PaperlayerView exact revision on arXiv

Source-linked AI summary

Disentangling Lung-Cancer CT/LDCT AI: A Systematic Evidence Map of Clinical Tasks, Evidence Chains, and Translational Gaps

Surajit Das

arXiv:2609.35240v1cs.LG

TL;DR

CT/LDCT lung-cancer AI literature groups clinically distinct tasks under broad prediction terminology, while multimodal evidence can mean different kinds of inputs. This evidence map classifies tasks and evidence roles, then finds a detection-dominated literature with few complete translational chains.

  • Problem

    CT/LDCT lung-cancer AI includes distinct clinical tasks, and molecular terminology can overstate multimodal integration when molecular data are not analytical inputs.

    Source: §I · p. 1

  • Method

    The review combines full-text task classification and role-aware modality/omics extraction with an MTEG that organizes evidence components by translational function.

    Source: D. Reporting Framework, PICOS-Based… · p. 3, J. Multi-Tier Evidence Graph · p. 6

  • Results

    Only 8 of 293 studies addressed future risk prediction, and only 31 completed all six substantive evidence tiers, with chain attrition greatest at reasoning/explanation.

    Source: §VI · p. 15, §IX · p. 16

  • Takeaways & Limitations

    The evidence map distinguishes genuine future risk prediction from detection and maps incomplete integration across translational components without treating chain completion as clinical quality or readiness.

    Source: §VI · p. 15, §IX · p. 16

    The review excluded inaccessible or insufficiently assessable full texts, so availability bias is possible; it also did not conduct a pooled clinical meta-analysis across heterogeneous tasks and outcomes.

    Source: §VIII · p. 16

Abstract

from arXiv · show

Artificial-intelligence studies using computed tomography (CT) for lung cancer are often broadly labelled "prediction" despite addressing clinically distinct tasks. We systematically mapped CT/low-dose CT (LDCT)-centered lung-cancer AI using five-database retrieval, full-text eligibility assessment, role-aware modality/omics extraction, clinical-task classification, and a Multi-Tier Evidence Graph (MTEG). The final corpus comprised 293 studies (2016-2026): 230 Detection, 8 future Risk-prediction, and 55 Other studies. Clinical variables (96.2%), 3D CT/LDCT (73.0%), and radiomics (63.5%) predominated, whereas external validation (29.0%), calibration (20.5%), decision-curve analysis (13.0%), longitudinal CT (17.7%), and saliency/attribution XAI (21.5%) were less frequent. The MTEG comprised 377 nodes and 3,444 edges; only 31 studies (10.6%) completed the six-tier substantive evidence chain, with greatest attrition at reasoning/explanation. Overall, the literature is detection-dominated, genuine future risk prediction remains uncommon, and complete translational evidence chains are rare.

I. INTRODUCTION

CT/LDCT lung-cancer AI studies address distinct clinical tasks and use “multimodal” evidence inconsistently, complicating synthesis and comparison. This review maps those tasks and evidence components, then examines whether they connect into translational pathways.

I. INTRODUCTION

  • Future cancer incidence, malignancy of an already detected nodule, and other outcomes are distinct tasks with different decision points and validation requirements.Pooling them without explicit classification may produce misleading estimates of modality use, performance, validation practice, and readiness.Source: §I · p. 1
  • Multimodality often means combining CT with clinical variables or radiomics; molecular terms alone do not establish molecular input integration.For example, predicting EGFR mutation from CT does not make genomics an input to risk estimation or lesion-malignancy assessment.Source: §I · p. 1
  • The review uses role-aware MTEGs to assess whether evidence components connect within studies, rather than merely counting models, modalities, or metrics.Its seven tiers distinguish evidence source, patient/context, input modality, representation/model, reasoning/explanation, validation/quality, and clinical output.Source: §I · pp. 1–2

II. REVIEW QUESTIONS

  • The review asks how components are represented and integrated, whether studies form complete translational chains, and how task-specific evidence architectures differ.It also examines the distribution of clinical tasks and the characteristics of CT/LDCT-based lung-cancer AI components.Source: §II · p. 2

