followed by an exercise where you loaded a 3D volume and extract slices from it.
• A walkthrough of a Notebook with an approach to doing exploratory data analysis on a single 3D CT volume, and then an EDA for a dataset of multiple volumes. After that, you worked on another exercise and worked out an approach to an EDA of your own, curating a “dirty” DICOM volume dataset.

Further Resources

Some great examples of EDA adventures could be found in the following competitions/challenges:

• The 2019 RSNA Intracranial Hemorrhage Detection Challenge
• 2018 RSNA Pneumonia Detection Challenge
• 2017 Data Science Bowl
• Multimodal Brain Tumor Segmentation Challenge 2019

And here are some notebooks presenting interesting tricks to try with DICOM data:

• As mentioned in the previous lesson, now that you know some DICOM, take a look at some creating windowing approaches: https://www.kaggle.com/jhoward/don-t-see-like-a-radiologist-fastai
• This one provides a good basis for windowing responsibly https://www.kaggle.com/dcstang/see-like-a-radiologist-with-systematic-windowing
• And this Notebook covers some useful tricks to keep handy as you are working through 3D imaging datasets: https://www.kaggle.com/jhoward/some-dicom-gotchas-to-be-aware-of-fastai

Vocabulary

• DICOM - Digital Imaging and Communication in Medicine. The standard defining the storage and communication of medical images.
• DICOM Information Object - representation of a real-world object (such as an MRI scan) per DICOM standard.
• IOD - Information Object Definition. Definition of an information object. Information Object Definition specifies what metadata fields have to be in place for a DICOM Information Object to be valid. IODs are published in the DICOM standard.
• Patient - a subject undergoing the imaging study.
• Study - a representation of a “medical study” performed on a patient. You can think of a study as a single visit to a hospital for the purpose of taking one or more images, usually within. A Study contains one or more series.
• Series - a representation of a single “acquisition sweep”. I.e., a CT scanner took multiple slices to compose a 3D image would be one image series. A set of MRI T1 images at different axial levels would also be called one image series. Series, among other things, consists of one or more instances.
• Instance - (or Image Information Entity instance) is an entity that represents a single scan, like a 2D image that is a result of filtered backprojection from CT or reconstruction at a given level for MR. Instances contain pixel data and metadata (Data Elements in DICOM lingo).
• SOP - Service-Object Pair. DICOM standard defines the concept of an Information Object, which is the representation of a real-world persistent object, such as an MRI image (DICOM Information Objects consist of Information Entities).
• Data Element - a DICOM metadata “field”, which is uniquely identified by a tuple of integer numbers called group id and element id.
• VR - Value Representation. This is the data type of a DICOM data element.
• Data Element Type - identifiers that are used by Information Object Definitions to specify if Data Elements are mandatory, conditional or optional.
• NIFTI - Neuroimaging Informatics Technology Initiative, is an open standard that is used to store various biomedical data, including 3D images.