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

Surajit Das

cs.LGSep 28, 2026

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

Authors: Surajit Das

Abstract

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.

Figures & tables

Element	Operational definition
Population	Individuals or cohorts evaluated in lung-cancer screening, pulmonary-nodule assessment, lung-cancer diagnosis, future cancer-risk estimation, prognosis, treatment-response assessment, or another prespecified clinically relevant thoracic-oncology setting.
Index / intervention	AI-, machine-learning-, deep-learning-, computer-vision-, or radiomics-based analysis in which CT or LDCT is an analytical input or central imaging component; eligible multimodal extensions could additionally incorporate clinical, longitudinal, molecular, image–text, or other complementary information.
Comparator	Radiologist assessment, pathological or clinical reference standards, conventional risk models, alternative AI/ML models, baseline algorithms, or no explicit comparator when the eligible report was primarily a model-development or validation study.
Outcomes	Future lung-cancer incidence/risk; current lesion or disease detection, segmentation, diagnosis, screening classification, or pulmonary-nodule malignancy; and prespecified related endpoints including prognosis, survival, recurrence, treatment response, staging/progression, lung function, or molecular status.
Study design	Peer-reviewed primary empirical studies reporting development, evaluation, validation, or clinically relevant application of an eligible CT/LDCT-centered computational approach. Reviews, surveys, editorials, dataset-only publications, abstract-only reports, retracted publications, and other non-eligible document types were excluded.

TABLE I: PICOS-Based Eligibility Framework for the CT/LDCT-Centered Lung-Cancer AI Evidence Map

Domain	Inclusion criterion	Exclusion criterion
Clinical scope	Lung cancer, pulmonary nodules, lung-cancer screening, or another directly relevant thoracic-oncology endpoint.	Primary focus outside lung cancer/pulmonary nodules or an off-target clinical problem without an eligible lung-cancer endpoint.
Imaging scope	CT/LDCT used as a primary analytical input or central imaging component, including eligible multimodal extensions.	No qualifying CT/LDCT contribution to the analytical model or evidence pathway.
Methodological scope	AI, machine learning, deep learning, computer vision, radiomics, or an eligible multimodal computational model.	No eligible computational modeling component or purely non-computational analysis.
Publication type	Peer-reviewed primary empirical research with sufficient methodological and outcome information.	Review/survey, editorial/commentary, dataset-only publication, abstract-only record, retracted article, non-peer-reviewed report, or another non-eligible document type.
Full-text availability	Full text sufficiently accessible to determine eligibility and extract the required evidence.	Unavailable/inaccessible full text or a document too incomplete for reliable eligibility assessment.
Language	Full text assessable under the review protocol which considers English or in a language that could be reliably assessed by the review team.	Full text not assessable under the review language protocol.

TABLE II: Predefined Full-Text Inclusion and Exclusion Criteria

Task	Classification criterion
Risk prediction	Estimates an individual’s susceptibility to, or probability of, developing lung cancer in the future, before an established malignant lesion is being evaluated. This category includes future lung-cancer incidence, susceptibility, individualized risk estimation, and risk stratification. The defining feature is the future occurrence of lung cancer as the predicted endpoint, rather than characterization of a lesion already present at the time of assessment.
Detection	Identifies, localizes, segments, diagnoses, or characterizes lung cancer or pulmonary lesions that are already present at the time of assessment. This category includes lesion/nodule detection and segmentation, screening detection, diagnostic classification, benign-versus-malignant classification, and malignancy assessment of an existing pulmonary nodule. Accordingly, models estimating “malignancy probability,” “risk of malignancy,” or the likelihood that an existing nodule is malignant are classified as Detection rather than Risk prediction, despite the use of the term “risk.”
Others	Includes clinically relevant endpoints that concern neither future occurrence of lung cancer nor detection/characterization of currently present cancer or pulmonary lesions. These include prognosis, overall/progression-free/disease-free survival, recurrence, mortality, treatment response, treatment toxicity or adverse events, cancer staging or progression, pulmonary/lung function, and molecular or mutation-status prediction (e.g., EGFR, ALK, KRAS, or PD-L1), as well as other eligible endpoints outside the preceding two categories.

