arxiv: 2604.23679 · v1 · submitted 2026-04-26 · 🧬 q-bio.GN

Recognition: unknown

Imaging Exploration of Molecular Subtypes in Tongue Squamous Cell Carcinoma

Hao Pan , Peipei Wang , Yajie Chang , Bingyi Lu , Yunyan Jiang , Mengfan Wang , Xinyue Wang , Xinrou Yang

show 4 more authors

Jiyuan Zhang Yu Liu Andrei Velichko Yuanjun Wang

Authors on Pith no claims yet

Pith reviewed 2026-05-08 04:44 UTC · model grok-4.3

classification 🧬 q-bio.GN

keywords radiomicsmolecular subtypestongue squamous cell carcinomatranscriptomicsradiogenomicstexture featureswavelet analysisnon-invasive imaging

0 comments

The pith

Radiomic texture features from preoperative imaging distinguish molecular subtypes of tongue squamous cell carcinoma.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper uses unsupervised consensus clustering on transcriptomic data from 60 TSCC cases to define two stable molecular subtypes, C1 and C2, whose main biological differences involve squamous epithelial differentiation, inflammatory signaling, lipid metabolism, and immune pathway enrichment in C2. It then extracts radiomic features from matched preoperative images with manually annotated tumor regions and finds ten features that differ significantly between the subtypes, primarily wavelet-derived texture measures from gray-level size zone, dependence, co-occurrence, and run length matrices. A sympathetic reader would care because molecular subtyping currently depends on invasive biopsies that capture only limited portions of heterogeneous tumors, whereas routine imaging is available for most patients. If the radiomic differences hold, imaging could provide a non-invasive window into the same biological distinctions that drive variable outcomes and treatment choices.

Core claim

Two stable molecular subtypes, C1 and C2, were identified from transcriptomic data. Their biological differences center on squamous epithelial differentiation, inflammatory signaling, and lipid metabolism, with C2 showing greater enrichment of immune-related pathways. Ten radiomic features differed significantly between the subtypes, mainly wavelet-derived texture features from gray-level size zone, dependence, co-occurrence, and run length matrices (P values 0.00202 to 0.0162). These results support radiomics as a non-invasive approach to characterize molecular heterogeneity in TSCC and supply an initial radiogenomic framework for preoperative assessment.

What carries the argument

The integrated transcriptomic-radiomics framework that defines subtypes via unsupervised consensus clustering on gene expression data and then compares them to PyRadiomics features extracted from manually segmented tumor regions on preoperative scans.

If this is right

Radiomics can serve as a non-invasive proxy for detecting molecular subtypes in TSCC.
Wavelet-derived texture features are the imaging markers most aligned with the transcriptomic differences in differentiation and immune pathways.
Preoperative imaging phenotypes can reflect intrinsic molecular heterogeneity across the whole tumor.
This framework enables biologically informed assessment before surgery or other interventions without additional tissue sampling.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

If validated, the approach could reduce reliance on focal biopsies by using whole-tumor imaging to infer subtype and guide initial therapy selection.
Spatial maps of the discriminating radiomic features might highlight intra-tumor regions with differing molecular profiles that a single biopsy would miss.
The same radiogenomic comparison could be tested in other head and neck squamous cell carcinomas or solid tumors that exhibit molecular subtype heterogeneity.

Load-bearing premise

Manually annotated tumor regions on preoperative imaging capture the same molecular profile as the tissue samples used for transcriptomic analysis, despite possible intra-tumor heterogeneity and sampling mismatch.

