LUCAID: Agentic Multimodal AI for Lung Cancer Precision Pathology

1 Introduction

• Manual evaluation is increasingly laborious and prone to interobserver variability as biomarker scoring systems assess different cell populations and compartments, while spatial relationships remain difficult to evaluate visually.Emerging TME features, including tumor-infiltrating lymphocyte density and spatial distribution, are impractical to assess manually at scale.Source: §1 · p. 2
• Existing AI tools often address isolated tasks, rarely reach pathologist-level performance, and have seldom been prospectively validated; pathology agentic systems lack validated quantitative measurements and clinical validation [42].LUCAID’s validated modularity differentiates it from vision-language models that remain unreliable for fine-grained perception and counting [44] [45] [47].Source: §1 · pp. 2, 4
• LUCAID orchestrates modules for quality control, tumor detection and subtyping, TME characterization, cellularity assessment, IHC phenotyping, and PD-L1, TROP-2, and MET scoring.The agent integrates quantitative module outputs into a structured report and supports case-specific natural-language queries.Source: §1 · p. 2, Figure 1 · p. 3
• 93.0% concordance with an expert-panel adjudicated reference standard across clinically actionable tasks exceeded the 68.3–81.1% concordance of individual pathologists.Validation used an independent prospective multicentric cohort.Source: §1 · p. 4

2 Results

• LUCAID combines nine on-demand or end-to-end AI modules for tissue quality control, tumor analysis, biomarker scoring, and structured case-level reporting.The modules cover H&E- and IHC-stained whole-slide images, from tissue profiling to immunohistochemical assessment and molecular testing.Source: §2 · p. 4
• Module testing used task-specific hold-out sets from a multicentric dataset of 1,620 lung cancer cases across three pathology institutes.Figure 2 describes 105,227 expert annotations across ten independent hold-out test sets, covering 26 prediction classes in four algorithm categories.Source: §2 · p. 4, Figure 2 · p. 5
• LUCAID was assessed at four levels: module performance against expert annotations, biological concordance with molecular measurements, report grounding, and prospective comparison with five thoracic pathologists.The prospective evaluation used an independent lung cancer cohort from the National Network Genomic Medicine (nNGM).Source: §2 · p. 4

2.1.1 Module 1: Quality Control

• Quality control achieved overall F1 scores of 0.91 on H&E and 0.88 on IHC slides, while valid tissue and non-tissue background each reached F1 0.99 in both modalities.The module identifies tissue regions and common artifacts in both staining modalities, providing a unified preprocessing step for downstream analyses.Source: §2.1.1 · p. 6

2.1.2 Module 2: Tissue Segmentation for Tumor Detection and Compartmentalization

• Across five lung cancer subtypes, tissue segmentation achieved average F1 0.93; carcinoma detection reached F1 0.98 (range 0.97–0.99) and metastatic carcinoma F1 0.99 (range 0.97–0.99).The model distinguishes seven tissue categories, including carcinoma, epithelium, stroma, necrosis, blood, vessel, and other tissue components.Source: §2.1.2 · p. 6, Figure 3 · p. 7

2.1.3 Module 3: Tumor Subtyping via Cell Phenotyping

• Lineage-marker scoring accuracies were 93.3% for TTF-1, 97.6% for CK7, 98.1% for p40, 100% for CK5/6, 95.5% for SYP, and 92.1% for CgA.Cases were classified as marker-positive when more than 10% of tumor cells were positive; marker profiles supported automated tumor classification using diagnostic criteria.Source: §2.1.3 · pp. 6–8

2.1.4 Module 4: Cell Phenotyping for TME Profiling

• H&E cell phenotyping achieved average F1 0.95 (subtype range 0.90–0.98), with carcinoma-cell F1 0.95 in primary and 0.99 in metastatic tumors.The test set comprised 24,880 annotations across 205 WSIs spanning all five major lung cancer subtypes.Source: Figure 2a · p. 8, §2.1.4 · p. 8

2.1.5 Module 5: Tumor Cellularity Assessment for Molecular Profiling

• Tumor cellularity assessment quantifies H&E tumor content using cell-count and nuclear-area approaches, and spatial heatmaps visualize regional variation in predicted tumor content.The maps illustrate intratumoral heterogeneity that may be difficult to appreciate by conventional H&E assessment alone and may support more standardized tissue-area selection for molecular profiling.Source: §2.1.5 · p. 8

