When central readers split on progression and a baseline series is missing, first establish what each reader could evaluate. Query the missing images, apply the charter’s evaluability rules, and adjudicate only the disagreement those rules identify. The final endpoint then follows the prespecified statistical analysis plan (SAP). An adjudicator cannot reconstruct an image that was never acquired; a missing upload may still be recoverable. Neither situation automatically determines the participant’s progression date or removes the participant from analysis.
Plan four pathways before reading starts: absent images, technically inadequate images, differences between evaluable reads, and follow-up that stops after local progression. Each needs an owner, an operational action, a rule connecting it to the endpoint, and a retained record. This article focuses on oncology blinded independent central review (BICR). Its Response Evaluation Criteria in Solid Tumours (RECIST) examples and the European Medicines Agency (EMA) appendix on progression-free survival (PFS) and disease-free survival (DFS) should not be transferred unchanged to other diseases or modalities.
What the Imaging Charter Must Settle Before the First Scan
FDA’s April 2018 final imaging-endpoint guidance recommends a charter covering acquisition, display, interpretation, archiving, and reconciliation of discordant reads. It recognizes that readers may justifiably disagree. The guidance contains nonbinding recommendations, principally for imaging used to assess a primary endpoint or its component in drug and biologic trials. It does not impose a universal double-read design.
The March 2023 accelerated-approval oncology draft remains labeled draft and not for implementation on FDA’s guidance page. For single-arm response assessment, it recommends BICR and making a charter with adjudication procedures available with a marketing application. This is a proposed recommendation, rather than a binding submission requirement. The draft cites Ford et al. (2009), whose review discusses central-review operating issues in phase II and III oncology studies.
For implementation, distinguish document responsibilities. Put acquisition and reader instructions in the charter, the assessment schedule and endpoint role in the protocol, and the analysis mapping in the SAP. Assign an owner to each handoff. A technically accepted scan, a completed read, and an analysis-ready endpoint are separate milestones; recording only the last one conceals where an exception arose.
Why Two Blinded Readers Disagree—and When That Is Normal
Schmid et al. (2021) describe a common oncology design with two independent readers and an adjudicator. The PINTAD narrative review includes pharmaceutical, consultancy, and imaging-core-lab affiliations; its authors declared no conflicts of interest. It recommends investigating unexpectedly high or low disagreement and cautions that reducing reader independence can hide variability. Neither its recommendations nor the FDA guidance establish a universal acceptable rate. A low rate alone does not demonstrate collusion or high quality.
A synthetic calculation shows how category differences can arise. Under RECIST 1.1, target-lesion progression requires at least a 20% increase from the smallest sum on study and an absolute increase of at least 5 mm. With baseline and nadir both 50 mm, sums of 59.5 and 60.5 mm imply increases of 19% and 21%. One falls below the target-progression threshold and the other crosses it. Calling the first overall stable disease additionally assumes no new lesion or unequivocal non-target progression and no other response category. This illustration does not establish either reader’s accuracy.
Denominators change the meaning of a benchmark. Beaumont et al. (2021) analyzed six lung-cancer trials, 1,833 participants, and 10,684 timepoints. They reported aggregate discordance of 34.3% at timepoint level and 59.2% at participant level, where a participant counted if any timepoint was discordant. Participant-level discordance restricted to the date of progression was 32.9%. All authors were Median Technologies employees, and the data came from one vendor’s trial pool. These are observed results for particular read designs, rather than regulatory limits or universal planning rates.
In a monitoring specification, name the unit, the adjudication trigger, the denominator, and the handling of incomplete reads before interpreting any percentage. Separate disagreement between central readers from disagreement between local and central assessments. Review case mix, imaging completeness, reader training, and independence when a signal changes; replacing the read design simply to lower a dashboard number changes what that number measures.
The Acquisition-to-Read Decision Ledger
The following ledger is an original editorial synthesis, not a regulator-issued template or a validated study instrument. Its rows turn the source recommendations into questions a team can assign and test. Fill in trial-specific owners, deadlines, criteria, and record locations. Quality control (QC) determines technical suitability; a non-evaluable (NE) assessment describes the limits of an endpoint evaluation. Query windows and adjudication algorithms should be defined for the trial.