A. Search Strategy and Corpus Assembly

  • Five databases were searched for lung-cancer CT/LDCT AI studies addressing clinically relevant endpoints, with a publication window of 2016–2026.The review describes database-specific strategies and retrieval totals in supplementary tables.Source: A. Search Strategy and Corpus Assembly · p. 2

B. Protocol Registration

The review used a retrospectively registered, auditable evidence-mapping protocol, with conservative record consolidation, explicit eligibility boundaries, and independent human review. Reviewers reported substantial pre-consensus agreement, while the analytical denominator remains publications rather than necessarily independent cohorts.

B. Protocol Registration

  • The review was retrospectively registered on 11 September 2026, after it commenced on 1 July, so the registration is not prospective preregistration.The registration records the review framework, eligibility, search and extraction strategies, task classification, evidence mapping, and planned synthesis.Source: B. Protocol Registration · p. 2

C. Deduplication and Record Consolidation

  • Conservative DOI/title matching reduced 16,916 records to 9,843, but publication deduplication does not establish whether records share patient cohorts.The final denominator represents publications; counting independent cohorts requires publication-to-cohort identity adjudication.Source: C. Deduplication and Record… · pp. 2–3

D. Reporting Framework, PICOS-Based Eligibility, and Predefined Criteria

  • An adapted PICOS framework made eligibility boundaries explicit without implying that every included study had an intervention-control comparison.Full peer-reviewed conference papers were eligible, while preprints and abstract-only records were excluded for lack of peer review or sufficient full-text evidence.Source: D. Reporting Framework, PICOS-Based… · p. 3

E. Independent Human Eligibility/Task Adjudication and Design-Specific Risk-of-Bias Assessment

  • Risk-of-bias assessment used QUADAS-3 for Detection studies and PROBAST+AI for Risk and Other studies, retaining domain judgments rather than pooling them into a numerical score.Comparisons with evidence-chain completeness were descriptive because the tools use partly different judgment labels.Source: E. Independent Human Eligibility/Task… · p. 3
  • 93.5% agreement for eligibility decisions (Cohen’s κ = 0.84) and 87.1% agreement for clinical-task classification (κ = 0.76) preceded consensus and adjudication.Disagreements involving 43 studies were resolved through discussion and, where required, third-author adjudication.Source: E. Independent Human Eligibility/Task… · p. 3

F. High-Specificity Automated Screening and Full-Text Eligibility

High-specificity screening retained only records meeting core eligibility criteria, and none of the near-boundary records entered the analytical corpus. Full-text classification then assigned each included study to one mutually exclusive task space.

F. High-Specificity Automated Screening and Full-Text Eligibility

  • 400 high-confidence candidates underwent full-text review, while 50 near-boundary records were retained only for audit and sensitivity documentation.No boundary records were promoted, and screening prioritized specificity rather than reaching a prespecified final study count.Source: F. High-Specificity Automated… · p. 3
  • Full-text eligibility review excluded 107 records, leaving a frozen 293-record corpus for downstream component, MTEG, and task-space analyses.The eligibility criteria and PICOS framework are titled in Tables I and II; the supplied table text does not provide their cell contents.Source: F. High-Specificity Automated… · p. 4, Table I · p. 4, Table II · p. 4

G. Full-Text Task Classification and Study Spaces

  • 230 studies were classified as Detection, 8 as Risk-prediction, and 55 as Other; the Non-Risk comparator combines Detection and Other and is not a fourth class.Each study was assigned by its primary predicted or analyzed endpoint rather than incidental terminology.Source: G. Full-Text Task Classification and… · p. 4

H. Role-Aware Omics Detection

The review counts omics only when molecular information is an input or predictor for an eligible lung-cancer outcome, then maps components and their connections across evidence tiers. It treats network fragmentation as descriptive and distinguishes a complete evidence chain from evidence of clinical actionability.