2.1.6 Module 6: IHC Cell Phenotyping for Biomarker Expression Scoring

• IHC cell phenotyping achieved all-cell F1 scores of 0.86 for TROP-2, 0.86 for MET, and 0.90 for PD-L1; carcinoma-cell F1 scores were 0.98, 0.97, and 0.95, respectively.The module classifies nine cell phenotypes, enabling biomarker expression to be assigned to specific cell populations.Source: §2.1.6 · pp. 8–9

2.1.7 Modules 7 and 8: Biomarker Expression Scoring

• Expression scoring reached average F1 0.93 for PD-L1, 0.87 for membranous MET and TROP-2, and 0.82 for cytoplasmic TROP-2.For the clinically most relevant moderate and strong expression categories, performance ranged from 0.86 to 0.93.Source: §2.1.7 · p. 9

2.2 Cellularity Estimates Correlate with Molecular Reference Data

• Nuclear area-based LUCAID cellularity achieved correlation with KRAS VAF of r = 0.46 (p < 0.001) and MAE of 17.1 ppt, reducing MAE 30.0% vs routine and 21.2% vs cell count.Quantification was performed within pathologist-annotated tumor regions selected for molecular testing.Source: §2.2 · p. 10
• Cell count-based LUCAID estimates correlated with KRAS VAF at r = 0.39 (p < 0.001) and had MAE of 21.7 ppt.Source: §2.2 · p. 10
• Routine pathologist assessment correlated weakly with KRAS VAF (r = 0.20, p = 0.033) and had MAE of 24.42 ppt.Source: §2.2 · p. 10

2.3 LUCAID Captures Patterns of Tumor Invasion and Identifies Prognostic TME Features

• Greater stromal TIL levels and carcinoma cell–plasma cell adjacency were associated with improved survival, while higher tissue NLR and endothelial cell–lymphocyte distance were associated with adverse outcome.The analyses were UICC-adjusted; TIL and NLR recapitulated established prognostic associations, while the spatial features provided additional prognostic information.Source: §2.3 · p. 10, Figure 6 · p. 13
• Pooled correlations between LUCAID and joint pathologist tissue-compartment assessments were Pearson r = 0.996 and Spearman ρ = 0.96, with carcinoma, necrosis and stroma MAEs of 1.6, 1.9 and 2.9 percentage points, respectively.The quantitative measurements support reproducible regression grading of resected specimens after neoadjuvant therapy following IASLC recommendations.Source: §2.3 · p. 10, Figure 6 · p. 13
• Integration of tissue segmentation and cell phenotyping enabled automated assessment of vascular invasion, lymphatic invasion and STAS, while cohort-wide tissue and cell distributions provided reference context for individual measurements.The distributions came from 1,001 discovery-cohort patients with follow-up and included features difficult or impractical to quantify by routine visual inspection.Source: §2.3 · p. 10, Figure 6 · p. 13
• Combining H&E-derived intratumoral lymphocyte density with IHC-based PD-L1 TPS stratified tumors into four TIL-high/low and PD-L1-high/low immune phenotypes associated with differential response to PD-1/PD-L1 blockade.LUCAID provides a standardized automated implementation of this combined immune phenotyping strategy for scalable cohort analysis.Source: Figure 6 · p. 13, §2.3 · p. 14

2.4 LUCAID Generates Structured Pathology Reports Grounded in Validated Modules

• Each report pairs module measurements with concise clinical interpretations, then integrates findings into a case-level assessment; measurements are transferred directly from validated modules before LUCAID contextualizes them.Source: §2.4 · p. 14
• Before generating interpretations, LUCAID retrieves potentially relevant PubMed records and selects citations from them; it also supports case-specific follow-up queries by calling relevant modules and integrating their outputs.Citation entries are taken directly from PubMed, preventing references to non-existent publications by construction.Source: §2.4 · p. 14
• The evaluation decomposed reports into statements rated for measurement groundedness, interpretation correctness, citation correctness, potential harm, and likelihood of influencing clinical decisions, while also checking omissions and contradictions.Source: §2.4 · pp. 14–15
• All 134 module-referencing statements were grounded, 1,194/1,231 interpretive statements (97.0%) were correct, and 131/159 citations (82.4%) fully or partially supported their associated statements.The 1,365 statements came from ten reports; 37 interpretive statements (3.0%) were incorrect.Source: §2.4 · p. 15
• Across 70 section-level assessments, no clinically relevant omissions were found, but 3 contradictions occurred in 2 reports; among 111 errors, 96 (86.5%) had no potential for harm, 15 (13.5%) mild-to-moderate potential, and none severe.Source: §2.4 · p. 15

