CARPL
CARPL is a vendor neutral platform that lets radiology departments and imaging groups buy, test, deploy and monitor other companies' imaging AI through one integration and one viewer. Its marketplace lists more than 300 applications from over 100 partners under a single procurement channel. A gateway on the hospital network accepts studies, identifies series automatically and routes them by configured rules, and the Universal AI Viewer shows findings from every application side by side, where radiologists accept, reject or amend them before approved outputs return to PACS as secondary captures, structured reports, GSPS or DICOM segmentations. FDA cleared the viewer and workflow device in 2024 under K232891, while each connected model keeps its own clearance.
The Test and Monitor workspace compares models on a hospital's own reference cases, with labels uploaded or extracted from free text reports by a built in language model. One click reports cover sensitivity, specificity, confusion matrices and ROC curves, with performance by age, sex, body mass index and scanner, and drift tracking after go live. PACS partners include Agfa, Fujifilm, Intelerad, RamSoft and Paxera.
CARPL says four of the world's five largest private radiology groups use it, among them RadNet and DeepHealth, alongside University Hospitals in Cleveland, I-MED, Fleury, QScan, Massachusetts General Hospital and government systems in Brazil, India, Singapore, Spain and the UAE. Chief executive Vidur Mahajan started it out of Mahajan Imaging, and the Cupertino based company raised a $10 million Series A led by the International Finance Corporation in July 2026.
Capability Axes
An AI Health Index grade measures what a buyer can verify from public sources on the date shown. It is not a rating of how good the product is. A vendor can build an excellent system and grade low on an axis because it publishes nothing an outsider can check. How grades read
CARPL is mainly a marketplace and control layer for other companies' imaging models. It routes studies, compares model performance and presents partner results for radiologists to review in one viewer, and it adds workflow features such as report auto population, triage and urgency flagging built on those partner outputs.
Its own learned component is the language model in the validation workspace, which extracts reference findings from free text radiology reports so hospitals can audit models without labeling cases by hand, though they can supply labels directly instead. The image interpretation and clinical flags come from the connected models, which each hospital chooses and licenses.
Hospitals set routing rules for incoming studies and choose which AI outputs the viewer presents, and radiologists accept, reject or change each finding and can switch configured outputs on or off. Before an application goes live, validation pairs its outputs with reference findings or radiologist opinions for review case by case, and after rollout monitoring flags changes in performance.
Routing can start automatically as images arrive or be triggered by a user. What each model does with a case is set by that model and its configuration, so the platform's oversight sits around the models rather than inside them.
The validation workflow uses a language model that extracts findings from narrative reports, and the source text and its use as reference data are described. The FDA filing explains routing by DICOM metadata and how the viewer handles partner outputs, and the validation documentation sets out the statistics used to compare tools and track drift, from confusion matrices to divergence measures.
The extraction model is not named, and its training material and observed error rate are not given. Diagnostic model documentation and version details come from whichever marketplace application the hospital selects, so transparency varies by vendor.
Marketplace listings identify the vendor behind every application a hospital can select, and hospitals can bring internal models into the same workflow. HOPPR supplies foundation models through a separate collaboration for customers building and tuning their own applications, PaxeraHealth provides a model authoring environment, and DeepHealth sits on both sides as a partner and, with RadNet, a customer.
The language model inside validation that extracts report findings has no named provider or version. The partner inventory covers the commercial marketplace, but no list maps the model operators, dependencies and update schedules that apply to a particular hospital's deployment.
Named adoption is broad. University Hospitals in Cleveland signed a multiyear validation and deployment agreement through its RadiCLE program, RadNet and DeepHealth chose CARPL as their AI platform partner, QScan runs a lung cancer screening program on it, and the company says four of the world's five largest private radiology groups are customers. The Algorithmic Audit paper sets out its method for evaluating third party models, and its research library compares imaging tools.
Those comparisons measure the selected models. No study measures what operating CARPL itself changes, such as reading time, missed findings or caught model failures, and the report extraction feature's accuracy is not reported.
The validation workspace deidentifies DICOM studies before use, and the gateway can strip identifying information from tags or from the image itself before a study reaches a partner model. Role based access control and single sign on limit who sees what, the privacy policy treats medical records as sensitive information, and radiologists can always read the image independently of any generated result.
How data flows to a given marketplace partner depends on the application and its routing. Whether the report extraction model can use submitted text for training, and how customers set retention and deletion across partners, are not stated.
CARPL describes HIPAA compliant DICOM deidentification in validation and publishes a privacy policy that treats medical records as sensitive information. In production, though, a connected application may receive identifiable imaging data through its configured route, and a marketplace listing or FDA status does not show which company handles protected health information on a hospital's behalf.
The business associate agreement language, when CARPL signs one and whether it binds each routed application and the report extraction provider are not stated. The Health AI Register reports a separate data agreement for the marketplace.
Chief executive Vidur Mahajan reports ISO/IEC 27001:2022 certification for the platform, though no current certificate showing the covered entity, systems and dates is available publicly. The platform offers single sign on, role based access control and gateway deidentification of incoming studies, and local deployment for institutions that keep imaging data on site.
Connected applications apply their own controls, and the ISO claim does not set out whether it covers the extraction model, the partner applications or their hosting. No SOC 2 report or penetration testing cadence is named, which leaves hospitals relying on the ISO statement and their own review of each partner.