| Workflow state | Detection and owner | First action | Prespecified rule | Endpoint consequence | Retained evidence |
|---|---|---|---|---|---|
| Scheduled scan never acquired | Schedule reconciliation; site coordinator and imaging operations | Confirm the reason and whether acquisition remains permitted | Protocol window; SAP handling of missing assessments | Record the missing assessment; apply the endpoint-specific analysis rule | Query response, actual visit status, relevant deviation record |
| Scan acquired but not received | Receipt log; central intake team | Locate the source series and verify a retransmission | Charter transfer, receipt verification, and escalation process | Assess recovered data under the timing rules; unresolved absence remains missing | Original acquisition date, transfer query, receipt verification |
| Required series incomplete | Completeness QC; core-lab intake team | Query the missing anatomical coverage or series | Charter coverage and criteria-specific evaluability rules | Use remaining evidence only as the criteria allow; no automatic global NE | QC finding, series inventory, documented evaluability decision |
| Series technically inadequate | Technical QC; qualified imaging reviewer | Identify the defect; assess permitted correction or repeat acquisition | Charter quality and repeat-imaging criteria; protocol and safety constraints | Read suitable recovered images, or apply the unevaluable-assessment rule | QC report, repeat decision and justification, retained images |
| Evaluable reads agree | Independent read completion; core-lab read operations | Retain and lock original read outcomes | Charter concordance definition and sponsor transfer specification | Agreed image outcome feeds the protocol/SAP endpoint derivation | Original read forms, lock record, transfer reconciliation |
| Evaluable reads disagree | Endpoint-relevant comparison; read operations | Apply the charter’s reconciliation process | Charter trigger, adjudicator role, allowed inputs, and decision rule | Reconciled outcome feeds the prespecified endpoint derivation | Original reads, adjudication outcome and rationale, audit trail |
| Required baseline comparison unavailable to a reader | Reader flags evaluability; intake/QC owner investigates | Check source availability and mapping before treating it as an interpretation dispute | Charter recovery, evaluability, and controlled-reread rules; SAP missing-data handling | A lesion or assessment may be unevaluable; other sufficient evidence can still determine an outcome | Reader flag, source query, versioned rereads or unresolved-data record |
| Imaging stops after local progression | Follow-up reconciliation; clinical operations and biostatistics | Record why follow-up ended and compare with the planned collection strategy | Protocol local/central assessment roles; SAP censoring and robustness analyses | Central non-confirmation can leave missing follow-up and potentially informative censoring | Local and central dates, discontinuation reason, follow-up status |
Missing Images: Establish Whether Acquisition or Transfer Failed
A missing central file is an observation about receipt, not proof of a missed scan. Query the site and distinguish non-acquisition from failed transfer or incorrect mapping. Record actual acquisition and receipt separately. A later upload preserves the original acquisition date; it does not turn a late-acquired examination into an on-time assessment. For site capability planning, see Site Feasibility: Can This Site Deliver the Endpoint Assessments?.
Once recovery is exhausted, biostatistics applies the planned endpoint handling. A missing timepoint is not automatically a censored participant, and an unevaluable response assessment is not automatically exclusion from the analysis population. Which observations remain usable depends on the endpoint, the response criteria, and the estimand. Document the reason for absence rather than replacing it with an unexplained “missing” flag.
The EMA anticancer PFS/DFS appendix, adopted in December 2012 and effective July 2013, calls for predefined, justified missing-data and censoring methods. It generally assigns progression to the first documented evidence, including after missed visits. It does not support a blanket rule censoring every participant after two missed scans. An alternative analysis date must be defined and justified; it must not overwrite the actual observation date. See Time-to-Event Endpoints: Writing Censoring Rules Readers Can Follow for the statistical handoff.
Unreadable Series: Technical QC and Re-acquisition Rules
A received series can still fail the planned measurement: motion, incomplete coverage, or inconsistent acquisition technique can prevent a reliable comparison. Record the defect and its effect on the required assessment. “Suboptimal” and “uninterpretable for this endpoint” are different conclusions. A reviewer should state what can and cannot be evaluated under the criteria rather than converting every technical imperfection into a global non-evaluable result.
FDA’s final imaging guidance recommends specifying query and repeat-imaging processes and identifying situations where technical failure makes repeat imaging critical, with justification for additional radiation exposure. It also recommends describing read locking, preserving original outcomes for comparison with rereads, and prespecifying the adjudication-rate determination process when that rate matters to interpretation. Repeat-acquisition windows and investigation thresholds should be defined for the trial rather than treated as universal regulatory limits.
Operationally, put the repeat decision with the authorized clinical and imaging roles, within the protocol and participant-safety constraints. Recovery might involve retransmission of an existing series rather than another exposure. Record which path was taken and why. Trend recurring technical failures by site, then investigate whether acquisition instructions, qualification, or retraining need attention; a repeated QC failure does not by itself authorize a new scan or site termination.
Reader Disagreement: Define the Reconciliation Procedure
Specify which differences affect the endpoint. A response-category difference, a different progression date, and a difference in lesion measurement need not all trigger the same process. The operational article Discordance Between BICR Readers recommends limiting adjudication variables to those affecting the primary endpoint and defining dating and missing-data conventions. It is imaging-CRO-authored trade coverage, not a regulatory requirement. Its PFS example focuses adjudication on the progression date.
Distinguish adjudication from another independent read. In a forced-choice design, a third reader selects between specified existing outcomes. In an independent third-read design, define in advance how that result combines with the original reads, including a three-way split. The methods differ in what the adjudicator sees and how the final outcome is chosen. The charter should also state the permitted clinical information, identity masking, and escalation process. Clinical-information access and image presentation order depend on the trial; neither universal exclusion of all clinical information nor universal chronological display is justified.