H. Role-Aware Omics Detection

  • Omics terms counted only when linked to an eligible lung-cancer outcome as an input, feature, predictor, or covariate; CT-to-mutation studies were not genomics-input studies when mutation was the target.The expanded molecular search characterized integration within the CT/LDCT-centered corpus, not all lung-cancer AI studies using omics.Source: H. Role-Aware Omics Detection · pp. 4–5

I. Component Co-Occurrence and Fragmentation

  • The bounded fragmentation score combines lack of integration density, lack of cross-community mixing, and isolated components, but is descriptive rather than a validated clinical metric.The component measures are interpreted alongside one another; the composite is not a stand-alone indicator of clinical or methodological quality.Source: §I · p. 5

J. Multi-Tier Evidence Graph

  • The MTEG assigns evidence-source, patient/context, input, model, reasoning, validation, and clinical-output concepts to typed tiers, with provenance retained for edges.The patient/context tier distinguishes future cancer-incidence risk from malignancy assessment of an already detected nodule.Source: J. Multi-Tier Evidence Graph · p. 5
  • A complete chain requires all six substantive tiers, while the explanation-optional five-tier chain is a separate sensitivity analysis.Clinical output describes the form of a model output, not clinical actionability or deployment readiness; even chain completion alone does not establish readiness.Source: J. Multi-Tier Evidence Graph · p. 6

K. Paper-Semantic Network Construction and Reproducibility

The paper-semantic network uses a fixed similarity threshold to create a weighted graph from the 293-study corpus. Its edges represent semantic similarity, unlike the directed, typed functional relations in the MTEG.

K. Paper-Semantic Network Construction and Reproducibility

  • The 293-paper semantic graph used normalized sentence-transformer embeddings and pairwise cosine similarity, with a 92nd-percentile threshold and a 0.25 minimum safeguard; the final threshold was 0.739488, yielding 293 nodes and 3,404 edges.The Detection, Risk, and Others subnetworks were induced from the same parent graph and used its global threshold.Source: K. Paper-Semantic Network Construction… · p. 6
  • Threshold robustness was examined across similarity quantiles 0.85, 0.88, 0.90, 0.92, 0.94, and 0.96.The MTEG is a separate directed, typed evidence graph.Source: K. Paper-Semantic Network Construction… · p. 6
  • Paper-semantic edges are undirected similarities, whereas MTEG edges point from studies to concepts or encode predefined directed cross-tier relations.MTEG does not add within-tier concept-to-concept co-occurrence edges under this rule.Source: K. Paper-Semantic Network Construction… · p. 6

L. MTEG Graph Measures and Evidence-Architecture Interpretation

The review characterizes MTEG measures as descriptive properties of evidence architecture, not clinical performance or study quality. It separately audits extraction, compares task-specific graphs, and reports how chain completeness varies across risk-of-bias groups.

M. Structured MTEG Extraction Validation

  • Full-cohort AI-assisted validation covered all 293 records across MTEG tier/concept, model/dataset, omics-role, and record-to-full-text alignment coding; it was not an independent double-human extraction round.A separate frozen 100-paper sample supported quantitative extraction-reliability analysis.Source: M. Structured MTEG Extraction Validation · pp. 6–7

N. Task-Specific MTEG Comparison

  • Task-specific MTEG projections compared Detection, Risk prediction, and Others descriptively; MTEG measures characterize evidence architecture rather than performance or study quality.The explanation-optional five-tier chain was evaluated separately and did not replace the primary six-tier definition.Source: L. MTEG Graph Measures and… · p. 6, N. Task-Specific MTEG Comparison · p. 7