TABLE III: Operational Criteria for the Three Mutually Exclusive Task Spaces

Measure	Definition	Interpretation in evidence synthesis
Component prevalence	$P(v)=n_{v}/N$	Frequency of each evidence component across studies.
Edge frequency	$w_{ij}=\sum_{s}I_{si}I_{sj}$ ; $\hat{w}_{ij}=w_{ij}/N$	Frequency/proportion with which two components are integrated.
Degree	$k_{i}=\sum_{j\neq i}A_{ij}$ [ 67 ]	Breadth of distinct direct component connections.
Strength	$s_{i}=\sum_{j\neq i}w_{ij}$ [ 68 ]	Cumulative recurrence of a component’s connections.
Betweenness	$BC(v)=\sum_{s\neq v\neq t}\sigma_{st}(v)/\sigma_{st}$ [ 69 ]	Bridging position between evidence regions.
Density	$D=2\|E\|/[\|V\|(\|V\|-1)]$ [ 67 ]	Proportion of possible pairwise connections observed.

TABLE IV: Core graph-theoretic and evidence-chain measures used to characterize the MTEG evidence architecture.

Characteristic	Included studies
Total analytical corpus	293
Detection	230 (78.5%)
Future Risk prediction	8 (2.7%)
Other clinical tasks	55 (18.8%)
Clinical variables	282 (96.2%)
3-D CT/LDCT	214 (73.0%)

TABLE V: Overview of the Included Analytical Corpus

Component	$N$	%	95% CI	Component	$N$	%	95% CI
Clinical variables	282	96.2	93.4–97.9	External validation	85	29.0	24.1–34.5
3-D CT/LDCT	214	73.0	67.7–77.8	Calibration	60	20.5	16.3–25.5
Radiomics	186	63.5	57.8–68.8	Decision-curve analysis	38	13.0	9.6–17.3
Saliency/attribution XAI	63	21.5	17.2–26.6	Longitudinal CT	52	17.7	13.8–22.5
Genomics	19	6.5	4.2–9.9	Transcriptomics	7	2.4	1.2–4.8
Multi-omics	11	3.8	2.1–6.6	Proteomics	1	0.3	0.1–1.9

TABLE VI: Principal Component Prevalence With Descriptive 95% Wilson Binomial Reference Intervals

Tier/role	Component	$N$	%	Tier/role	Component	$N$	%
Input	Clinical variables [ 1 ]	282	96.2	Validation	External validation [ 8 ]	85	29.0
Input	3-D CT/LDCT [ 16 ]	214	73.0	Validation	Calibration [ 6 ]	60	20.5
Input	Radiomics [ 6 ]	186	63.5	Validation	Decision-curve analysis [ 6 ]	38	13.0
Input	2-D CT/LDCT [ 2 ]	108	36.9	Reasoning	Saliency/attribution XAI [ 9 ]	63	21.5
Input	Longitudinal CT [ 24 ]	52	17.7	Reasoning	First-order logic/rules † [ 48 ]	32	10.9
Input	Genomics [ 77 ]	19	6.5	Input	Multi-omics [ 21 ]	11	3.8

TABLE VII: Selected Modalities, Methods, and Translational Components in the 293-Record Corpus. Citations illustrate one representative study per component;

N

and percentages are corpus-wide frequencies.

Fig. 2: Projected component co-occurrence network. Node size reflects prevalence; edge weight reflects paper-level co-use. Omics concepts are visible as a small, weakly connected molecular subnetwork.