Lock and preserve the original assessments before reconciliation. A revised assessment based on newly available information needs a defined reread pathway and an attributable record, rather than an overwritten measurement. Charter changes likewise need a prospectively described modification process, version control, and documented reasons.
When Local Progression Stops Imaging
If local progression ends imaging and later central review does not confirm progression, the trial may have no subsequent central event to observe. Dodd et al. (2008) explain that censoring such participants can be informative because stopping follow-up is associated with prognosis. Their methods analysis favors well-blinded designs and considers BICR useful as an audit, especially for marginally positive findings. It does not establish that central review always causes more bias than it removes, or that the bias always has one direction.
The EMA appendix discusses real-time BICR and additional scans after locally designated progression as ways to reduce that information gap. It also describes sample-based BICR to investigate directional local-assessment bias, with the sample and escalation rules planned beforehand. A sample audit does not recover missing follow-up. None of these options creates a universal requirement to continue treatment until central confirmation or an obligation to keep scanning someone who withdraws consent.
For local–central comparison, early and late discrepancy rates describe local progression being declared respectively earlier or later than BICR. They are distinct from disagreement between the two central readers. State their precise definitions and denominators in the analysis specification. Differences across treatment arms warrant investigation of potential bias, assessment timing, and missing follow-up; they do not by themselves prove that one assessor is correct.
What Central Review Does Not Fix
Amit et al. (2011) combined simulations with a meta-analysis of 27 randomized phase III trials through a PhRMA-sponsored working group. Local and central treatment-effect estimates were strongly correlated: R = 0.947, with a reported 95% confidence interval of 0.88–0.97. The authors argued that BICR may not be warranted in blinded trials or when a large PFS effect is observed, and that sample-based review may suffice in some large open-label trials. These historical, industry-associated summary results do not establish equivalence for every trial or validate a particular local reader.
Choose the review strategy from the endpoint’s subjectivity, masking feasibility, expected bias, and consequences of incomplete follow-up. Full central review adds a second assessment process; it cannot replace uncollected images. For a proposed sample audit, specify the sampling and escalation logic before seeing discordance results and discuss its suitability with the relevant regulator. An open-label design alone does not establish a universal requirement for two readers plus an adjudicator.
Locking, Rereads, and the Inspection Trail
ICH E6(R3), consolidated Step 4 guideline, June 2026, section 3.6, addresses agreements, service-provider suitability, and sponsor oversight. Transferring work to an imaging vendor does not transfer the sponsor’s ultimate responsibility. Its data-governance and essential-record provisions support retention, traceability, and access. They do not require every imaging source file to be copied into a single trial master file (TMF) folder.
Maintain an accessible record chain: acquisition and receipt history, QC decisions, original locked reads, any rereads, adjudication outputs, applicable charter versions, and the sponsor’s endpoint derivation. Identify where records reside and who can retrieve them. A core-lab repository can hold the source images while the TMF identifies the arrangements and related essential records. Align the filing plan with TMF Reference Model: Filing the Evidence Behind Endpoint Decisions and the oversight signals in Risk-Based Monitoring: Turning Endpoint Risks Into Review Signals.
Review a completed exception record by asking whether another qualified reviewer could reconstruct why that outcome entered the analysis. Reader-workload controls, substitute-reader arrangements, and retraining plans should fit the read task; the guidance and cited methods papers do not prescribe a universal session length. Use traceable workload and quality records to evaluate the plan.
What the SAP Must Say Before Unblinding
Resolve the analysis strategy during planning, before treatment-aware results can influence decisions. ICH E9(R1) distinguishes intercurrent events from missing data and requires sensitivity analyses to address robustness for the same estimand. A missing image is not itself an intercurrent event; treatment discontinuation or switching may be. Specify both the event strategy and what data are needed after it. See ICH E9(R1) Estimands: Specifying Intercurrent Events Before Endpoint Collection.
Define the estimand’s treatment conditions, population, variable, intercurrent-event strategies, and population-level summary. State the role of local and central outcomes in deriving the endpoint.
Define how absent or unevaluable assessments affect the analysis, including censoring where applicable. Preserve actual observation dates separately from any derived analysis dates.
If adjudication frequency matters to interpretation, specify its calculation, unit, denominator, and relevant summaries. Document any study-specific investigation thresholds and their rationale rather than presenting them as regulatory acceptance limits.
Prespecify clinically plausible assumptions and same-estimand sensitivity analyses. Comparing a local-assessed endpoint with a central-assessed endpoint may be supplementary because the variable changes; calling both “sensitivity” does not make them target the same treatment effect. See Missing Data vs Intercurrent Events: Estimand-Aligned Sensitivity Analysis.
The final cross-check is operational: take one missing scan, one QC failure, and one disagreement through the charter, protocol, and SAP and confirm that the resulting records explain the endpoint assignment. Statistical Analysis Plans for Endpoints: What Must Be Locked Before Unblinding covers that analysis specification in more detail. The ledger’s value is making these handoffs reviewable, not promising that prespecification eliminates missing data or guarantees regulatory acceptance.