What would settle it

An independent cohort of TSCC cases with both transcriptomic profiling and preoperative imaging in which the ten reported radiomic features show no statistically significant differences between the same subtypes.

read the original abstract

Tongue squamous cell carcinoma (TSCC) is an aggressive malignancy with marked biological heterogeneity and variable clinical outcomes. Although molecular profiling has improved understanding of TSCC heterogeneity, its clinical use remains constrained by invasive tissue sampling and limited representation of whole-tumor spatial complexity. Meanwhile, most radiomics studies in TSCC have focused on downstream clinical endpoints, and whether imaging can non-invasively reflect intrinsic molecular subtypes remains unclear. In this study, an integrated transcriptomic-radiomics framework was used to investigate the relationship between preoperative imaging phenotypes and molecular subtypes in TSCC. Transcriptomic data from 60 TSCC cases in The Cancer Genome Atlas were analyzed using unsupervised consensus clustering, followed by differential expression and functional enrichment analyses. Matched preoperative imaging data from The Cancer Imaging Archive were manually annotated for primary tumor regions, and radiomic features were extracted using PyRadiomics; group differences were assessed with the U-test. Two stable molecular subtypes, C1 and C2, were identified. Their biological differences were mainly associated with squamous epithelial differentiation, inflammatory signaling, and lipid metabolism, with C2 showing greater enrichment of immune-related pathways. In addition, 10 radiomic features differed significantly between the two subtypes, mainly wavelet-derived texture features from gray-level size zone, dependence, co-occurrence, and run length matrices (P=0.00202-0.0162). These findings support the potential of radiomics as a non-invasive approach for characterizing molecular heterogeneity in TSCC and provide an initial radiogenomic framework for biologically informed preoperative assessment.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

This paper applies standard consensus clustering and PyRadiomics to public TCGA/TCIA data and reports ten texture features that differ between two TSCC molecular subtypes, but the link rests on a small unvalidated cohort with an untested assumption about tumor sampling alignment.

read the letter

The core finding is that two transcriptomic subtypes (C1 and C2) show differences in epithelial differentiation, inflammation, and lipid metabolism, and that ten mostly wavelet-derived radiomic features extracted from manually segmented preoperative images differ between them at nominal p-values between 0.002 and 0.016. The work uses off-the-shelf tools on public datasets, which keeps the analysis reproducible in principle and gives a concrete example of radiogenomics applied to tongue cancer rather than a new method or framework. That is the main thing the paper actually delivers: an initial set of candidate imaging markers tied to unsupervised molecular clusters in this one cancer type. The execution is competent for an exploratory study; the clustering appears stable and the functional enrichment is reported clearly enough to show biological plausibility. The radiomic feature list is also specific enough that others could test the same features on new data. The sample is only 60 cases with no held-out or external validation cohort. No correction for multiple testing across the large radiomic feature space is described, so the reported p-values are hard to interpret at face value. The central assumption—that the full manually annotated tumor volume on imaging matches the molecular profile from the (usually small, non-image-guided) tissue aliquot used for RNA-seq—is not tested. Tongue tumors are known to be spatially heterogeneous, and without multi-region sampling or spatial registration the group differences could reflect sampling bias rather than a stable radiogenomic signal. This is a real soft spot, not a minor one, because it directly undercuts the claim of non-invasive subtype characterization. The paper is mainly useful to radiogenomics groups working on head and neck cancers who want to see the standard pipeline run on TSCC and to generate hypotheses for follow-up. It is not strong enough on its own for broad clinical claims. A serious editor should send it for peer review because the data are public, the question is well-posed, and the reported associations are falsifiable with new cohorts; referees will need to press on validation, multiple-testing control, and the sampling alignment issue. I would not cite it yet and would only bring it to a reading group to discuss the heterogeneity problem.

Referee Report

3 major / 3 minor

Summary. The manuscript identifies two stable molecular subtypes (C1 and C2) of tongue squamous cell carcinoma via unsupervised consensus clustering on TCGA RNA-seq data from 60 cases, characterizes their differences through differential expression and pathway enrichment (focusing on epithelial differentiation, inflammation, and lipid metabolism), extracts radiomic features from manually segmented primary tumor regions on matched TCIA preoperative images using PyRadiomics, and reports that 10 features (primarily wavelet-derived texture features from GLSZM, GLDM, GLCM, and GLRLM matrices) differ significantly between subtypes by Mann-Whitney U-test (p=0.00202–0.0162), proposing radiomics as a non-invasive tool for molecular subtyping.