Fig. 3: Role-aware Multi-Tier Evidence Graph (MTEG) of the lung-cancer CT/LDCT AI evidence architecture. Concepts are arranged from left to right by functional role: Evidence source identifies study provenance; Patient/context defines the clinical population, setting, prediction target, and horizon; Input modality represents clinical, imaging, radiomic, longitudinal, molecular, and multimodal information supplied to models; Representation/model captures machine-learning and representation-learning approaches; Reasoning/explanation includes attribution, explainability, rule-based, ontology, knowledge-graph, and grounding mechanisms; Clinical output represents decision-relevant predictions such as diagnosis, malignancy probability, cancer-risk score, and uncertainty; and Validation/quality captures internal/external validation, calibration, decision analysis, prospective evaluation, and related reliability evidence. Node size reflects PageRank-based structural prominence. Directed edges denote paper-supported relationships oriented according to the predefined functional cross-tier relation map. Edge opacity is reduced in highly congested regions and increased in sparsely connected regions solely to improve visual readability and should not be interpreted as a quantitative measure of evidence strength. The graph therefore visualizes how methodological components connect across the translational pathway and where evidence chains remain incomplete.

Fig. 4: Weighted MTEG tier-connectivity matrix. Strong input-to-model and input-to-output links contrast with sparse reasoning-related pathways.

Measure	Value
Mapped papers	293
MTEG nodes	377
MTEG edges	3,444
Complete six-tier chains	31
Complete-chain prevalence	10.6%
Median substantive tiers covered	5

TABLE VIII: MTEG Evidence-Chain Summary

Concept	Papers	Degree	Between.
Classical ML	165	0.253	0.004
CNN	210	0.253	0.009
Binary diagnosis	113	0.398	0.000
Cancer-risk score	61	0.434	0.000
Uncertainty estimate	86	0.398	0.000
Nodule malignancy prob.	70	0.422	0.000

TABLE IX: Selected central concepts in the updated role-aware MTEG. Supporting-paper counts and retained centrality measures are taken from the frozen 293-paper MTEG export.

Network aspect	Observed result	Biomedical/AI interpretation
Paper-level semantic network	293 papers; 3,404 edges; density 0.080	The literature is selectively interconnected rather than uniformly similar; most possible paper–paper relationships are absent.
Connected components	56 connected components	The evidence architecture contains multiple structurally separated regions.
Local organization	Mean clustering coefficient 0.346	Similar papers tend to form local methodological neighborhoods, consistent with recurring combinations of related approaches.
MTEG evidence architecture	377 nodes; 3,444 edges	The extracted evidence forms a large relational architecture linking clinical context, inputs, models, reasoning/explanation, outputs, and validation components.
Evidence-chain completeness	31/293 complete chains (10.6%); median 5 tiers	Most studies cover several translational elements, but only about one in ten connects all six substantive tiers end-to-end.
Largest chain bottleneck	Model $\rightarrow$ reasoning/explanation: 249 $\rightarrow$ 61 studies; conditional retention 24.5%	Explicit reasoning/explanation is the sharpest break in the cumulative translational chain, despite broad representation/model coverage.

TABLE X: Summary of principal findings in the final 293-study corpus and their interpretation for biomedical-AI evidence synthesis.

Tier	Precision	Recall	F1	$\kappa$
Patient/context	0.928	0.978	0.952	0.144
Input modality	0.990	1.000	0.995	0.000
Representation/model	1.000	0.948	0.973	0.593
Reasoning/explanation	1.000	0.919	0.958	0.935
Clinical output	0.893	0.827	0.859	0.362
Validation/quality	1.000	0.828	0.906	0.555

TABLE XI: Frozen 100-Paper Tier-Level MTEG Extraction Audit. Automated tier assignments are compared with the frozen reference full-text coding used for the structured extraction audit; this audit is distinct from the completed independent human eligibility/task/risk-of-bias review.

Fig. 5: Stage-wise attrition across the substantive translational evidence chain. The dominant loss occurs at the transition from representation/model to reasoning/explanation.

	Detection		Risk		Others
Tier	$N$	%	$N$	%	$N$	%
Patient/context	225	97.8	8	100.0	50	90.9
Input modality	227	98.7	8	100.0	55	100.0
Representation/model	207	90.0	5	62.5	46	83.6
Reasoning/explanation	50	21.7	2	25.0	18	32.7
Clinical output	179	77.8	8	100.0	26	47.3

TABLE XII: MTEG Tier Coverage Across the Three Full-Text Task Spaces. These are marginal per-tier coverage counts; they should not be confused with cumulative evidence-chain traversal, which requires all preceding tiers.

