APR May/June 2023 - 52

» FORMULATION AND DEVELOPMENT
»
medical center is more useful when accompanied by clinical reports
and image annotations.
Such a platform should have ingestion and curation capabilities that
can bring in the necessary datasets efficiently, with audit trails and
repeatable processes. As noted, the datasets involved in R&D are at the
petabyte level, so automating and parallelizing ingestion and curation
wherever possible can help the data-centralization step move faster.
De-identification of data is of course critical, and tools for de-id should
be deployable at the edge of a data platform whenever possible to
minimize risk. Because the types of personal health information (PHI)
at risk of exposure can vary widely from modality to modality, teams
need various de-id tools. Text-based anonymization can be applied to
data in the DICOM header at scale in a repeatable method. Imaging
data can also contain PHI that goes beyond text and hence pixel-based
anonymization techniques (e.g., redaction of burned-in text, " defacing "
of the exterior layer of the skin in brain images) are important to better
ensure compliance.
Where multimodal data is being used, there is an additional layer of
complexity in de-id strategies. Anonymization and PHI removal should
be performed on the imaging data in a way that maintains the ability
to link it with the other types of data pertaining to the same patient.
Once data is centralized in a platform, the metadata embedded
within the images can be extracted. This is most efficiently done with
automated tools that can index large volumes of data via a scalable
method. It is also useful to have the ability to perform validation and
verification on the data. The quality of imaging data from real-world
datasets can vary, so the ability to automate quality control is valuable.
This step also enables teams to correct issues in the data or metadata
early and/or eliminate some of the datasets in the process, helping
prevent problems that can arise from using poor-quality data to train
ML models.
The ultimate goal of these data management strategies is to turn
complex imaging data into analysis-ready datasets that can be used
and reused over time. So, while there may be an initial lift to implement
platforms and tools that can perform these functions, the automation
they can bring to what were previously very manual, time-intensive
procedures can yield dividends.
Imaging R&D Successes in
Life Sciences
Real-world imaging data is utilized heavily in AI/ML initiatives at
pioneering companies and within academia. One top-five pharma
organization has used a data management platform to power the
digital transformation of its R&D department. After centralizing its
imaging assets alongside its legacy data, this organization had more
than 50 million standardized and organized imaging studies. This
data is now accessible to the many members of the enterprise's R&D
52 |
| May/June 2023
teams and can be used and reused without creating duplicate data.
Research teams are now leveraging this data for complex analysis
and machine learning.
In another application of imaging RWD, the University of Wisconsin
Medical School studied 50,000 datasets during the early days of the
COVID pandemic to create an AI model to assist in diagnosing COVID-19
via X-ray. This research united chest X-rays from multiple hospital PACS
with EMR data showing PCR test results. It also used a database of chest
X-rays from the NIH for control data. Researchers efficiently organized
the images and created workflows for labeling and AI training. The
resulting deep neural network was able to differentiate COVID-19
from other types of pneumonia " with performance exceeding that of
experienced thoracic radiologists. "
As these examples illustrate, imaging data can be leveraged to
address both broad organizational goals, as well as pressing public
health emergencies.
Putting Data at Researchers' Fingertips
Data science teams consistently report that the majority of their time
in machine learning projects is spent on mundane data wrangling.
This is especially true when utilizing messy, non-standard imaging
RWD-which, if it can be tamed, is a powerful component in R&D.
Organizations, therefore, need a framework for efficiently incorporating
imaging RWD into their research workflows. This holds the potential for
enabling rich pathways for drug discovery research and accelerating
the development of AI applications. While imaging has previously
been underleveraged due to the complexities of managing the data,
modern tools are finally making it feasible to work with at scale.
For forward-thinking pharma leaders, the implications are clear: the
efficiency with which an organization can handle data management
is a major determining factor in the productivity and the cost of R&D.
The promise that RWD holds for research breakthroughs demands
thoughtful strategies for putting this data within reach for researchers
as efficiently as possible.
Author Biography
Elif Sikoglu is the Senior Director of Life Sciences Solutions
at Flywheel, a biomedical research data platform. She holds
a Ph.D. in biomedical engineering and was previously the
medical director of an imaging CRO. Elif has expertise in
utilizing advanced imaging approaches in neuropsychiatric indications
and significantly contributed to clinical trials with imaging-supported
endpoints across multiple therapeutic areas.
APR May/June 2023

Table of Contents for the Digital Edition of APR May/June 2023