FDA cleared CARPL's web based PACS and radiology workflow management device under K232891, covering image viewing and assessment in specified modalities and the gateway and viewer workflow. The filing treats connected models as third party systems, each governed by its own FDA record and the hospital's use of it, and the company markets itself as the largest FDA cleared radiology AI marketplace.
The clearance does not name the language model that extracts reference findings for validation as a cleared function. That feature supports evaluation rather than diagnosis and has no separate indication in the published description.
Governance of other vendors' models is the product's core. The validation workspace runs models against local reference findings, lets reviewers inspect false positives and negatives case by case, and applies an Aequitas based parity analysis by age, sex, body mass index and scanner type. One click reports summarize sensitivity, specificity, confusion matrices and ROC curves for internal committees, and monitoring calculates divergence measures to catch drift after deployment. Tan Tock Seng Hospital used it to validate chest X ray and thyroid models on one platform.
The same subgroup results and a monitoring protocol are not described for CARPL's own report extraction model.
Errors are visible at two stages. In the viewer, radiologists inspect, reject or revise partner results before they reach PACS, and during validation, governance staff examine every false positive and negative against the reference set.
The public terms disclaim responsibility for third party resources and make no promise to correct errors in outputs, and they cap aggregate liability at fees paid or payable over the preceding 12 months, to the extent the law allows. A separate subscription agreement can add terms. No extraction accuracy commitment, turnaround for model corrections or compensation for a missed finding appears in public materials.
The FDA submission describes the local gateway collecting DICOM studies from an imaging system and sending secondary capture images back to the connected PACS, and outputs can also return as structured reports, DICOM PDF, GSPS and segmentation objects. RamSoft's integration covers PowerServer and OmegaAI, Intelerad includes CARPL in its InteleGence ecosystem, and Agfa, Fujifilm, Paxera and Adva PACS appear as partners. Validation accepts browser uploads, API transfers and DICOM push.
No public case pairs a named institution with a named PACS and documents writeback in production, and connections reach the imaging record rather than the EHR.
The imaging gateway sits on the customer side of the modality and PACS network. The rest of the deployment can run on premises or on any cloud, and a local arrangement keeps images inside the institution for those that require it, with hybrid and virtualized local configurations also listed in the Health AI Register. Deployments run in the United States, Brazil, Singapore, Australia, India and elsewhere.
The route to an individual connected model can use that partner's own infrastructure, so data locality depends on the applications chosen as well as the hospital's setup. No standard list of cloud regions or committed backup location is given for cloud customers.
Enterprise access is sold through a sales conversation, with a single contracting channel for buying applications from the marketplace, so a department signs one agreement instead of one per AI vendor. The service terms describe a subscription plus separately purchased services, and hospitals can compare several applications in the validation workspace before committing to any.
No platform subscription price, testing cost, rate per study or standard price for marketplace applications is published. Whether validation, production monitoring, local deployment, integration work and added applications sit in the same order and renewal is not stated.
Hospital radiology departments, independent imaging networks and validation research teams are the users, across North America, Latin America, Europe, the Middle East, India, Southeast Asia and Australia. The cleared viewer handles CT, MRI, ultrasound, digital radiography and 2D and 3D mammography. Named users span University Hospitals in Cleveland, RadNet, I-MED in Australia, Fleury in Brazil, Tan Tock Seng and Changi General in Singapore, and the Royal Marsden.
The platform works before production to evaluate products and afterward in the reading workflow. Coverage by body region or disease depends on the applications chosen, and planned moves into pathology, genomics and general medicine are not yet products.
Alternatives to CARPL
Other products that do the same kind of work as CARPL, and what each one does.
Enlitic builds software that cleans up and standardizes medical imaging data for hospitals, radiology groups and archive operators.
TeraRecon, a ConcertAI company, sells advanced visualization and AI workflow software for radiology and cardiology.
MosaicOS is a cloud radiology platform from Mosaic Clinical Technologies, the technology subsidiary Radiology Partners launched in July 2025.
DeepHealth is RadNet's digital health business and the entity that contracts RadNet's clinical AI and imaging informatics software to third party providers.
Romanian imaging AI company based in Timisoara whose Rayscape CXR analyzes chest X-rays, detecting and localizing 148 findings across 17 clinical classes, including tuberculosis, with bone suppressed views, an automatic cardiothoracic ratio and a normal or abnormal flag for worklist…
Better Medicine develops software that flags, measures and characterizes renal lesions on CT scans. BMVision Kidney uses deep learning to locate cystic and solid lesions and generate findings a radiologist can inspect in BMViewer.
Pricing
Vendor-published figures are labeled as such. Figures labeled “Estimated” are derived from third-party sources and have not been confirmed by the vendor.
| Entry Price | Pricing Basis | BAA Tier | Implementation | Source |
|---|---|---|---|---|
|
Contact CARPL for pricing
|
CARPL sells an enterprise subscription with one procurement channel for connected apps. Public terms do not disclose a seat, site or study rate. | The product describes HIPAA compliant deidentification in validation. No business associate agreement tier, or its reach to marketplace vendors, is stated in the public terms. | The service integrates a gateway and viewer with local imaging systems, on premises or in the cloud. Fees for installation, testing or a different deployment configuration are set in the quote. | Vendor Published |
Validation, monitoring and the marketplace can be bought through a single point of contact. Rates for the platform, individual applications and added services are not published.