Significance. If the radiogenomic associations hold after addressing methodological gaps, the work offers an initial framework linking preoperative imaging phenotypes to transcriptomic subtypes in TSCC, potentially reducing reliance on invasive sampling and enabling biologically informed preoperative planning. The use of public datasets, standard consensus clustering, and PyRadiomics extraction are strengths, but the small n and empirical nature limit broader impact without validation.

major comments (3)

[Results (radiomic feature differences)] In the radiomics results section: the 10 reported U-test p-values (0.00202–0.0162) are given without multiple-testing correction despite extraction of a high-dimensional feature set (typically >100 features via PyRadiomics); this is load-bearing for the central claim of significant radiomic differences and risks false positives.
[Methods (data integration) and Discussion] In the methods and discussion of imaging-transcriptomic integration: the claim that radiomic features from manually annotated full tumor volumes reflect the molecular subtypes rests on the untested assumption that these annotations match the (typically small, non-image-guided) tissue aliquots used for TCGA RNA-seq, without addressing intra-tumor heterogeneity or spatial mismatch; this directly undermines interpretability of the group differences.
[Results (molecular subtypes and radiomics)] In the subtype identification and radiomics association results: with n=60, no independent validation cohort, and post-hoc feature comparison after clustering, the stability of C1/C2 and the reported radiomic associations lack external confirmation and are vulnerable to overfitting or sampling bias.

minor comments (3)

[Methods (radiomics)] Specify the exact imaging modalities (CT vs. MRI) and any PyRadiomics preprocessing parameters (e.g., bin width, resampling) used for feature extraction to improve reproducibility.
[Abstract and Results] The abstract lists p-value range without indicating whether these are the extremal values among the 10 or providing the full list; add a supplementary table with all tested features and exact p-values.
[Methods (radiomics)] Clarify the number of radiomic features initially extracted versus those retained after any filtering, and state whether feature selection occurred before or after the U-tests.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for the detailed and constructive review of our manuscript. We have addressed each of the major comments point-by-point below, agreeing where appropriate and outlining revisions to improve the rigor and clarity of our work.

read point-by-point responses

Referee: In the radiomics results section: the 10 reported U-test p-values (0.00202–0.0162) are given without multiple-testing correction despite extraction of a high-dimensional feature set (typically >100 features via PyRadiomics); this is load-bearing for the central claim of significant radiomic differences and risks false positives.

Authors: We acknowledge the importance of multiple testing correction in high-dimensional data analysis. The original manuscript presented uncorrected p-values to highlight candidate features in an exploratory context. In the revised version, we will implement the Benjamini-Hochberg procedure for FDR correction and include both raw and adjusted p-values in the Results section. If fewer than 10 features remain significant, we will update the text and figures accordingly and discuss the exploratory nature of the findings. revision: yes
Referee: In the methods and discussion of imaging-transcriptomic integration: the claim that radiomic features from manually annotated full tumor volumes reflect the molecular subtypes rests on the untested assumption that these annotations match the (typically small, non-image-guided) tissue aliquots used for TCGA RNA-seq, without addressing intra-tumor heterogeneity or spatial mismatch; this directly undermines interpretability of the group differences.

Authors: We agree that this represents a significant interpretative caveat. The TCGA samples are not spatially matched to the full tumor volumes segmented on imaging. We will add a new paragraph in the Discussion explicitly addressing intra-tumor heterogeneity, the potential for spatial mismatch between the RNA-seq aliquots and the radiomic regions of interest, and the implications for our conclusions. This will temper our claims and suggest directions for future research involving spatially resolved data. revision: yes
Referee: In the subtype identification and radiomics association results: with n=60, no independent validation cohort, and post-hoc feature comparison after clustering, the stability of C1/C2 and the reported radiomic associations lack external confirmation and are vulnerable to overfitting or sampling bias.

