This paper considers multimodal radiology generation from either a 2D chest radiograph (CXR) or a 3D chest CT volume. The system is required to produce two outputs: (i) a set of structured findings that can be checked against structured labels, and (ii) a free-form impression-style report for human readability and downstream use.

The key design requirement is that the system should be optimized and evaluated under open-access supervision, using released anchors: structured labels, reference-standard subsets when available, and reference regions when available. A second requirement is non-circularity: text/extractor signals may be used as auxiliary constraints during training, but quantitative correctness evaluation is anchored to held-out structured labels and reference regions under predefined rules.

These goals are implemented by introducing an ontology-derived atomic-finding interface as a bottleneck between vision and free-form text. This interface serves two roles. During training, it provides a structured channel for supervision and for counterfactual constraints that directly manipulate evidence. During evaluation, it provides a consistent target space for measuring correctness and completeness under a shared protocol.

The interface is defined at two levels. The supervised prediction index is an atomic slot (1)f=⟨c,r⟩,where c is a biomedical concept identifier (e.g., RadLex) and r is an anatomy region identifier. For consistent bookkeeping and cross-dataset conversion, each slot is associated with tuple states (2)af=⟨c,r,s,u⟩,s∈{pos,neg}, u∈{certain,uncertain}.

Thus, polarity and uncertainty are explicit in the tuple schema, while model supervision and primary metrics use the positive projection of each slot.

First, syntactic validity is required to be checkable without dataset-specific heuristics. A tuple record is valid if and only if all four fields are present and each value lies in a fixed, known inventory. This keeps conversions consistent under domain shift in report style.

Second, cross-dataset comparability is needed. Public datasets differ in label definitions and in whether uncertainty is encoded. By mapping each dataset’s released structured labels into this common tuple schema, a single slot space ℱ can be evaluated under multiple anchors with consistent rules.

The inventory of valid concepts and regions is constructed from (i) a public biomedical concept ontology (RadLex) and (ii) the concept inventories covered by two open-source report labelers used as fixed extractors (CheXpert labeler [2] and NegBio [15]). Across datasets, the mapping from released structured labels (e.g., CheXpert observations, VinDr labels, CT-RATE abnormalities) to atomic tuples is predefined.

All label-to-tuple and report-to-tuple conversions are rule-based with no learned thresholds beyond the global slot-level decision rule pf≥0.5 used in the evaluation protocol below. Table 1 summarizes the complete mapping logic used by training and evaluation.

Mapping ambiguity is accounted for explicitly rather than hidden in dataset-specific scripts. If anatomical region information is absent, the default protocol maps the finding to a thoracic-global identifier so that the study remains countable under the shared slot inventory; if different released sources disagree, a fixed precedence rule is used once and then shared across all methods. Because the same tuple interface is used for TA-DECT, internal baselines, and parser-converted external rows, this mapping layer can shift absolute numbers but does not selectively favor one method family. The experiments therefore include compact sensitivity analyses under alternate region, uncertainty, and conflict-resolution rules.

A multimodal encoder–decoder is trained to produce (i) atomic finding probabilities and (ii) a free-form impression-style report. The architecture is intentionally simple in order to isolate the effect of the proposed training objective and evaluation protocol.

The visual backbone is BiomedCLIP’s image encoder (ViT-B/16, about 86M parameters) [16]. For CXR, inputs are resized to 448 × 448 and encoded into patch tokens. For CT, volumes are clipped to Hounsfield units (HU) [−1000,400], linearly rescaled, and 32 axial slices are sampled uniformly per study; each slice is encoded by the same ViT-B/16 and slice tokens are aggregated by a two-layer attention-pooling block. This design is intentionally conservative: it preserves a matched image-to-token interface across CXR and CT and isolates the training objective from a larger architecture search, but it does not claim that 32-slice sampling is the best possible three-dimensional encoder. The revision therefore includes a higher-slice CT variant while keeping the objective fixed.

The atomic head is a multi-label classifier that predicts a probability pf(x)∈[0,1] for each atomic slot f=⟨c,r⟩. Here pf(x) denotes confidence for the slot-level positive state used by the label-anchor protocol (with uncertain/blank handled by predefined exclusion/mapping rules). The output of this head is the primary vehicle for truth-anchored correctness and for evidence sensitivity under interventions.