A. PRISMA, Independent Review, and Risk-of-Bias Assessment

  • Two reviewers agreed on 93.5% of eligibility decisions (κ = 0.84) and 87.1% of task classifications (κ = 0.76) before discussion and adjudication of disagreements involving 43 studies.Risk-of-bias assessment covered all 293 studies: QUADAS-3 v1.2 for the 230 Detection studies and PROBAST+AI (2025) for the other 63.Source: A. PRISMA, Independent Review, and… · p. 7
  • 22/115 Low-risk studies (19.1%) completed the six-tier chain, compared with 4/88 High-risk studies (4.5%), 4/72 QUADAS-3 II studies (5.6%), and 1/18 PROBAST+AI Unclear studies (5.6%).With the explanation-optional five-tier definition, the corresponding values were 62/115 (53.9%), 25/88 (28.4%), 23/72 (31.9%), and 6/18 (33.3%); these descriptive patterns do not support causal or pooled cross-tool quality interpretation.Source: A. PRISMA, Independent Review, and… · p. 7

B. Review Accounting, Full-Text Eligibility, Task Classification, and Publication Growth

The 293-study corpus was detection-dominated, with far fewer future risk-prediction studies. Across the corpus, clinical, imaging, and radiomic components predominated, while molecular integration and several translational components were less common.

B. Review Accounting, Full-Text Eligibility, Task Classification, and Publication Growth

  • Table V summarizes the included corpus and selected methodological and translational characteristics, complementing the detailed component, semantic-network, and MTEG analyses.Source: B. Review Accounting, Full-Text… · p. 7, Table V · p. 8
  • 230 of 293 studies were Detection, 8 were future Risk-prediction, and 55 were Other; none of the 50 near-boundary records entered the analytical corpus.The 293 records followed retrieval of 16,916 records, deduplication to 9,843, and full-text review of 400 high-confidence candidates, of which 107 were excluded.Source: B. Review Accounting, Full-Text… · p. 7, Figure 1 · p. 8
  • Publication volume rose sharply after 2019 and was largest in 2025; the apparent 2026 decline reflects an incomplete calendar year in the search export.Source: B. Review Accounting, Full-Text… · p. 8

C. Data Modalities and Translational Components

  • Clinical variables appeared in 282/293 studies (96.2%), 3-D CT/LDCT in 214/293 (73.0%), and radiomics in 186/293 (63.5%).External validation appeared in 85 papers (29.0%), calibration in 60 (20.5%), decision-curve analysis in 38 (13.0%), longitudinal CT in 52 (17.7%), and saliency/attribution XAI in 63 (21.5%).Source: C. Data Modalities and Translational… · p. 8

D. Cross-Scale Integration Was Sparse

  • The strongest component co-occurrences were 3-D CT/LDCT with clinical variables (207 papers), radiomics with clinical variables (181), and 3-D CT/LDCT with radiomics (147), while molecular pairings were much smaller.The observed taxonomy coverage was 85.7%, integration density 0.565, cross-community mixing 0.100, and exploratory fragmentation score 0.445; no component node was isolated.Source: D. Cross-Scale Integration Was Sparse · p. 8

E. The MTEG Exposed an Evidence-Chain Bottleneck

The MTEG shows that components are more strongly connected from inputs to models and clinical outputs than through reasoning and validation, and that complete evidence chains are uncommon. Centrality describes graph structure rather than clinical superiority, while role-aware coding reveals that keyword searches overcount molecular inputs.

E. The MTEG Exposed an Evidence-Chain Bottleneck

F. Central Concepts and Translational Bottlenecks

G. Role Awareness Prevented Systematic Omics Inflation

  • Keyword-only searches inflated molecular-input counts: proteomics rose from 1 role-aware positive to four keyword hits, the largest reported increase.Genomics rose from 19 to 30, transcriptomics from 7 to 9, and multi-omics from 11 to 13; molecular terminology alone is not a reliable proxy for predictor inputs.Source: G. Role Awareness Prevented Systematic… · pp. 9–10, Figure 2 · p. 10

H. Structured Extraction Audit Characterized MTEG Reliability

The extraction audit found heterogeneous agreement across MTEG tiers, while stage-wise traversal localized the largest chain loss at reasoning/explanation. Full-text classification also showed that the corpus is dominated by Detection studies, with few future Risk-prediction studies.

