The Long, Strange Trip of Medical Images in Clinical Trials

Terminology

CRO: CRO is short for contract resource organization. Pharmaceutical companies typically outsource some or most of the management of their clinical trials to CROs. A subset of clinical trials leverages medical imaging. The three biggest medical imaging CROs are Icon, Clario, and Calyx. Many big pharmaceutical companies will outsource their clinical trials to all three big imaging CROs as a diversification strategy.

eCRF: eCRF is short for electronic case report form. In a clinical trial, once you have the imaging data, you need the imaging read, typically by radiologists. Similar to a report in the clinical world, the read is captured in an eCRF. In the clinical trial world, there are standardized reading protocols like RECIST, which is a protocol that defines how well a cancer patient is responding to treatment…are the tumors growing, staying the same or shrinking? In addition to standardized protocols like RECIST, many eCRFs are customized per trial.

EDC: EDC is short for electronic data capture. This is the clinical trial sister of the EMR in the clinical world. Trial details like sites, subject, eCRFs, consolidated eCRF output, and other relevant clinical trial artifacts are stored in the EDC. It is the system of record.

Group 12 tags: For studies/timepoints that are part of a clinical trial, typically the trial, subject, site, and visit details are captured in the group 12 DICOM tags. The subject is captured in (0012,0040); the sponsor is captured in (0012,0010); the site is captured in (0012,0030); the visit is captured in (0012,0051).

Site: A site is the physical location where the imaging for a subject is done. In one clinical trial, it is not unusual to have hundreds of sites.

Subject: A subject is a patient who participates in a clinical trial.

Timepoint: A timepoint is the clinical data that is captured for a subject at a visit.

Trial: A clinical trial is sometimes called a study. A clinical trial typically attempts to prove the efficacy and safety of a pharmaceutical. At any given time globally, there are about 250k interventional pharmaceutical or biologic studies ongoing, and about 40k new clinical trials per year. 15% to 20% of all clinical trials are imaging trials. Clinical trials go through preclinical, phase I, phase II, phase III, FDA review, and phase IV clinical processes. From molecule to market takes 13 years on average with an average cost of $1B. 90% of these new compounds will fail to make it to market.

Visit: Typically, a subject participating in a clinical trial has a defined number of visits as part of a trial’s protocol. The visits vary per trial, but a normal cadence is baseline, week 4, week 8, week 12.

Workflow

A typical workflow for an imaging clinical trial follows this crazy path (in the really crazy old days, which was 2000 - 2015, most studies were transported to the CRO on CD via courier):

Phase 1 – Getting the Study to the CRO

Subject is imaged
Tech burns the study (MR, CT, etc) to CD or thumb drive
Tech loads the study to an exchange platform. Key attributes of the exchange platform include:
The ability to code the site id, subject id, visit description, and sponsor to the relevant group 12 tags.
The ability to client side de-id other tags as part of doing the upload (i.e. patient name, dob, gender, MRN, all private tags, and a bunch more).
No installation at the site level. Important to remember that one trial can easily have hundreds of sites and doing hundreds of hospital installs is forever and a day.
The ability to check for valid DICOM. Lots of checks, but things like does the preamble start with 128 null bytes followed by DICM, does each file have a study instance uid, etc..
The study is sent to the image management platform at the CRO. Typically, the study is sent via DIMSE DICOM to the image management platform. Unlike the sites, it is normal to have some type of DICOM protocol adapter installed locally for this purpose. The DICOM protocol adapter can do other DICOM transformations while the study is in route.

Phase 2 – Reading the Study/Gathering Objective Evidence that your Pharmaceutical Works and Is Safe