Authors: The modest sample size and absence of an external validation set are acknowledged limitations of this study, which relies on publicly available data. We will enhance the Methods section with additional details on clustering stability metrics (e.g., consensus matrix, PAC score) and expand the Discussion to emphasize the need for independent validation in future work. As an initial radiogenomic framework, the current findings are hypothesis-generating rather than confirmatory. revision: partial

Circularity Check

0 steps flagged

No circularity: purely empirical comparison on public datasets

full rationale

The paper applies standard unsupervised consensus clustering to TCGA RNA-seq data to define molecular subtypes C1/C2, performs differential expression and enrichment, then extracts PyRadiomics features from manually segmented TCIA images and compares them via Mann-Whitney U-test. No equations, no parameter fitting that defines the outcome, no self-citations invoked as uniqueness theorems or ansatzes, and no renaming of known results. The reported feature differences are direct statistical observations, not reductions to the clustering inputs by construction. The analysis is self-contained against external benchmarks.

Axiom & Free-Parameter Ledger

1 free parameters · 2 axioms · 0 invented entities

The claim rests on standard bioinformatics assumptions about clustering and radiomics without new entities or heavy parameter fitting beyond cluster count.

free parameters (1)

number of clusters
Set to 2 based on consensus clustering stability in the transcriptomic analysis

axioms (2)

domain assumption Unsupervised consensus clustering on transcriptomic profiles identifies biologically meaningful molecular subtypes
Invoked to define C1 and C2 from TCGA data
domain assumption Radiomic features from manually segmented tumor regions on imaging reflect underlying transcriptomic differences
Central premise linking imaging phenotypes to molecular subtypes

pith-pipeline@v0.9.0 · 5617 in / 1375 out tokens · 73495 ms · 2026-05-08T04:44:03.610119+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

34 extracted references

[1]

Ghazi, N., Saghravanian, N., Anvari, K., Saghafi Khadem, S., Hoseinzadeh, M., & Barzanouni, R. (2025). Evaluation of survival rate in patients with tongue squamous cell carcinoma: a retrospective single-center study. BMC Oral Health, 25(1), 658

2025
[2]

J., Punjabi, N

Damazo, B. J., Punjabi, N. A., Liu, Y. F., & Inman, J. C. (2024). Histopathologic predictors of recurrence and survival in early T stage oral tongue squamous cell carcinoma. Frontiers in oral health, 5, 1426709

2024
[3]

H., Ji, Y

Kwak, J. H., Ji, Y. B., Song, C. M., Lee, Y. J., Park, H. J., & Tae, K. (2025). Long-term oncologic outcomes and prognostic factors related to recurrences in pathologic Stage I/II early oral tongue cancer. Frontiers in surgery, 12, 1534274

2025
[4]

R., Kolegova, E

Patysheva, M. R., Kolegova, E. S., Khozyainova, A. A., Prostakishina, E. A., Korobeynikov, V. Y., Menyailo, M. E., ... & Denisov, E. V. (2024). Revealing molecular mechanisms of early-onset tongue cancer by spatial transcriptomics. Scientific Reports, 14(1), 26255

2024
[5]

& Zhang, Y

Lin, W., Huang, W., Mei, S., Lu, X., Zheng, J., Wang, H., ... & Zhang, Y. (2025). Transdifferentiation of tongue muscle cells into cancer-associated fibroblasts in response to tongue squamous cell carcinoma. Nature Communications, 16(1), 6753

2025
[6]

Liang, L., Li, Y., Ying, B., Huang, X., Liao, S., Yang, J., & Liao, G. (2023). Mutation-associated transcripts reconstruct the prognostic features of oral tongue squamous cell carcinoma. International Journal of Oral Science, 15(1), 1

2023
[7]

Astreidis, I., Kostidis, I., Malousi, A., Paraskevopoulos, K., Andreadis, D., Vahtsevanos, K., & Vizirianakis, I. (2026). Omics-Derived Prognostic Biomarkers in Tongue Squamous Cell Carcinoma: A Systematic Review with Risk -of-Bias Appraisal and Translation al Prioritization. Current Issues in Molecular Biology, 48(4), 389

2026
[8]