H. Structured Extraction Audit Characterized MTEG Reliability

  • Reasoning/explanation and validation/quality extraction achieved F1 scores of 0.958 and 0.906, with Cohen’s κ of 0.935 and 0.555, respectively.Patient/context and clinical-output labels had lower κ despite relatively high recall; the audit used a frozen 100-paper subset and was distinct from human eligibility and task review.Source: H. Structured Extraction Audit… · p. 10, Table XI · p. 13

I. Stage-Wise Evidence-Chain Attrition Localized the Main Bottleneck

  • Only 61 papers remained when reasoning/explanation was added to the cumulative chain, a 24.5% conditional retention from the preceding stage.The chain then reached 46 papers with clinical output and 31 with validation; omitting reasoning/explanation yielded 116/293 (39.6%) completing the five-tier chain.Source: §I · pp. 10, 12, Figure 5 · p. 12

J. Full-Text Classification Revealed Three Clinically Distinct Study Spaces

  • Full-text classification found 230 Detection studies, 8 future Risk-prediction studies, and 55 Other applications among 293 records.Risk-prediction studies were 8/293 (2.7%; 95% CI: 1.4–5.3%), while Detection studies were 230/293 (78.5%; 95% CI: 73.4–82.8%).Source: J. Full-Text Classification Revealed… · p. 12
  • All 8 Risk-prediction papers included context, input, and clinical-output evidence, but only 2 (25.0%) included reasoning/explanation.These are marginal tier-coverage figures, not cumulative chain traversal.Source: J. Full-Text Classification Revealed… · pp. 12–13, Table XII · p. 13

K. Task-Specific MTEGs Revealed Different Translational Concept Architectures

Task-specific MTEGs describe different concept architectures, while the evidence map identifies sparse translational components and few complete evidence chains. The discussion distinguishes genuine future risk prediction from detection and emphasizes that graph structure is descriptive, not a measure of clinical maturity.

K. Task-Specific MTEGs Revealed Different Translational Concept Architectures

  • Risk had the highest concept-network density (0.786), compared with Detection (0.709) and Others (0.640).Risk, Detection, and Others had 21, 44, and 39 concept nodes, respectively; density is architectural compactness, not clinical maturity.Source: K. Task-Specific MTEGs Revealed… · p. 13

V. RESEARCH GAPS AND EVIDENCE PRIORITIES

  • External validation (29.0%), calibration (20.5%), decision-curve analysis (13.0%), and longitudinal CT (17.7%) were identified as translational gaps.The review also identified low analytical-input frequencies for genomics (6.5%), transcriptomics (2.4%), multi-omics (3.8%), and proteomics (0.3%), and a major break at reasoning/explanation and clinically grounded output.Source: §V · pp. 13, 15

VI. DISCUSSION

  • 8/293 (2.7%) studies addressed genuine future Risk prediction, versus 230/293 (78.5%) Detection and 55/293 (18.8%) Other studies.The review treats the scarcity of future-risk studies as a central clinical finding.Source: §VI · p. 15
  • 31 studies completed the six-tier substantive chain; 116/293 completed the explanation-optional C →I →M →O →V chain.The increase when explanation is optional highlights the structural importance of reasoning/explanation, while incompleteness remains in the five-tier chain.Source: §VI · p. 15
  • Role-aware extraction found that genomics, transcriptomics, proteomics, and multi-omics were uncommon as analytical inputs, limiting cross-scale integration.Molecular terms were not counted as inputs unless their analytical role was supported by source evidence.Source: §VI · p. 15
  • Task-specific Risk architecture was smaller but relatively dense; with a small Risk sample, these differences are descriptive, not evidence of superiority or clinical quality.Graph architecture and component frequency address different questions, and density or clustering should not be equated with clinical maturity.Source: K. Task-Specific MTEGs Revealed… · p. 13, §VI · p. 15
  • The review prioritizes clearer endpoint separation, longitudinal risk modeling, external and prospective validation, calibration, decision analysis, role-aware integration, and interpretable reasoning.These priorities reflect observed gaps, not proof that any model family or modality will improve outcomes.Source: §VI · p. 15

VII. VALIDATION AND GENERALIZABILITY OF THE MTEG FRAMEWORK

The framework was assessed through human review, structured extraction checks, and robustness analyses, with stronger support for broad evidence patterns than fine-grained concepts. Interpretation is bounded by publication-level counting, role-reporting and mapping uncertainty, a small Risk-prediction space, and unestablished external portability.

