Central Review in Clinical Trials

Central review in a clinical trial is the process of having imaging data assessed by an independent group of qualified readers located away from the trial sites, rather than by the investigator at each participating centre. It standardises endpoint assessment across all sites, reduces inter-reader variability, and eliminates local investigator bias — making imaging-derived endpoints more credible in regulatory submissions.

 

What is central review in a clinical trial?

Central review — also called central imaging review, centralized radiology review, or independent central review — is the process of having trial imaging data assessed by a small team of qualified specialist readers at a central facility, rather than by the radiologist or clinician at each investigator site.

In a multi-site trial, endpoint assessment distributed across dozens of local radiologists introduces two systematic problems. First, different readers apply the same criteria differently — inter-reader variability that inflates measurement noise and reduces the reliability of the endpoint. Second, in open-label trials, site investigators know which treatment their patient is receiving, creating the risk of unconscious bias in endpoint assessment. Central review reduces both of these problems: a single team of readers, applying standardised criteria and standardised software, assesses all images without knowledge of treatment assignment or patient history. Central readers can still disagree with each other — which is why double reads and adjudication exist — but the degree of variability and the bias risk are substantially lower than with distributed site reads.

This is why central review is particularly important in multi-site imaging trials: the more sites in a trial, the greater the variability introduced by distributed local reading — and the greater the benefit of centralising assessment.

Site read, central review, and BICR — what is the difference?

These three terms describe related but distinct approaches to imaging endpoint assessment. Understanding the difference is essential for protocol design and regulatory planning.

Site read (local read/investigator assessment)

Images are reviewed by the radiologist or clinician at each trial site. The reader knows the patient, their history, and in open-label trials, their treatment assignment.

When it may be sufficient: early-phase trials, double-blind designs, exploratory imaging endpoints, or studies where imaging is not the primary efficacy endpoint. Regulators generally accept local reads in well-controlled blinded trials.

Limitations: introduces inter-reader variability across sites; creates bias risk in open-label designs; does not provide the independent verification regulators expect in pivotal submissions.

Central review

All images are routed to a central facility where a qualified team of specialist readers assesses them using standardised criteria, standardised software, and standardised protocols. Readers are not necessarily blinded to treatment assignment — blinding depends on trial design.

When it is used: multi-site trials where endpoint consistency across sites is a priority; neurological trials using complex MRI biomarkers; any trial where the sponsor wants a consistent, documented assessment record independent of site variability.

Limitations: more expensive and operationally complex than site reads; requires an image management system capable of receiving, anonymising, and routing images from all sites.

BICR — blinded independent central review

Central review performed under specific regulatory conditions: readers are blinded to treatment assignment, prior assessments, and patient-identifying information; they are independent of the sponsor and the investigator.

When it is required or strongly recommended: pivotal unblinded oncology trials where progression-free survival or objective response rate is the primary endpoint. FDA guidance¹ and EMA guidance² recommend BICR in these settings. Sponsors who omit BICR in pivotal unblinded oncology trials should be prepared to justify the decision to the regulatory agency, which may request BICR data during review.

Limitations: most resource-intensive option; introduces the risk of informative censoring in progression-free survival analysis when central and local assessments disagree.

 

Setting up central review for your next trial?
QMENTA's imaging scientists can advise on reading plan design, reader qualification requirements, blinding architecture, and platform configuration for your protocol and regulatory jurisdiction.
Talk to an expert →

 

When is central review used?

Central review — whether blinded or unblinded — is used in a range of settings beyond pivotal oncology trials.

Neurological trials increasingly use centralised reading to standardise MRI-based endpoint assessment across sites. The 2024 revised McDonald Criteria incorporated the Central Vein Sign and paramagnetic rim lesions as diagnostic markers — both require specific MRI sequences and trained specialist readers to evaluate reliably. Assessing these markers consistently across a multi-site trial operationally favours centralised reading, though the criteria themselves do not mandate a specific review structure and further prospective validation of these markers continues.

Multi-site oncology trials use centralised reading to enforce consistent application of RECIST 1.1 or disease-specific response criteria across sites with varying levels of expertise in the specific indication.

Medical device trials — hardware/therapeutic devices Pivotal trials of hardware-based therapies — radiation devices, ablation systems, intratumoral delivery platforms — face the same open-label bias problem as unblinded pharma trials: site investigators who administer or monitor the device cannot be meaningfully blinded, making independent central review the only credible mechanism for unbiased endpoint assessment. Mixed-modality requirements, such as tracking response across CT and clinical photography within the same assessment, add complexity that centralised infrastructure handles more reliably than distributed site reads. For hardware device sponsors, establishing centralised image collection from study start is operationally critical — without it, retrospective BICR is infeasible and FDA may have no accepted sensitivity analysis to fall back on.

AI Software as a Medical Device (SaMD) validation studies Validating an AI diagnostic or imaging algorithm requires a reference standard that is independent of the data and institutions the model was trained on — otherwise, the validation is circular. Central review by independent specialist readers, applying standardised criteria across a representative image set, produces the clean ground truth that regulators require. FDA's evolving AI/ML SaMD guidance and the IMDRF framework both point toward independent performance evaluation; central review is the operational mechanism that makes that independence real and auditable.