2.5 LUCAID Shows High Concordance across Clinically Actionable Tasks

• LUCAID ranked third of six raters in continuous-score correlation with the reference, with all raters’ correlations significant at p < 0.001.LUCAID’s MAEs were 3.46 percentage points for tumor cellularity (ρ = 0.975), 6.70 for PD-L1 (ρ = 0.959), 19.11 for MET (ρ = 0.926), 20.82 for membranous TROP-2 (ρ = 0.776), and 24.20 for cytoplasmic TROP-2 (ρ = 0.846). For cellularity, pathologist MAEs were 8.9–21.23 percentage points; LUCAID ranked among the lowest-error raters across tasks.Source: §2.5 · p. 15
• Against KRAS variant allele frequency, the nucleus area-based cellularity approach had MAE = 10.91 percentage points and r = 0.776, versus MAE = 18.25 and r = 0.760 for tumor cell counting.Pathologist estimates had MAEs of 14.89–23.68 percentage points and correlations of r = 0.18–0.70.Source: §2.5 · p. 15
• 93.0% overall clinical action concordance was achieved by LUCAID, compared with 68.3–81.1% for individual pathologists against the expert-panel reference standard.The evaluation covered 338 assessments by LUCAID and five experienced pathologists across 70 prospectively enrolled patients.Source: §2.5 · p. 15
• Concordance was 98.6% for molecular testing eligibility, 89.9% for PD-L1 treatment stratification, and 92.5% for MET classification, exceeding pathologist ranges.For cellularity eligibility, pathologist concordance was 89.9–92.9%; for PD-L1 it was 70.1–76.8%, and for MET it was 56.7–75.8%.Source: §2.5 · pp. 15, 17
• TROP-2 concordance was 95.2% for membranous expression and 88.7% for cytoplasmic expression, versus pathologist ranges of 48.4–93.5% and 50.0–74.2%, respectively.TROP-2 showed the greatest interobserver variability among the evaluated tasks.Source: §2.5 · p. 17
• Across cases, 122 assessments (36%) differed between evaluators, with disagreement varying markedly by case and threshold-related discrepancies affecting cellularity, PD-L1, MET, and TROP-2.In low-cellularity tumors, pathologists more often estimated tumor content above the 10% testing threshold, while LUCAID remained below it; LUCAID also detected small PD-L1-positive tumor-cell populations crossing 1% that visual assessment missed. These differences could affect molecular testing or immunotherapy eligibility.Source: §2.5 · p. 17

3 Discussion

• LUCAID’s validated modules span the diagnostic workflow, unlike prior tools such as GrandQC, Kludt et al., Spronck et al. [67], and HistoPLUS [68], which address isolated steps.The discussion states that those prior approaches had not been prospectively validated against pathologists on threshold-based treatment decisions.Source: §3 · p. 17
• 93.0% agreement with the adjudicated expert reference standard was achieved by LUCAID, versus 68.3–81.1% for five experienced thoracic pathologists across clinically actionable decisions.The comparison was made in prospective, multicenter validation on a real-world lung cancer cohort.Source: §3 · p. 17, Figure 8 · p. 18
• 122 of 338 assessments by five pathologists disagreed on clinical action category, underscoring the potential value of standardized, quantitative, reproducible scoring for actionable decisions such as molecular-testing eligibility and PD-L1 assessment.The authors suggest AI’s principal benefit may be standardization and reproducibility across institutions and observers, rather than automation alone.Source: §3 · pp. 17, 19
• LUCAID’s biological-primitives approach supports cohort-scale biomarker validation and discovery, including assessments not yet established in clinical practice.It reproduced joint pathologist assessments of viable tumor, necrosis, and stroma to within 1.6–2.9 percentage points; combined lymphocyte density and PD-L1 scoring to define four immune phenotypes; and computed reference distributions across 1,001 patients for case contextualization and candidate prognostic features.Source: §3 · p. 19
• The authors advocate comprehensive, reliable biological readouts across indications and institutions over task-specific models, paired with whole-slide analysis and a natural-language interface for pathology workflows.They envision pathologists selecting and directing independently validated tools through natural language, extending the principle demonstrated in lung cancer diagnostics.Source: §3 · pp. 19–20
• LUCAID grounds reports in validated module measurements, unlike image- or literature-grounded report systems [42] and evidence-retrieval approaches [75] [76] that do not reliably quantify underlying biology.The agent contextualizes module values against clinical thresholds, while the values themselves are filled directly from the modules; this avoids hallucinating biomarker values from diagnostic labels by construction.Source: §3 · pp. 19–20
• 97.0% of clinical interpretations in the ten-report assessment were judged correct; 9.8% of statements referenced module outputs and all were faithful, while 82.4% of citations supported or partially supported their statements.Among errors rated for harm, 15 (13.5%) carried mild-to-moderate potential and none was rated severe; the authors characterize this evaluation as preliminary.Source: §3 · p. 20
• The study is limited to lung cancer, and real-world effects on clinical decisions and routine pathologist interaction remain unestablished; generated-report evaluation covered only ten cases.The orchestrating agent uses a general-purpose model, with domain-specific alternatives, fine-tuning, and human-feedback refinement identified as possible next steps.Source: §3 · p. 20