VII. VALIDATION AND GENERALIZABILITY OF THE MTEG FRAMEWORK

  • Validation combined independent human adjudication, a frozen 100-paper quantitative extraction audit, full-cohort checks across 293 records, and downstream robustness analyses.Full-cohort validation was AI-assisted, not an independent double-human extraction round; design-specific risk-of-bias assessment remained separate.Source: §VII · p. 15
  • Robustness checks tested similarity thresholds, graph stability, missing-link score components, and analytical feature-role collapsing.Temporal benchmarking used only information available before each historical cutoff.Source: §VII · p. 15

VIII. LIMITATIONS

  • The 293-record denominator counts publications, not necessarily independent patient cohorts; cohort independence requires publication-to-cohort identity adjudication.Source: §VIII · p. 16
  • Aggregate-tier extraction achieved macro-F1 = 0.940, while fine-grained concept normalization achieved macro-F1 = 0.713.Broad tier-level and higher-order patterns are more strongly supported than claims relying on individual normalized concepts; implicit terminology and context-dependent roles may cause mapping error.Source: §VIII · p. 16
  • Conservative role-aware omics coding may underestimate molecular integration when predictor roles are incompletely reported.Molecular evidence was counted as an input only when source evidence supported that analytical role.Source: §VIII · p. 16
  • Graph structure remains representation-dependent, Risk-prediction has a small sample, and network properties do not measure quality, causal importance, effectiveness, or deployment readiness.External portability of the taxonomy and task-specific findings has not been established.Source: §VIII · p. 16, §VII · p. 16
  • No pooled clinical meta-analysis was undertaken across heterogeneous tasks, endpoints, designs, and performance definitions; forecasting superiority was not claimed.The hand-weighted MTEG ranking did not outperform the frequency baseline on average across informative historical cutoffs.Source: §VIII · p. 16
  • Unavailable or inaccessible full texts, language-assessment limits, and exclusion of abstract-only or non-peer-reviewed reports may reduce coverage of newer work.Source: §VIII · p. 16

IX. CONCLUSION

The conclusion characterizes CT/LDCT lung-cancer AI as extensive but heterogeneous, with sparse translational components and few complete evidence chains. It emphasizes role-aware task and evidence mapping, while supporting materials document the review workflow and corpus.

IX. CONCLUSION

  • Only 31 papers formed complete six-tier substantive evidence chains, with the largest stage-wise attrition at reasoning/explanation.The broad MTEG contained 377 nodes and 3,444 edges; role-aware omics coding limited keyword-driven inflation of molecular integration.Source: §IX · p. 16
  • The corpus comprised 230 Detection, 8 Risk-prediction, and 55 Other studies; genuine future Risk prediction represented 2.7%.Task-specific MTEGs describe concept architecture, not topological superiority between differently sized task spaces.Source: §IX · p. 16
  • The contribution is a role-aware evidence architecture distinguishing task spaces and mapping translational structure, not a claim of forecasting superiority.The paper identifies an opportunity to connect context, longitudinal imaging and molecular evidence, reasoning, calibrated output, and external validation in clinically coherent studies.Source: §IX · p. 16

DATA AND CODE AVAILABILITY

  • Reproducibility materials include the frozen 293-study corpus, validation and risk-of-bias workbooks, and analysis outputs; restricted source content should not be redistributed.Source: DATA AND CODE AVAILABILITY · p. 17

Related papers

Loading 2609.35240v1…