The text decoder is Mistral-7B-v0.1 [17] with added cross-attention blocks that consume visual tokens and an atomic-summary token stream. Decoder adaptation uses LoRA [18] (rank r=16, α=32, dropout 0.05) on attention and MLP projections; base decoder weights remain frozen. This yields about 7.24B total decoder parameters with about 93M trainable decoder-side parameters. In the default implementation, cross-attention adapters are inserted every 4 decoder blocks (8 insertion points over the 32-layer stack), the atomic summary is capped at 64 tokens, and pooled visual conditioning tokens are capped at 256 tokens. Conditioning on the structured predictions encourages report–structure consistency, while the report output remains free-form.

Fig. 1 summarizes the TA-DECT architecture and training objective, highlighting the atomic bottleneck, occlusion interventions, and the detached dual-extractor training auxiliary.

Let x be an image (CXR) or volume (CT), y be public structured labels when available, and r be the generated report. The TA-DECT objective couples four loss terms with an optional report likelihood term. The terms are designed to address distinct but interacting risks: label-supervised correctness, extractor-specific gaming, non-selective evidence sensitivity, and omission.

Training begins with structured supervision when labels are available. On examples with structured labels, define the slot-level positive target yf+=⊮[yf=pos] on the evaluable label subset yf∈{pos,neg} (uncertain/blank excluded by a predefined policy). A weighted multi-label binary cross-entropy over slots is then optimized:(3)ℒstruct=∑f∈ℱwfBCE(pf(x),yf+),where labels with state {uncertain, blank} are excluded under a predefined policy (reported as an excluded fraction). Weights are set to wf∝1/πf+ϵ where πf is the training-set prevalence of slot f (computed on the available structured-label subset for that dataset) and ϵ=10−3. This weighting reduces the tendency of rare findings to be ignored while preserving a consistent training setup.

To reduce circularity from a single evaluator, a dual-extractor minimax constraint is imposed on generated text. Two fixed report-to-finding extractors E1 (CheXpert labeler) and E2 (NegBio) are applied to the generated report r, producing extracted finding indicators y~f(k)(r)∈{0,1}. The worst-case mismatch between atomic predictions and extracted indicators is minimized:(4)ℒminimax=maxk∈{1,2}∑f∈ℱparse(x)mf(k)(x)BCE(pf(x),y~f(k)(r)).

Here ℱparse(x)⊆ℱ denotes slots covered by the fixed extractor inventories for the report domain of x. On CT-RATE, minimax updates are restricted to this parser-supported subset; slots outside ℱparse are optimized only by ℒstruct, ℒcf, and ℒcomp. The mask mf(k)(x) is an anchor-gating rule:(5)mf(k)(x)={1,if yf is unavailable,1,if yf∈{pos,neg} and y~f(k)(r)=yf+,0,otherwise.

Hence, on labeled samples, extractor outputs that conflict with the structured anchor do not contribute gradient. In optimization, this term is implemented as a deterministic detached-target step because report decoding and extractor parsing are non-differentiable. For each minibatch, one greedy report r^ is decoded per sample, E1/E2 are run on r^, and the resulting y~f(k)(r^) values are treated as constants when optimizing ℒminimax. Gradients from ℒminimax update the visual encoder and atomic head (through pf(x)), while decoder parameters are updated by ℒtext and by the shared conditioning on the atomic summary. This block-coordinate implementation makes the gradient path explicit: ℒminimax never backpropagates through parser outputs or report text, and it cannot override available structured labels because the anchor gate suppresses conflicting parser targets. Extractor states are converted to binary indicators using the fixed rule in Table 1 (positive/uncertain ↦1, negative/blank ↦0); the maximization over k applies the larger mismatch between the two fixed extractors for each update. Even with this safeguard, the extractors remain part of the training auxiliary loop, so this dependence is stated explicitly in the revised discussion and limitations rather than treated as invisible infrastructure.

Next, a counterfactual selectivity constraint is added, which directly tests whether predicted confidences respond selectively to the removal of evidence. An intervention family i∼ℐ(x) removes evidence by occluding either (a) a fixed image grid region (used in training and evaluation) or (b) a reference bounding box (evaluation only on held-out ROI splits).