4.1 Clinical Cohorts; 4.1.1 Test Cohorts for Quality Control, Tissue Segmentation, Cell Phenotyping and Biomarker Scoring

• Test cohorts spanned primary and metastatic lung tumors, QC slides, and IHC biomarker tasks, with tissue and imaging diversity intended to support generalization.The H&E segmentation and cell-phenotyping cohort included 158 primary and 47 metastatic tumor slides; QC cohorts included 132 H&E and 106 IHC slides, and IHC task cohorts covered TROP-2, MET, and PD-L1 phenotyping and scoring.Source: §4.1 · p. 20, §4.1.1 · p. 21

4.1.2 Discovery Cohort; 4.1.3 Prospective Clinical Validation Cohort

• The discovery cohort comprised 1,001 resected NSCLCs, while prospective validation enrolled 70 consecutive nNGM lung cancer cases across Berlin and Brandenburg sites.The discovery cohort included 581 LUAD, 402 LUSC, and 18 ASC; all prospective cases underwent comprehensive molecular profiling.Source: §4.1.2 · p. 21, §4.1.3 · p. 21

4.2 Immunohistochemical Tumor Classification

• IHC tumor subtypes were assigned by integrating marker combinations, with ≥10% positive tumor cells defining marker positivity and morphology contributing to neuroendocrine subtype criteria.The classification used TTF-1, p40, CK7, CK5/6, synaptophysin, and chromogranin A to distinguish LUAD, LUSC, ASC, LCNEC, and SCLC.Source: §4.2 · p. 22

4.3 Comparison of Tumor Cellularity Assessment Variants

• Tumor cellularity estimates from pathologists and AI were compared with KRAS variant allele frequency using Spearman correlation and mean absolute error.The AI pipeline estimated cellularity both as carcinoma cells among detected cells and as carcinoma nuclear area among total nuclear area, using the same pathologist-annotated regions.Source: §4.3 · pp. 22–23

4.4 Immunohistochemical Analysis

• IHC slides were stained using an automated immunostainer, and reference labels used binary calls for most markers, four intensity categories for MET and TROP-2, H-scores for those markers, and TPS for PD-L1.Five pathologists independently assessed membranous and cytoplasmic TROP-2 and MET by H-score during prospective validation; PD-L1 expression was assessed by tumor proportion score.Source: §4.4 · p. 23

4.5 Image Analysis Pipeline

• The image pipeline sequentially applies quality control, tissue segmentation, cell phenotyping, and—on IHC slides—expression scoring, passing masks and coordinates to restrict downstream analyses.The image-processing modules use the Atlas foundation model and follow its H&E-TME development framework; IHC cell phenotyping and PD-L1, MET, and TROP-2 scoring were additionally developed for LUCAID.Source: §4.5 · p. 23

4.6 IHC Expression Scoring

• IHC scoring classified carcinoma-cell staining intensity and aggregated cell-level predictions into slide-level H-scores or tumor proportion scores.MET and TROP-2 used negative, weak, moderate, and strong membranous categories; TROP-2 was also assessed cytoplasmically, while PD-L1 used positive or negative membranous categories.Source: §4.6 · p. 24