Fig. 6: Task-specific MTEG concept networks for Detection, Risk prediction, and Other applications. Numbered nodes represent methodological and translational concepts, with the corresponding concept names provided in the panel legends. Node size reflects weighted concept strength, while edge width reflects the strength of concept co-occurrence. Detection uses a three-column legend, whereas Risk and Others use two-column legends. These networks describe the organization of methodological and translational concepts rather than the organization of individual papers.

Metric	Detection	Risk	Others
Papers	230	8	55
Concept nodes	44	21	39
Concept edges	671	165	474
Density	0.709	0.786	0.640
Connected components	1	1	1
Average clustering	0.053	0.259	0.077

TABLE XIII: Task-Specific MTEG Concept-Network Metrics. Density and clustering refer to the task-specific concept graphs and are therefore distinct from the paper-level semantic-subnetwork metrics in the network comparison.

D1 (All 16916 records)	D2 (Deduplicated records 9843)	D3 (Auto excluded record 9393)	D4 (Eligible Candidates 293)
Database-specific search strategies used for literature retrieval.	Obtained after the deduplication process.	Exclusion following low-confidence text assessment.	Frozen analytical corpus containing the final included studies and stable Record_ID information used for downstream analyses..
D5 (Inelligible Candidates 107)	D6 (Backup records 50)	D7 (MTEG validation 293)	D8 (Risk of Bias 293)
Individual exempted according to specified exclusion principles).	PRISMA 2020 reporting crosswalk for the systematic evidence map.	Consolidated MTEG manual validation Consolidated Risk of Bias (RoB) report

TABLE S1: Master index of supplementary data, review records, validation materials, and reproducibility artifacts. All documents are available at https://github.com/SurajitDaz/MTEG_review .

