Accelerating AI-Driven Imaging Triage for Faster Time to Care
Customer
Harrison.ai builds next-generation medical imaging
AI tools to power clinicians and ensure that no patient outcome depends on the capacity limits of the
healthcare system.
The Clinical Problem: Imaging Backlogs and Time-to-Read Delays
Radiology departments around the world are facing mounting pressure. Imaging volumes continue to rise year over year, while radiologist capacity struggles to keep pace. This imbalance has made effective prioritization one of the most pressing challenges in medical imaging today.
Clinical literature consistently shows that delays between image acquisition and radiologist review are associated with delayed diagnosis and treatment, particularly for acute and high-risk conditions. Triage tools have emerged as a critical lever—helping to ensure that urgent cases surface quickly in increasingly crowded worklists, rather than waiting behind routine exams.
For Harrison.ai, solving this problem meant developing an AI-driven triage algorithm capable of identifying urgent cases and automatically elevating them to the top of the radiologist’s workload.
While triage algorithms do not directly deliver treatment, their impact is profound: by shortening time-to-read, they indirectly accelerate downstream clinical decisions and patient care.
The Hidden Cost of “Available” Imaging Data
To validate this triage algorithm for clearance in the United States, Harrison.ai needed large volumes of high-quality imaging data. The team initially procured data from two imaging data vendors—but quickly ran into quality and quantity issues.
Key issues included:
Data corruption and mislabeling
Images that appeared successfully ingested later failed during analysis—such as studies labeled as chest imaging that were actually from the wrong anatomical region.Hidden quality defects
Pixel-level corruption that only surfaced during manual inspection or radiologist case reads, forcing repeated rework and significant time and labor costs.
“I spent a full Saturday spot-checking images for pixel corruption, and months later I’m still replacing bad data.”
Limited volume of positive cases
Enriching the data set with imaging studies that are positive for a specified finding required identifying sufficient volume of cases containing the rare critical finding that also met all precise technical inclusion criteria. Other vendors had cases that contained the finding or met the technical criteria, but not both.
Risk of unintended bias
When hundreds of cases had to be removed post-ingestion, the team faced a new concern:
Were the discarded cases disproportionately coming from a single hospital or system?
Were they inadvertently introducing bias into the training set?
Despite having already invested budget and effort into two vendors, the team lacked confidence in the integrity and representativeness of the data—creating risk for both model performance and regulatory submission.
The Turning Point: Partnering with Avandra
When Harrison.ai turned to Avandra, the experience fundamentally changed.
Avandra provided clean, complete, and regulatory-grade imaging data, purpose-built for advanced AI validation. Unlike marketplace-style or self-serve vendors, Avandra operates as a true imaging data partner—working directly with health systems and owning the infrastructure required to deliver high-fidelity data at scale. The difference was clear:
Curated, Not Crowdsourced
Avandra’s imaging data is sourced through trusted health system partnerships and undergoes rigorous curation. Studies arrive correctly labeled, anatomically consistent, and ready for analysis—eliminating the need for downstream triage of the data itself.Expert-Led Quality Assurance
Rather than placing the QA burden on Harrison.ai’s team, Avandra applied multi-layer quality checks across ingestion, de-identification, and delivery. The result: confidence that data quality wasn’t dependent on a single set of eyes but validated by a dedicated team with deep imaging and clinical expertise.
Reduced Bias Risk
Because Avandra delivers complete, balanced datasets, Harrison.ai no longer had to remove large subsets post hoc—mitigating concerns about skewed representation from specific sites or systems.
Operational Efficiency
Clean data meant Harrison’s engineers and clinicians could redirect time away from manual review and remediation—and back toward algorithm development, validation, and performance optimization.
Built for Regulatory Readiness
With an FDA submission planned this Spring, data quality was non-negotiable. Avandra’s provenance, consistency, and documentation provided confidence that the data met the standards required for regulatory review.
“Even after contracting with two other vendors, the difference was clear: when we got data from Avandra, it was clean and complete.”
The Impact: Faster Development, Stronger Confidence, Clearer Path to FDA
With Avandra as a data partner, Harrison.ai was able to move forward with confidence—at a critical moment in their development timeline.
Key outcomes included:
Accelerated validation of an AI algorithm intended for triage of chest CTs
Reduced manual QA effort and internal resource drain
Increased confidence in dataset integrity and representativeness
Stronger readiness for FDA submission, supported by high-quality data
Most importantly, this enabled Harrison.ai to stay focused on their mission:
Delivering solutions to help radiologists identify urgent imaging cases faster—
so patients can receive timely care when it matters most.