While CROs’ image management platforms have typically used block storage, I am certain that for security and scalability reasons the CROs have moved or are moving their image management platforms to object storage and the three big cloud providers.
In the image management platform, CRO staff does QC on each study. The QC questions are trial specific. The bottom line is that the CRO is getting ready for a central read, and to do this, they are making sure the data is consistent and of a certain quality before moving it through to the central read step. These are checks like (not an exhaustive list):
Does slice thickness meet protocol parameters?
Does spacing between slices meet protocol parameters?
Are the images clear?
Does the right anatomy show?
Merge or delete any series?
Delete non-relevant SOPs/images?
Crop any images?
Is the study instance uid consistent across SOPs?
The CRO gets ready for the central read. A central read is the process of bringing in radiologists or other specialists to read the imaging and complete the associated eCRF (electronic case report form). The CROs will work with the sponsor to structure the central read so that the central read output is supportive of the clinical trial’s hypothesis (that the pharmaceutical is efficacious and safe) and scientifically rigorous so that pharmaceutical can pass the FDA’s (and other regulatory agencies’) clearance process. They make decisions like: 1) anonymized read orders per radiologist, 2) a double blind with adjudication approach, 3) defining fields to capture as part of the read, 4) capturing orthogonal axes of a lesion, and 5) comparing orthogonal axes for target lesions across visits for growth, stability, or shrinkage.
The readers are given training on the protocol.
The readers view the images on a 510(k) cleared viewer and fill out the related eCRF. The study to eCRF relationship is similar to the study to report relationship (typically in Powerscribe 360) in the clinical world. Sometimes the relationship between the viewer and the eCRF in the clinical trial world is more integrated than the viewer to report relationship in the clinical world. For example, if you execute an orthogonal axis measurement in the clinical trial viewer and label it as TL1 (or target lesion 1), the measurements will auto populate the right field in the eCRF. This is pretty sweet.
CROs are actively working on strategies to have the capabilities to plug in the right viewer for a given trial. For example, you might use an advanced volumetric viewer like Mint Medical for a solid tumor oncology trial. You might use a simple web viewer with integrated measurement to eCRF capabilities like Ambra Health for reads that are more 2D.
The output of the central reads is extracted from the eCRFs and loaded into the EDC.
The output of the central reads is packaged up with a bunch of other material and submitted to the FDA or like regulatory agency for review.

Topics for Cloud Providers to Consider

Although each of the big three probably has some type of meaningful presence at most sites that participate in a clinical trial, that is not enough. None of the big three will have presence at all sites that participate in a clinical trial. The act of extracting a study from PACS, sticking it on a thumb drive, and then loading it via a web uploader is better than having to install software that is DIMSE DICOM connected to PACS at hundreds of sites. Getting the studies from the sites to the CRO’s image management platform is a hard problem, and I don’t see the big three’s footprints as much of an advantage for this step in the process.
The above being said, once the images are in the CRO’s image management system, the big three could provide VNA like capabilities to the CROs. After the studies make it into the CRO’s image management system, the workflow steps are specific to CROs, but at a high level, they are pretty similar to imaging stuff that happens in the clinical world. The capabilities should be purpose built, and I believe that the software should either be built out by dedicated software vendors and/or the CROs themselves.
The ability to ‘turn on’ the right viewer per trial is an opportunity. The big three are perfectly positioned to enable such a service.
One of the huge opportunities is that the pharma companies want the data that is in the CROs’ image management systems. This is big data. One MR (or CT) that is part of one visit in one of these trials typically has over 1,000 slices and is 300 MB to 1 GB in size. Today, the CROs have a very hard time moving the data to the pharma companies, and this is a source of friction. The big three should be able to make this a lot easier.
Interestingly, the sponsor pharma companies want access to some of the DICOM data, but not all of it. It takes 13 years on average from IND patent application to FDA clearance. The IND patent lasts for 20 years so the pharma companies have seven years to market and sell the cleared pharmaceutical free and clear. This seven year period is critical to pharma companies’ success, and they do not take risks during this window. One of the strategic ideas that pharma companies have pursued is creating data lakes to facilitate exploratory analysis. Novartis’ Data42 (https://www.novartis.com/stories/data42-program-shows-novartis-intent-go-big-data-and-digital) is a good example of this. However, pharma companies will not make any data tied to a successfully cleared pharmaceutical available for this type of exploratory analysis because if exploratory analysis finds anything that contraindicates the intended use as defined in the regulatory application, this new information must be submitted to the FDA by law and could result in the original clearance being rescinded or modified. The data from failed trials is surprisingly more pliable to pharma’s exploratory analysis efforts than data from successful trials.
In RECIST, the reader will execute an orthogonal axis measurement on a target lesion. The orthogonal axis measurement is the lesion’s length at its longest point and its width at its widest point. It is not a pixel perfect polygon segmentation around the tumor. For those hoping to use RECIST labels to train AI/ML models, too bad. You are going to have to re-label all the lesions.