Database-specific search strategy
PubMed ( "Lung Neoplasms"[Mesh] OR "lung cancer"[Title/Abstract] OR "lung carcinoma"[Title/Abstract] OR "pulmonary cancer"[Title/Abstract] OR "lung neoplasm"[Title/Abstract] OR "pulmonary nodule"[Title/Abstract] OR "lung nodule"[Title/Abstract] OR "non-small cell lung cancer"[Title/Abstract] OR NSCLC[Title/Abstract] OR "small cell lung cancer"[Title/Abstract] OR SCLC[Title/Abstract] OR "lung adenocarcinoma"[Title/Abstract] OR "lung squamous cell carcinoma"[Title/Abstract] ) AND ( "Tomography, X-Ray Computed"[Mesh] OR "computed tomography"[Title/Abstract] OR "low-dose computed tomography"[Title/Abstract] OR "low-dose CT"[Title/Abstract] OR LDCT[Title/Abstract] OR "chest CT"[Title/Abstract] OR radiomics[Title/Abstract] OR "CT imaging"[Title/Abstract] ) AND ( "Artificial Intelligence"[Mesh] OR "artificial intelligence"[Title/Abstract] OR "machine learning"[Title/Abstract] OR "deep learning"[Title/Abstract] OR "neural network"[Title/Abstract] OR "computer vision"[Title/Abstract] OR "computer-aided diagnosis"[Title/Abstract] OR "large language model"[Title/Abstract] OR "vision-language model"[Title/Abstract] OR multimodal[Title/Abstract] OR "multi-modal"[Title/Abstract] ) AND ( predict[Title/Abstract] OR prognos[Title/Abstract] OR "risk prediction"[Title/Abstract] OR "risk stratification"[Title/Abstract] OR "malignancy prediction"[Title/Abstract] OR "malignancy classification"[Title/Abstract] OR diagnos*[Title/Abstract] OR screening[Title/Abstract] ) AND ("2016/01/01"[Date - Publication] : "2026/12/31"[Date - Publication])
IEEE Xplore ("All Metadata":"lung cancer" OR "All Metadata":"lung carcinoma" OR "All Metadata":"pulmonary cancer" OR "All Metadata":"lung neoplasm" OR "All Metadata":"pulmonary nodule" OR "All Metadata":"lung nodule") AND ("All Metadata":"computed tomography" OR "All Metadata":"low-dose computed tomography" OR "All Metadata":"low-dose CT" OR "All Metadata":LDCT OR "All Metadata":"chest CT" OR "All Metadata":radiomics OR "All Metadata":"CT imaging") AND ("All Metadata":"artificial intelligence" OR "All Metadata":"machine learning" OR "All Metadata":"deep learning" OR "All Metadata":"neural network" OR "All Metadata":"computer vision" OR "All Metadata":"large language model" OR "All Metadata":"vision-language model" OR "All Metadata":multimodal) AND ("All Metadata":prediction OR "All Metadata":prognosis OR "All Metadata":"risk prediction" OR "All Metadata":"risk stratification" OR "All Metadata":"malignancy prediction" OR "All Metadata":"malignancy classification" OR "All Metadata":diagnosis OR "All Metadata":screening)
ACM Digital Library [[All: "lung cancer"] OR [All: "lung carcinoma"] OR [All: "pulmonary cancer"] OR [All: "lung neoplasm"] OR [All: "pulmonary nodule"] OR [All: "lung nodule"]] AND [[All: "computed tomography"] OR [All: "low-dose computed tomography"] OR [All: "low-dose ct"] OR [All: ldct] OR [All: "chest ct"] OR [All: radiomics] OR [All: "ct imaging"]] AND [[All: "artificial intelligence"] OR [All: "machine learning"] OR [All: "deep learning"] OR [All: "neural network"] OR [All: "computer vision"] OR [All: "large language model"] OR [All: "vision-language model"] OR [All: multimodal]] AND [[All: prediction] OR [All: prognosis] OR [All: "risk prediction"] OR [All: "risk stratification"] OR [All: "malignancy prediction"] OR [All: "malignancy classification"] OR [All: diagnosis] OR [All: screening]] AND [E-Publication Date: (01/01/2016 TO 12/31/2026)]
Embase/Ovid ( exp lung cancer/ OR exp lung tumor/ OR exp non small cell lung cancer/ OR exp small cell lung cancer/ OR exp pulmonary nodule/ OR ((lung OR pulmonary) adj3 (cancer $OR carcinoma$ OR neoplasm $OR tumor$ OR tumour $OR malignan$ )).ti,ab,kw. OR (lung adj3 nodule $).ti,ab,kw. OR NSCLC.ti,ab,kw. OR SCLC.ti,ab,kw. ) AND ( exp computed tomography/ OR exp low dose computed tomography/ OR ("computed tomography" OR "low-dose computed tomography" OR "low dose computed tomography" OR "low-dose CT" OR "low dose CT" OR LDCT OR "chest CT" OR radiomics OR "CT imaging").ti,ab,kw. ) AND ( exp artificial intelligence/ OR exp machine learning/ OR exp deep learning/ OR ("artificial intelligence" OR "machine learning" OR "deep learning" OR "neural network$ " OR "computer vision" OR "large language model $" OR "vision-language model$ " OR multimodal OR "multi-modal").ti,ab,kw. ) AND ( (predict $OR prognos$ OR "risk prediction" OR "risk stratification" OR "malignancy prediction" OR "malignancy classification" OR diagnos$ OR screening).ti,ab,kw. ) AND ( 2016:2026 ).yr.

TABLE S2: Database-specific search strategies used for literature retrieval.