Grid occlusion (train+eval). Training uses a fixed 3×3 grid: for 2D CXR, exactly one grid cell is occluded by replacing all pixels in that cell with the per-image mean intensity; for 3D CT, the same cell index is occluded across all axial slices. The 3×3 design is used in training because it is aligned to the coarse 9-region anatomy partition used by the atomic interface: (left/center/right) × (upper/middle/lower). In the revision experiments, a targeted retraining variant additionally mixes 3×3 and 7×7 interventions while keeping the rest of the objective unchanged.

ROI occlusion (eval-only). For MS-CXR and VinDr-CXR, given a held-out ground-truth box b, the pixels inside b are occluded (again replaced by the per-image mean intensity). ROI occlusion is never used for training.

Intervention-to-region and ROI-category-to-concept mappings follow the predefined rules in Table 1. For grid interventions, the targeted set is Treg(i)={f∈ℱ:region(f)=r(i)}, and for ROI interventions with category labels (MS-CXR and VinDr-CXR), the targeted set is Troi(i)={f∈ℱ:region(f)=r(i) ∧ concept(f)∈C(i)}.

Intervention distribution. For the default training setup (3×3), ℐ(x) is uniform over the 9 grid cells and one cell is sampled per example. At evaluation, metrics are averaged over all interventions in the tested grid family (3×3, 7×7, or 16×16) or over all held-out ROI boxes (ROI occlusion). Interventions with |T(i)|=0 are excluded from both training and evaluation averages.

The counterfactual term enforces selective decreases for targeted findings while stabilizing untargeted findings:(6)ℒcf=Ei∼ℐ(x)[1|T(i)|∑f∈T(i)max(0,m−(pf(x)−pf(x∖i)))⏟targeted drop (7)+λ1|ℱ∖T(i)|∑f∉T(i)|pf(x)−pf(x∖i)|⏟untargeted stability],with margin m>0. The experiments use m=0.05 and λ=1.0 throughout. Intuitively, the first term demands a minimum decrease for findings logically linked to the occluded region, while the second term penalizes global instability that would be consistent with non-specific confidence collapse.

Finally, to explicitly control omission, a label-anchored completeness constraint is added on examples with public structured labels:(8)ℒcomp=∑f∈ℱαyf+BCE(pf(x),1).

α is set to 1 for all findings, and the overall strength is controlled via δ in the total objective. This term adds a direct penalty on false negatives under the structured anchor: low-confidence predictions on positive labels increase this loss.

On datasets with paired report text, negative log-likelihood ℒtext is also optimized. The full objective is (9)ℒ=ℒstruct+βℒminimax+γℒcf+δℒcomp+ηℒtext.

Unless explicitly varied in ablations, weights are set to β=0.5, γ=1.0, δ=0.5, and η=1.0 on paired-text datasets; for image-only datasets, η=0. Text losses are applied only when paired text exists; on image-only datasets (e.g., VinDr-CXR), training uses only atomic losses.

The evaluation protocol is designed to make correctness and completeness jointly visible under a consistent rule set, while keeping extractor-based signals diagnostic. The evaluation emphasizes micro-averaged, label-anchored outcomes because they directly quantify omitted positives and unsupported positives in a form that remains interpretable across datasets.

Label-anchored indicators (micro-averaged) are used on held-out truth anchors. For each example x and slot f with ground-truth yf(x)∈{pos,neg}, predicted-positive is defined as y^f(x)=⊮[pf(x)≥0.5]. The default threshold pf≥0.5 is fixed once and shared across datasets, modalities, and internal methods so that comparisons do not inherit post hoc threshold tuning. Because different threshold choices can materially change the miss–unsupported balance, the revision experiments include a dedicated robustness sweep rather than treating 0.5 as intrinsically optimal. (10)TP=∑x,f⊮[yf+=1∧y^f=1], FP=∑x,f⊮[yf+=0∧y^f=1], (11)FN=∑x,f⊮[yf+=1∧y^f=0].

Here TP, FP, and FN denote true positives, false positives, and false negatives. Missed-finding rate =FN/(TP+FN), unsupported-assertion rate =FP/(TP+FP), and F1=2TP/(2TP+FP+FN). For anchors with explicit uncertain/blank states (CheXpert-style labels), uncertain/blank are excluded under a fixed policy. The excluded fraction (Excl.) is computed once from ground-truth labels on each anchor split and is therefore method-invariant within that split. This policy improves comparability under a binary slot projection, but it may hide clinically relevant uncertainty failures; it is therefore treated as a limitation rather than as a claim that uncertain labels are unimportant in practice.