Trials with imaging as a secondary endpoint may use central review as a quality-control measure even when local reads are used for primary efficacy assessment — providing an independent dataset for post-hoc analysis or regulatory queries.

Central review is generally not required in: early-phase trials, double-blind designs where blinding eliminates the bias concern, or studies where imaging is purely exploratory and not a regulatory endpoint.

How central review works — step by step

Reader qualification

Before being assigned patient cases, all central readers must pass a qualification set — a batch of test images demonstrating they can apply the trial's endpoint criteria within an acceptable variability range. Reader qualification is an important entry-level control for endpoint consistency within the central review team, but it is not sufficient on its own. Ongoing calibration, intra-reader consistency monitoring, adjudication rate tracking, and interim performance audits are all part of maintaining consistent endpoint assessment throughout a long trial. Unqualified readers are a common root cause of high adjudication rates in double-read studies.

Step 1 — Site Image submission
Images are collected from trial sites via PACS integration or manual upload through a site uploader tool to the central platform. For retrospective or mass migrations, images are batch-uploaded.

Step 2 — Anonymization at receipt
Images are automatically de-identified upon receipt before any reader can access them. PHI is removed from both DICOM header tags and pixel data. See: DICOM anonymization.

Step 3 — Image QC
Images pass an automated check validating DICOM integrity, modality, and series completeness. Images that pass then undergo a manual visual QC review, typically with a structured pass/fail assessment covering image quality, acquisition compliance, and eligibility for reading under the trial's criteria (for example, confirming MRI sequences meet RANO requirements). Any image that fails at either stage triggers a query back to the site or sourcing vendor.

Step 4 — Case assignment
QC-cleared images are assigned to qualified readers according to the trial's reading plan. Three common reading plan structures are used:

• Single read — one reader assesses each image. Used for exploratory endpoints or when resource constraints apply, and the regulatory context permits.
• Double read with adjudication — two independent readers assess each image; a third adjudicator resolves any discrepancy. Standard for pivotal BICR in oncology.
• Sequential read — the second reader reviews after the first, with defined rules for when adjudication is triggered. Used in some neurological and complex-biomarker settings.
• AI-assisted workflows – the AI output is prepared alongside the image set before reader assignment. Multiple readers can review the same case in parallel.

Step 5 — Reading and assessment
Readers access images through a validated reading platform and record structured assessments in an electronic case report form (eCRF). Readers in BICR studies have no access to site read data, patient treatment information, or prior central read results at the time of assessment.

• Standard BICR (2+1): Reader 1 and Reader 2 independently assess each case, recording structured measurements and response determinations in the eCRF. Neither reader has access to the other's assessment, site read data, or treatment information.
• AI-assisted reading: Readers complete an unaided session first (Session 1), then a second aided session (Session 2) where AI pre-segmentation and measurements are available. This two-session structure preserves the ability to measure the AI's impact on reader performance.

Step 6 — Result QC (eCRF QC)
Completed assessments are reviewed through automated edit checks before entering the adjudication or results pipeline — validating field completeness, measurement plausibility, and internal consistency across timepoints. Flagged assessments can generate a query and are returned to the reader for correction or escalated for adjudication.

Step 7 — Adjudication (if applicable)
In double-read designs, where the two independent readers disagree beyond a protocol-defined threshold, a third adjudicator reviews the discordant case and renders a final determination, typically choosing one of the two existing reads rather than performing an independent assessment from scratch. The adjudicator's assessment also passes through QC before the result is locked. Adjudication rates are tracked throughout the study as an ongoing performance indicator for the reader panel.

Step 8 — Dataset compilation and delivery
Locked assessments are compiled into the central read dataset and exported in the agreed format for statistical analysis and regulatory submission. The full audit trail — image receipt, QC decisions, reader assignments, assessments, adjudications — is preserved in the audit trail for inspection readiness.

Key takeaways

• Central review, site read, and BICR are three distinct approaches to imaging endpoint assessment — distinguished by who performs the read, whether they are blinded, and whether they are independent of the sponsor
• BICR is central review performed under specific blinding and independence conditions — it is commonly expected by regulators in pivotal unblinded oncology trials, though the need depends on trial design and blinding status
• Central review eliminates two problems simultaneously: inter-reader variability across sites and open-label investigator bias
• Reader qualification — passing a test set before being assigned patient cases — is the primary mechanism for ensuring consistent endpoint assessment within the central review team
• Double read with adjudication is the standard reading plan for pivotal BICR; single reads and sequential reads are used in other regulatory and operational contexts
• Retrospective BICR can be used as a sensitivity analysis in appropriate circumstances when images were systematically collected throughout the trial — it is not a blanket option for all trials
• Every step of the central review workflow — from image receipt to final dataset delivery — must be captured in an audit trail for regulatory submissions