Lai, J., Yang, H., & Chen, J. (2024). Predicting radiotherapy efficacy and prognosis in tongue squamous cell carcinoma through an in-depth analysis of a radiosensitivity gene signature. Frontiers in Oncology, 14, 1334747

2024
[9]

Wang, D., He, X., Huang, C., Li, W., Li, H., Huang, C., & Hu, C. (2024). Magnetic resonance imaging-based radiomics and deep learning models for predicting lymph node metastasis of squamous cell carcinoma of the tongue. Oral Surgery, Oral Medicine, Oral Pathology and Oral Radiology, 138(1), 214-224

2024
[10]

& Liu, F

Liu, S., Zhang, A., Xiong, J., Su, X., Zhou, Y., Li, Y., ... & Liu, F. (2024). The application of radiomics machine learning models based on multimodal MRI with different sequence combinations in predicting cervical lymph node metastasis in oral tongue squamous cell carcinoma patients. Head & Neck, 46(3), 513-527

2024
[11]

Yao, Y., Jin, X., Peng, T., Song, P., Ye, Y., Song, L., ... & An, P. (2024). A novel nomogram for predicting overall survival in patients with tongue squamous cell carcinoma using clinical features and MRI radiomics data: a pilot study. World Journal of Surgical Oncology, 22(1), 227

2024
[12]

& Ansarin, M

Tagliabue, M., Ruju, F., Mossinelli, C., Gaeta, A., Raimondi, S., Volpe, S., ... & Ansarin, M. (2024). The prognostic role of MRI-based radiomics in tongue carcinoma: a multicentric validation study. La radiologia medica, 129(9), 1369-1381

2024
[14]

& Covello, R

Vidiri, A., Marzi, S., Piludu, F., Lucchese, S., Dolcetti, V., Polito, E., ... & Covello, R. (2023). Magnetic resonance imaging -based prediction models for tumor stage and cervical lymph node metastasis of tongue squamous cell carcinoma. Computational and Structural Biotechnology Journal, 21, 4277-4287

2023
[15]

& Liu, Y

Li, W., Li, Y., Wang, L., Yang, M., Iikubo, M., Huang, N., ... & Liu, Y. (2025). Evaluating fusion models for predicting occult lymph node metastasis in tongue squamous cell carcinoma. European Radiology, 35(9), 5228-5238

2025
[16]

Li, Y., Huang, N., Wang, L., Xiao, H., Chen, W., Xing, Y., ... & Li, W. (2026). An interpretable machine learning model using SHapley Additive exPlanations for preoperative cervical lymph node metastasis risk stratification in tongue squamous cell carcinom a: a multicenter study. BMC Oral Health, 26(1), 185

2026
[17]

ConsensusClusterPlus: a class discovery tool with confidence assessments and item tracking

Wilkerson, Matthew D., and D. Neil Hayes. "ConsensusClusterPlus: a class discovery tool with confidence assessments and item tracking." Bioinformatics 26.12 (2010): 1572-1573

2010
[18]

Liu, S., Wang, Z., Zhu, R., Wang, F., Cheng, Y., & Liu, Y. (2021). Three differential expression analysis methods for RNA sequencing: limma, EdgeR, DESeq2. JoVE (Journal of Visualized Experiments), (175), e62528

2021
[19]

A., Gao, Y., & Gerig, G

Yushkevich, P. A., Gao, Y., & Gerig, G. (2016, August). ITK -SNAP: An interactive tool for semi-automatic segmentation of multi -modality biomedical images. In 2016 38th annual international conference of the IEEE engineering in medicine and biology society (EMBC) (pp. 3342-3345). IEEE

2016
[20]

J., Fedorov, A., Parmar, C., Hosny, A., Aucoin, N., Narayan, V.,

Van Griethuysen, J. J., Fedorov, A., Parmar, C., Hosny, A., Aucoin, N., Narayan, V., ... & Aerts, H. J. (2017). Computational radiomics system to decode the radiographic phenotype. Cancer research, 77(21), e104-e107

2017
[21]