For counterfactual evidence sensitivity, the study reports the selectivity metric. For input x and intervention i, let pf(x) be confidence for finding f, and pf(x∖i) after evidence removal. Define (12)targeted_drop(x,i)=1|T(i)|∑f∈T(i)max(0,pf(x)−pf(x∖i)), (13)untargeted_change(x,i)=1|ℱ∖T(i)|∑f∉T(i)|pf(x)−pf(x∖i)|,and report selectivity S=E[targeted_drop−untargeted_change]. For grid interventions, T(i)=Treg(i); for ROI interventions with category labels, T(i)=Troi(i). The subtraction makes the metric sensitive to two common failure patterns: (i) insufficient response for targeted findings and (ii) non-specific instability that changes many untargeted findings.

Evaluation pathways follow method family. Methods with a native atomic head are evaluated directly from pf(x) (including calibration). Report-only baselines are converted to binary findings using the same dual-parser rule in Table 1; because this pathway does not expose calibrated pf(x), expected calibration error (ECE) is not reported for report-only external baselines.

Calibration is evaluated using expected calibration error (ECE, 15 bins) for methods with native atomic probabilities, while extractor disagreement on generated reports is reported separately as a diagnostic signal.

TA-DECT is evaluated on public datasets chosen to cover paired image–report supervision, structured-label anchors, region-level evidence anchors, and cross-modality transfer. Where available, the study reports three complementary channels: label-anchored finding quality, region-of-interest (ROI)-based evidence linkage, and counterfactual selectivity.

Table 2 summarizes, for each dataset, modality/text availability, applied text losses, evaluation anchors, reference-region availability, intervention type, and train/test role.

CheXpert provides uncertainty-aware labels and an official radiologist benchmark subset. In this manuscript, primary CheXpert counting uses a held-out split of released structured labels aligned to the full atomic slot inventory; the smaller radiologist benchmark (validation 200, test 500 in the official protocol) is used as a secondary diagnostic reference. When text losses are applied in the CheXpert domain, paired reports are sourced from CheXpert Plus and aligned by released image paths.

MIMIC-CXR serves as the primary paired CXR training domain and as a source/target for report-style shift. Structured-label evaluation for MIMIC uses released MIMIC-CXR-JPG benchmark labels (mimic-cxr-2.1.0-test-set-labeled.csv); consistent with the official PhysioNet release, some auxiliary metadata files in v2.1.0 retain 2.0.0 prefixes, and the released naming is kept unchanged. Train/validation/test partitions follow the official MIMIC-CXR-JPG split file, with paired free-text reports joined from MIMIC-CXR. Because MS-CXR is derived from MIMIC-CXR, the study enforces study-level de-overlap: any MIMIC study appearing in the MS-CXR evaluation pool is removed from MIMIC training and validation for all compared methods. MS-CXR and VinDr-CXR boxes are then used only for held-out ROI evaluation (evidence linkage and ROI occlusion), never as training supervision. CT-RATE provides paired 3D CT volumes, reports, and structured labels, and is used for both in-domain testing and cross-modality shift experiments.

All internal variants use the same backbone family and decoding policy unless a specific ablation requires a structural change: BiomedCLIP ViT-B/16 with Mistral-7B-v0.1 and LoRA adaptation (rank 16, α=32, dropout 0.05). CXR inputs are resized to 448×448; CT volumes are clipped to HU [−1000,400], normalized, and sampled at 32 axial slices. This matched-backbone design is used to attribute performance differences to objective terms rather than to capacity changes.

Optimization uses AdamW (2×10−4 base LR), cosine decay, and 2k warmup. Batch sizes are 64 (CXR) and 8 (CT, with gradient accumulation to effective 64). Training runs for 80k steps on CXR and 40k on CT. Mixed-precision BF16, gradient checkpointing, and distributed sharding (FSDP-style) are used under the same configuration for all internal runs. Validation and checkpoint selection follow a single fixed rule across methods: select the checkpoint with highest validation anchor F1 under the same deterministic thresholding and exclusion policy used at test time.

