ohdsi / imagewg

Repository for medical image working group

Medical Imaging Working Group

The Medical Imaging Working Group (MI WG) for the OHDSI community was formed in 2021, comprised of imaging research scientists and observational health researchers familiar with OMOP CDM. The working group evaluated standard vocabularies, defined fields containing key imaging events, and identified limitations of OMOP CDM for imaging observational research.

The working group started with the R-CDM in the development of the medical imaging extension. Imaging researchers across the field were consulted to gather requirements and gain insights into the structure and usability of the proposed model. The principal clinical use case focused on longitudinal tracking of multiple lung nodules. Important attributes included CT acquisition parameters, nodule diameter, location, density, shape, and other phenotypes.

A prototype using CT lung nodules was developed and demonstrated at the 2023 Society of Imaging Informatics in Medicine (SIIM) conference Hackathon.

Please check our Page for more information.

imagewg's Issues

Anatomic Site Concept ID (Entire vs Structure)

In the MI-CDM paper (Fig3), the anatomic_site_concept_id for lung nodule use case was designated as '4118108 (entire thorax)'. If we obtain anatomic site information from DICOM, '4184181 (thoracic cavity structure)' might be more appropriate.
p.s. I am seeking input on this matter to avoid confusion for researchers who are following MI-CDM based on our paper.

Rationale:

• (DICOM) DICOM generally uses anatomic concepts with "structure" rather than "entire". For instance, the BodyPartExamined Tag 'CHEST' is defined as SCT '43799004 (thoracic cavity structure)'.
• (OMOP CDM) Both codes are included as Standard Concept IDs in the Specimen Table's Accepted Concepts, but 'Entire' lacks descendants compared to 'Structure'.

Concept Type	OMOP Concept ID	SCT Code	Description
Structure	4184181	43799004	Thoracic cavity structure (body structure)
Entire	4118108	302551006	Entire thorax

References:

• DICOM PS3.16 > Correspondence of Anatomic Region Codes and Body Part Examined Defined Terms
• OMOP CDM v5.4 > Specimen Table > ETL Conventions

Anatomic Site Concept ID: BodyPartExamined (0018,0015) vs AnatomicRegionSequence (0008,2218)

For future automation of ETL processes, we may need to select appropriate standard tags as sources. Attributes like 'BodyPartExamined' and 'AnatomicRegionSequence' contain information about the anatomic site of the imaging.

Key Points:

• (DICOM Standard) According to DICOM PS3.16, 'AnatomicRegionSequence' defines a more diverse and detailed Context Group compared to 'BodyPartExamined', which reflects historical or clinically well-recognized usage.
• Insights from our institution's sample data indicate 'BodyPartExamined' captured more information for Chest CT and Brain MR, but results may vary across institutions and modalities.

Modality	Captured Values for 'BodyPartExamined'	Captured Values for 'AnatomicRegionSequence'
Brain MR	'BRAIN', 'NECKCHESTABDOMEN', 'HEAD', nan	{Code Value: T-D1100, Code Meaning: HEAD}, nan
Chest CT	'CHEST', 'ABDOMEN', nan	nan

Discussion Point:

• I am curious about how this appears in other institutions and different imaging studies.

References:

• DICOM PS3.16 > Correspondence of Anatomic Region Codes and Body Part Examined Defined Terms

ETL of DICOM Header: Raw vs Cleansed Data

When ETL-ing DICOM headers, should we use raw data or cleanse it before uploading?

Considerations:

1. Modality Concept ID: Across similar medical images, the 'Modality' tag (0008,0060) exhibits diverse representations.

Modality	Captured Values for 'Modality (0008,0060)' Tag
Mammography	MG (Mammography), CR (Computed Radiography)
Chest X-ray	CR (Computed Radiography), DR (Digital Radiography), DX (Digital X-ray)

For instance, reviewing sample data from Korea, we observed a mixture of MG and CR in Mammography studies. Similarly, Chest X-ray studies showed a mix of CR, DR, and DX.

2. Errors in Raw Data: Instances where 'Body Part Examined' in Chest X-ray headers was reported as 'Skull'.

Efficiency of Image_occurrence Table (Series-level)

Although previously discussed, I would like to revisit the decision to structure the Image_occurrence table at the series-level. When structured at the series-level, each DICOM Study typically contains numerous Series (varies by modality), resulting in multiple identical rows in the Image_occurrence table distinguished only by 'Image_series_UID'.

Discussion Points:

• When institutions are deciding whether to introduce an extension table in addition to the existing CDM table, we understand that DB storage is also considered. If there is no specific reason to maintain it at the Series-level, could we consider switching to study-level to eliminate duplicate rows?
• Additionally, if maintained at series-level, should we include information to uniquely identify contents of series (SOP Class UID or IOD — e.g. CT Image Storage, MR Image Storage, Secondary Capture Image Storage, ...)? As you know, through the required tag (Type 1) SOP Class UID (0008,0016), each DICOM file can identify the mandatory tags for each IOD.

References:

• DICOM PS3.4 > B.5 Standard SOP Classes