Item	Reported information	Status / qualification
Final search date	21-Aug-2026	Same date reported in the full manuscript.
Publication window	2016–2026	Represented in the query as PUBYEAR > 2015 and PUBYEAR < 2027 .
Database-specific retrieved records	Scopus: 6,963; PubMed: 3,184; Embase: 4,707; IEEE Xplore: 1,774; ACM: 288; total: 16,916	Database-specific search strategies are reported in Supplementary Table S2 .
Language rule	Full text assessable under the review protocol: English or another language that could be reliably assessed by the review team.	Not described as an English-only database search filter.
Publication-type rule	Peer-reviewed primary empirical studies; eligible full conference papers allowed; reviews/surveys, editorials/commentaries, dataset-only publications, abstract-only records, retracted articles, preprints/non-peer-reviewed reports excluded.	From the predefined full-text criteria in the prior manuscript.
Deduplication	Hierarchical exact normalized DOI followed by exact normalized title; title match required $\geq 20$ normalized characters and no non-empty DOI/year conflict; transitive groups consolidated by Union–Find; bibliographically richest record retained; ambiguous conflicts retained for audit.	Conservative deduplication; 16,916 raw records reduced to 9,843.

TABLE S3: Search, screening, deduplication, and full-text eligibility procedures used to construct the analytical corpus.

Exclusion category	Count
Conference-abstract-only	20
Residual duplicate	3
Anonymous/inadequately attributable document	0
Document shorter than four pages	10
Non-peer-reviewed report	5
Unavailable full text	12

TABLE S4: Aggregate reasons for exclusion following full-text assessment, reconstructed from the final eligibility and selection-comment fields.

Record ID	Year	Study	Model/approach	Venue
REC_05652	2020	Convolutional Neural Network ensembles for accurate lung nodule malignancy prediction 2 years in the future	CNN; ResNet; VGG; ensemble	Computers in Biology and Medicine
REC_04313	2022	Deep Learning to Optimize Candidate Selection for Lung Cancer CT Screening: Advancing the 2021 USPSTF Recommendations	CNN	Radiology
REC_04101	2023	Sybil: A Validated Deep Learning Model to Predict Future Lung Cancer Risk from a Single Low-Dose Chest Computed Tomography	CNN	Journal of Clinical Oncology
REC_01400	2025	Radiomics for Dynamic Lung Cancer Risk Prediction in USPSTF-Ineligible Patients	LASSO/radiomics-based modelling	Cancers

TABLE S5: Individual Risk-Prediction Studies in the Final Analytical Corpus (Part I)

Record ID	Year	Study	Model/approach	Venue
REC_01476	2025	Significance of Image Reconstruction Parameters for Future Lung Cancer Risk Prediction Using Low-Dose Chest Computed Tomography and the Open-Access Sybil Algorithm	AI-based future-risk prediction	Investigative Radiology
REC_01674	2025	External Testing of a Deep Learning Model for Lung Cancer Risk from Low-Dose Chest CT	Deep-learning risk model; external testing	Radiology
REC_01865	2025	Deep learning-based lung cancer risk assessment using chest computed tomography images without pulmonary nodules $\geq 8$ mm	CNN; DenseNet; EfficientNet; U-Net/nnU-Net	Translational Lung Cancer Research
REC_02099	2025	Predicting Future Lung Cancer Risk in Low-Dose CT Screening Patients with AI Tools	CNN; VGG	Proceedings of SPIE

TABLE S6: Individual Risk-Prediction Studies in the Final Analytical Corpus (Part II)

Item	Reporting topic	Location/status
1–4	Title, abstract, rationale, objectives	Title/Abstract, Introduction, and five Review Questions.
5–7	Eligibility, information sources, search	Tables I – II ; five databases; final search 21 August 2026; Fig. .
8–10	Selection, collection, data items	Deduplication/screening/full-text Methods; role-aware extraction; Table III .
11	Study risk of bias	Completed for all 293 studies: QUADAS-3 v1.2 for 230 Detection studies and PROBAST+AI (2025) for 63 Risk/Other studies.
12–13	Effect/synthesis methods	No pooled clinical effect measure; descriptive prevalence, semantic/network analyses, permutation and matched resampling specified.
14–15	Reporting bias/certainty	No pooled publication-bias or GRADE claim; selection and evidence limitations are explicitly reported.