For internal comparisons, preprocessing, augmentation, decoding, and intervention settings are shared across all rows. Inference uses beam size 3 and max report length 192 tokens, and atomic findings are positive at pf≥0.5. Calibration settings (ECE with 15 bins), intervention families, and region-mapping rules are unchanged across datasets. Unless explicitly stated otherwise, reported point estimates are the mean of five independent seeds and confidence intervals use two-stage bootstrap resampling. The same uncertainty protocol is used for aggregate counterfactual and localization summaries; when a compact table reports point estimates only, the corresponding values are interpreted as descriptive operating-point summaries rather than standalone significance claims. Training hardware is 8×A100-80GB (or equivalent).

Optimization remained stable across the reported internal runs under the same warmup, sharding, and checkpoint-selection protocol. The main additional training cost of TA-DECT relative to the direct-to-text baseline comes from one greedy decode per sample plus two fixed extractor passes for the detached minimax target; the inference architecture itself is unchanged. Wall-clock comparisons are not reported here because they are cluster-dependent, but the later experiments isolate the extra components by matched-backbone ablations and targeted retraining variants.

Baselines are organized into two groups: internal matched-backbone baselines and external reference baselines. The internal group preserves backbone capacity, data splits, and decoding policy, and changes only objective components. The primary comparator is direct-to-text, which removes atomic conditioning while retaining the same visual and decoder family. Additional internal controls include Atomic+report without dual minimax, single-extractor consistency (E1 only), and a grounding-only auxiliary model that emphasizes localization without explicit semantic completeness constraints.

External baselines provide reference points against established methods. The comparison includes R2Gen, R2GenGPT, KARGEN, LLaVA-Med, CT2Rep, and continual tuning (all introduced and cited in Section 2). Because several external methods are report-only and do not expose calibrated slot probabilities, label-anchor outcomes are computed via the same dual-parser conversion used for all report-only rows, and ECE is omitted in those rows. The internal matched-backbone comparisons are therefore treated as the strongest causal evidence for the proposed objective, while the external rows serve as practical reference comparisons under a necessarily less symmetric evaluation interface.

Ablation design follows a one-factor-at-a-time protocol around the full TA-DECT objective. The ablation study removes each major term (ℒminimax,ℒcf,ℒcomp), tests single-extractor replacement for minimax, and tests removal of anchor masking in minimax. This protocol is intended to reveal whether gains arise from one dominant component or from interaction among complementary constraints.

Table 3 reports micro-averaged finding quality on the primary structured-label anchors (CheX-hold., MIMIC-CXR-JPG curated labels, and CT-RATE structured labels), including F1, precision, recall, missed-finding rate, unsupported-assertion rate, ECE (when available), and exclusion fraction. Secondary report-derived rows are included for scale and extractor-drift diagnostics.

The central question is whether unsupported-assertion reductions can be obtained without increasing omissions. Across matched-path internal rows, TA-DECT improves F1 while reducing both missed-finding and unsupported-assertion rates relative to direct-to-text and other internal controls. For the primary matched-backbone comparison against direct-to-text, the 95% bootstrap CIs for F1 are non-overlapping on all three primary anchors (CheX-hold., MIMIC-cur-hold., and CT-RATE), indicating that the observed F1 gains exceed the estimated run-and-sample uncertainty under the shared protocol. Missed-finding and unsupported-assertion rates remain part of the same operating-point description rather than separately tuned hypothesis-test endpoints. The CheX-hold. anchor has a non-zero excluded fraction under the fixed uncertain/blank policy, whereas MIMIC-cur-hold. and CT-RATE are binary anchors in this protocol and therefore have Excl. = 0.00. As noted above, these internal rows provide the strongest evidence for the proposed objective because backbone family, split files, decoding policy, and intervention settings are matched.

Table 4 provides TP/FP/FN counts from a fixed reference checkpoint (seed 1) on the held-out primary anchors. Seed variation is reported as across-seed SD (five runs). Confidence intervals in Table 3 use two-stage bootstrap resampling.

Anchor composition and counting scope are summarized in Table 5.

Fig. 2 provides a complementary view of the omission–hallucination trade-off, plotting missed-finding rate vs. unsupported-assertion rate with uncertainty intervals.

To keep the revision compact, the newly requested sensitivity and retraining analyses are summarized in one table. The first block re-scores existing checkpoints under alternate mapping and threshold rules without retraining. The second block reports the two targeted retraining variants requested by the reviewers: mixed-granularity counterfactual training and a higher-slice CT pathway.