4.7 Agentic Orchestration and Report Generation

• LUCAID exposed each diagnostic module as an MCP tool, then used Claude Opus 4.8 to retrieve validated outputs and generate constrained readout interpretations and case-level reports.The language model did not produce quantitative values; a separate synthesis step condensed the module-derived findings into a diagnosis and treatment considerations. Threshold-free metrics were contextualized deterministically against an independent lung cancer reference distribution.Source: §4.7 · p. 24

4.8 Molecular Analysis

• Pathologists selected tumor-rich FFPE regions for DNA extraction and sequencing with a 53-gene cancer hotspot panel; variants were filtered at a 5% minimum allele frequency.Libraries used 80 ng DNA input; reads were aligned to hg19, and pathogenic or likely pathogenic variants were retained for downstream analysis.Source: §4.8 · pp. 24–25

4.9 Expert-Panel Adjudicated Reference Standard

• Adjudicated tumor cellularity was lower than the individual raters’ mean by a median −10 ppt in 86% of cases (p < 10^-9).All five pathologists jointly reviewed 70 cases after a washout of at least three months; qualitative LUCAID heatmaps were shown, but numeric model readouts and classifications were withheld, and the panel determined final scores by discussion.Source: §4.9 · pp. 25–26

4.10 Clinical Action Concordance

• The clinical action concordance index measured the proportion of cases in which each rater’s categorical assessment matched the reference classification.Categories covered tumor cellularity adequacy (< 10% versus ≥10%), PD-L1 TPS (< 1%, 1–49%, ≥50%), and MET and membranous or cytoplasmic TROP-2 H-score groups.Source: §4.10 · p. 26

4.11 Pathologic Regression Grading After Neoadjuvant Therapy

• In 140 resected cases after neoadjuvant therapy, pathologists and LUCAID assessed viable tumor, necrosis, and stroma, with agreement evaluated by correlations and mean absolute error.Pathologists estimated the three tumor-bed components, which summed to 100%, in 10% increments except amounts of 5% or less; LUCAID quantified corresponding segmented compartments.Source: §4.11 · p. 26

4.12 Spatial Feature and Survival Analysis

• Spatial niche, adjacency, and nearest-cell distance features were related individually to overall survival using UICC-stage-adjusted Cox models with false-discovery control.Features were derived from seven AI-phenotyped cell types; lymphoid niches were defined by lymphocytes plus plasma cells reaching ≥40% of community cells.Source: §4.12 · pp. 26–27

4.13 Statistical Analysis

• Module performance was assessed with per-class and macro-averaged F1 scores, while continuous predictions were evaluated by correlation and mean absolute error.Continuous comparisons used Pearson and Spearman coefficients and MAE with 95% confidence intervals from 1,000 bootstrap iterations; slides with fewer than 100 tumor cells were excluded from percentage-based analyses.Source: §4.13 · p. 27

4.14 Ethical Approval and Consent

• The study received Charité ethics approval, followed the Declaration of Helsinki, and obtained written informed consent for scientific use of archived tissue and clinical data.Race, ethnicity, and socioeconomic status were unavailable because they were not collected in routine diagnostic documentation.Source: §4.14 · p. 27

• F.K., M.A., and K.R.M. are Aignostics co-founders; F.K. is its lead medical advisor and a board member, K.R.M. its lead technical advisor, M.A. its CTO, S.S. part-time employees, and DH a scientific advisory board member.Source: Competing interests · p. 28
• The other authors declare no conflict of interest.Source: Competing interests · p. 28

Supplementary Figure 1

• Supplementary Figure 1 illustrates IHC scoring for four cytoplasmic and two nuclear markers, classifying cases with more than 10% positive tumor cells as marker-positive.Cytoplasmic positivity uses LUCAID module 3 expression scoring, while TTF-1 and p40 positivity uses nuclear DAB intensity thresholding.Source: Figure 1 · p. 42

Supplementary Figure 2

• In metastatic specimens spanning five lung carcinoma subtypes, LUCAID overlays show seven tissue compartments and nine cell classes, with F1 scores reported overall and by subtype.The figure includes H&E tumor-detection and cell-phenotyping examples from metastases of LUAD, LUSC, ASC, SCLC, and LCNEC.Source: Figure 2 · p. 43