TABLE S7: PRISMA 2020 Reporting Crosswalk for This Systematic Evidence Map

Explore similar work

Benchmarking Deep Learning Models for Laryngeal Cancer Staging Using the LaryngealCT Dataset
Oct 13, 2025Nivea Roy, Son Tran, Atul Sajjanhar +4Deep Learning
School of Information Technology, Deakin University, Burwood, Australia · Department of Head and Neck Surgery, Kasturba Medical College, Manipal Academy of Higher Education, Manipal, India · Department of Radiodiagnosis and Imaging, Kasturba Medical College, Manipal Academy of Higher Education, Manipal, India +1
Foundation Models vs. Radiomics for Lung Computed Tomography: A Benchmark of Feature Extractors, Classification Heads, and Segmentation Choices
Jul 1, 2026Nils Neukirch, Martin Maurer, Nils StrodthoffRadiomics Cohort
Division AI4Health, Carl von Ossietzky Universität Oldenburg, Ammerländer Heerstraße 114-118, Oldenburg 26129, Niedersachsen, Germany · University Institute for Diagnostic and Interventional Radiology, Klinikum Oldenburg AöR, Rahel-Straus-Straße 10, Oldenburg 26133, Niedersachsen, Germany
Trusting Right Predictions for Wrong Reasons: A LIME Based Analysis of Deep Learning Interpretability in Lung Cancer Diagnosis
Jun 14, 2026Samarpan Poudel, Vladislav D VekslerCancer Detection Clinical Prediction
Caldwell University School of Business and Computer Science Caldwell, NJ 07006

Oct 13, 2025, cs.CVJ/K move · Enter open · S save

Benchmarking Deep Learning Models for Laryngeal Cancer Staging Using the LaryngealCT Dataset

Nivea Roy, Son Tran, Atul Sajjanhar, K. Devaraja+3

School of Information Technology, Deakin University, Burwood, Australia · Department of Head and Neck Surgery, Kasturba Medical College, Manipal Academy of Higher Education, Manipal, India · Department of Radiodiagnosis and Imaging, Kasturba Medical College, Manipal Academy of Higher Education, Manipal, India+1

Deep Learning

Open paper AskPDF arXiv

Laryngeal cancer imaging research lacks standardised public datasets to enable reproducible deep learning (DL) model development. We present LaryngealCT, a curated benchmark of 1,029 computed tomography (CT) scans aggregated from six collections from The Cancer Imaging Archive (TCIA). Uniform 1 mm isotropic volumes of interest encompassing the larynx were extracted using a weakly supervised parameter search framework validated by clinical experts. Six 3D DL architectures (custom 3D CNN, ResNet18,50,101, DenseNet121 and MedicalNet-pretrained ResNet50) were benchmarked on (i) early (Tis,T1,T2) vs. advanced (T3,T4) and (ii) T4 vs. non-T4 classification tasks. On the independent test set, the 3D CNN achieved the strongest overall performance across global and per-class metrics (Accuracy 0.854, F1-macro 0.841) in early vs. advanced classification. In the T4 task, AU-ROC values exceeded 0.82 for most models, but sensitivity for T4 disease remained limited (less than or equal to 0.412), with ResNet101 showing the most promising calibrated T4 recall (0.706. Model explainability assessed using GradCAMpp with thyroid cartilage overlays for T4 classification task revealed anatomically plausible peri-cartilage activations, although spatial overlap was modest. Through open-source data, pretrained models, and integrated explainability tools, LaryngealCT offers a reproducible foundation for AI-driven research to support future clinical decision-making in laryngeal oncology.

Abstract

Explore similar work

Benchmarking Deep Learning Models for Laryngeal Cancer Staging Using the LaryngealCT Dataset

Foundation Models vs. Radiomics for Lung Computed Tomography: A Benchmark of Feature Extractors, Classification Heads, and Segmentation Choices

Trusting Right Predictions for Wrong Reasons: A LIME Based Analysis of Deep Learning Interpretability in Lung Cancer Diagnosis