These additions are intentionally narrow. They are included to test whether the main directional conclusions persist under reasonable protocol changes, not to introduce a new benchmark or a broader architecture search. Across the re-scoring settings, the default protocol remains a stable operating point with the expected miss-vs.-unsupported trade-off under threshold sweeps. Table 6 summarizes both the re-scoring sensitivity analyses and the targeted retraining variants.

Label-anchor metrics quantify correctness and completeness, whereas counterfactual tests quantify evidence sensitivity. Fig. 3 reports selectivity for held-out ROI occlusion (MS-CXR, VinDr-CXR) and grid occlusion (CT-RATE). These intervention results are intended as coarse evidence-sensitivity diagnostics rather than lesion-level causal localization tests. Because targeted sets are defined by fixed region/category mappings, improved selectivity should be read as a more appropriate confidence response under the released anchors, not as evidence that every fine-grained localization failure is removed.

To connect selectivity with spatial evidence behavior, patch-level gradient saliency maps are computed and localization is evaluated on held-out ROI data using pointing game accuracy (PGA) and intersection-over-union at threshold 0.1 (IoU@0.1). For each target finding, the corresponding slot logit is backpropagated to the visual patch tokens, the absolute gradient is taken, the result is averaged across the channel dimension, normalized to [0,1], and upsampled to image resolution. PGA counts whether the peak response lies inside the reference ROI, and IoU@0.1 thresholds the normalized heatmap at 0.1 before measuring overlap with the held-out ROI mask. Table 7 reports localization and counterfactual outcomes jointly.

The later failure review makes this limit visible: anatomy mislocalization/region mismatch and uncertainty-collapse cases remain even when coarse selectivity and PGA/IoU improve.

Table 8 isolates each objective component. Removing the counterfactual term consistently lowers selectivity, showing that targeted evidence sensitivity does not emerge reliably from label supervision and report likelihood alone. Removing completeness increases missed-finding rates, consistent with under-reporting behavior.

Weakening dual-extractor constraints (removing dual minimax, replacing minimax with single-extractor consistency, or removing anchor masking) increases extractor disagreement and reduces label-anchor quality. Together, these ablations are consistent with complementary roles across objective terms rather than a single dominant term.

The study evaluates a pre-specified shift suite: MIMIC→CheXpert, CheXpert→MIMIC, and two CXR/CT modality shifts via CT-RATE. Fig. 4 decomposes drift into label, evidence, and extractor components; Table 9 reports deltas relative to each method’s in-domain baseline.

All methods degrade under shift, but the decomposition clarifies where degradation concentrates. In several settings, TA-DECT shows smaller label-anchor drift than direct-to-text under identical decoding and intervention rules. Extractor drift remains diagnostic rather than primary evidence, but it helps separate text-parser instability from anchor-level performance change.

Fig. 5 tests whether anchor metrics can be improved by degenerate policies such as forced short reports, uncertainty-heavy phrasing, and global confidence shrinkage (pf←κpf, fixed κ<1).

Across settings, these stress tests produce compensating failures: lower unsupported-assertion rates are typically accompanied by higher miss rates or weaker selectivity. This supports interpreting the main endpoint as a coupled profile (F1, miss, unsupported, selectivity) rather than any single scalar.

Fig. 6 provides representative qualitative examples that complement the aggregate metrics. Each panel summarizes supported/unsupported/missed findings together with confidence changes before and after ROI or grid occlusion under the same mapping rules used in the quantitative protocol.

These examples are illustrative diagnostics rather than primary evidence. They are included to make mixed error behavior within a single study visually explicit while keeping the main claims anchored to structured-label and ROI-based results.

To make residual safety-relevant errors more explicit, and to expose failure modes that can be obscured by coarse region partitions or fixed uncertain-label exclusion, this revision reports a compact manual review of 20 errorful studies per primary anchor (CheX-hold., MIMIC-cur-hold., CT-RATE), where “errorful” means at least one false positive or false negative under the default threshold. Each sampled study receives one primary label: omission-dominant, unsupported-assertion-dominant, anatomy mislocalization/region mismatch, or uncertainty/wording collapse. Table 10 summarizes the resulting error categories within each reviewed anchor.