Supplementary Figure 3; Supplementary Tables 1–2

• Supplementary Figure 3 shows molecular-validation examples of module 4 TME profiling and module 5 tumor-cellularity heatmaps, while Supplementary Tables 1–2 report cohort clinicopathological characteristics.The figure includes six representative cases selected for molecular analysis and cellularity outputs in two representative cases.Source: Figure 3 · p. 44, Table 1 · p. 47, Table 2 · p. 48

Supplementary Figure 4

• Supplementary Figure 4 extends prospective validation with rater-correlation matrices and MAE analyses across cellularity, PD-L1, MET, and TROP-2 scoring, alongside sequential module outputs.The MAE comparisons include tumor cellularity against both the expert-panel reference and KRAS allele frequency; LUCAID area-based cellularity is also shown.Source: Figure 4 · p. 45

Supplementary Figure 5

• In representative discordant cases, LUCAID estimated tumor cellularity at 7.0% and 8.8%, while all pathologists estimated 15.0–40.0%.For PD-L1, LUCAID (TPS 1.6%) and P1 (5.0%) exceeded the 1% threshold, while P2–P5 scored 0%; for c-MET, LUCAID (H-score 239.0) and P1 (260) exceeded H-score ≥200, while P2–P5 scored 170–190. For TROP-2 membranous H-score, LUCAID (270.3), P2 (220), and P4 (240) were positive; for cytoplasmic H-score, LUCAID (223.9), P1 (210), and P4 (230) were positive.Source: Figure 5 · p. 46

Supplementary Table 3

• Supplementary Table 3 reports a forest-result estimate for tissue NLR, defined as dens(granulocyte)/dens(lymphocyte): 1.37 (95% CI 1.16–1.63; p<0.001; q=0.009).Source: Table 3 · p. 49, Supplementary Tables And Supplementary… · p. 50

• Each statement is checked for faithfulness to its referenced model output and correctness of any clinical inference.Source: B Report Evaluation · p. 51
• At report level, reviewers check internal consistency and whether clinically relevant findings appear where readers would expect them.Source: B Report Evaluation · p. 51
• Reviewers split reports into atomic statements, grade each claim, assess errors’ clinical harm, and then evaluate the report as a whole.Source: B Report Evaluation · p. 51
• Errors are rated for potential clinical harm, including its extent and likelihood, so they can be prioritized by potential effect on patient care.Source: B Report Evaluation · p. 51
• Groundedness is rated as grounded, partially grounded, or ungrounded based on whether wording preserves the module output’s direction and magnitude or lacks support or contradicts it.Source: B Report Evaluation · p. 51

Diagnostic Marker Section

• Diagnostic markers above 10% are positive, and subtype classification follows marker profiles without therapeutic interpretation or molecular-testing recommendations.The prompt specifies ADC profiles using CK7, p40, and TTF1; LCNEC using CK7, synaptophysin, and chromogranin A; and SqCC using CK5/6, p40, and TTF1. Report each marker as positive or negative, without naming scoring methods.Source: C Report Prompt · p. 54

Prognostic Marker Section

• PD-L1 expression of 50–100% supports first-line immune monotherapy, while expression above 0% and below 50% supports immune therapy plus chemotherapy.The prompt also associates membranous cMET H-score > 200 and TROP-2 H-score >= 200 with potentially improved response to corresponding ADCs in clinical-trial contexts.Source: C Report Prompt · p. 54

Molecular Panel Section

• For the nNGM panel, report only genes with detected mutations, identify actionable alterations and key thresholds, and briefly note TP53 or KEAP1 as non-actionable.The prompt lists KRAS G12C, EGFR exon 19 deletion/L858R, BRAF V600E, ALK/ROS1/RET fusions, MET exon 14 skipping, ERBB2, and NTRK fusions as actionable; allele frequency >20% is clonal, while <10% may be subclonal or artifact.Source: C Report Prompt · p. 54

Reporting Instructions

• Assign predictive markers green when above therapeutic thresholds and amber when below, never red; reserve red for poor prognosis or insufficient cellularity.The stated green thresholds are PD-L1 >= 50%, MET H-score >= 200, and TROP-2 H-score >= 200; cellularity is green at >= 20%, amber at 10–19%, and red below 10%.Source: C Report Prompt · pp. 54–55
• Tissue-composition and immune-infiltration rows use null colors, while directional prognostic rows require concise, case-specific interpretation consistent with the annotation.The prompt asks these rows to describe the measure, its stated prognostic direction, and the case's qualitative value, without stating cohort percentile or tertile.Source: C Report Prompt · p. 54