Ren, J., Yang, G., Song, Y., Zhang, C., & Yuan, Y. (2024). Machine learning -based MRI radiomics for assessing the level of tumor infiltrating lymphocytes in oral tongue squamous cell carcinoma: a pilot study. BMC Medical Imaging, 24(1), 33

2024
[22]

D., Mattavelli, D., Ravanelli, M., Poli, T.,

Corti, A., Lenoci, D., Corino, V. D., Mattavelli, D., Ravanelli, M., Poli, T., ... & Mainardi, L. (2025). Interplay between MRI radiomics and immune gene expression signatures in oral squamous cell carcinoma. Scientific Reports, 15(1), 12622

2025
[23]

& Mainardi, L

Corti, A., De Cecco, L., Cavalieri, S., Lenoci, D., Pistore, F., Calareso, G., ... & Mainardi, L. (2023). MRI -based radiomic prognostic signature for locally advanced oral cavity squamous cell carcinoma: development, testing and comparison with genomic prognostic signatures. Biomarker Research, 11(1), 69

2023
[24]

Näsiaho, J., Nissi, L., Ventelä, S., Irjala, H., & Salmi, M. (2026). Spatial single -cell analysis reveals tumor microenvironment signatures predictive of oral cavity cancer outcome. Cell Reports Medicine, 7(2)

2026
[25]

& Sun, S

Sun, L., Kang, X., Wang, C., Wang, R., Yang, G., Jiang, W., ... & Sun, S. (2023). Single -cell and spatial dissection of precancerous lesions underlying the initiation process of oral squamous cell carcinoma. Cell discovery, 9(1), 28

2023
[26]

Jin, W., Yu, Q., Yu, L., Zhou, T., Wang, X., Lu, W., ... & Ni, Y. (2025). HAS1high cancer associated fibroblasts located at the tumor invasion front zone promote oral squamous cell carcinoma invasion via ECM remodeling. Journal of Experimental & Clinical Cancer Research, 44(1), 238

2025
[27]

Y., Zhu, W

Zhang, J. Y., Zhu, W. W., Wang, M. Y., Zhai, R. D., Wang, Q., Shen, W. L., & Liu, L. K. (2021). Cancer-associated fibroblasts promote oral squamous cell carcinoma progression through LOX-mediated matrix stiffness. Journal of Translational Medicine, 19(1), 513

2021
[28]

Liu, X., Wang, D., Cheng, X., Ma, Y., Li, S., Deng, J., & Qiao, C. (2025). Fibroblast–immune crosstalk in oral squamous cell carcinoma: from tumor promotion to immune evasion. Frontiers in Immunology, 16, 1703277

2025
[29]

& Goto, A

Oikawa, Y., Umakoshi, M., Suzuki, K., Kudo-Asabe, Y., Miyabe, K., Koyama, K., ... & Goto, A. (2025). Prognostic significance of cancer-associated fibroblasts and tumor-associated macrophages in the tongue squamous cell carcinoma and their correlation with tumor budding. Oral Oncology, 165, 107295

2025
[30]

& Mainardi, L

Corti, A., Cavalieri, S., Calareso, G., Mattavelli, D., Ravanelli, M., Poli, T., ... & Mainardi, L. (2024). MRI radiomics in head and neck cancer from reproducibility to combined approaches. Scientific Reports, 14(1), 9451

2024
[31]

& Xiao, M

Song, H., Lou, C., Ma, J., Gong, Q., Tian, Z., You, Y., ... & Xiao, M. (2022). Single -cell transcriptome analysis reveals changes of tumor immune microenvironment in oral squamous cell carcinoma after chemotherapy. Frontiers in cell and developmental biology, 10, 914120

2022
[32]

& Marzi, S

Vidiri, A., Dolcetti, V., Mazzola, F., Lucchese, S., Laganaro, F., Piludu, F., ... & Marzi, S. (2025). MRI in Oral Tongue Squamous Cell Carcinoma: A Radiomic Approach in the Local Recurrence Evaluation. Current Oncology, 32(2), 116

2025
[33]

K., Hedou, J., Feyaerts, D., Ozawa, M

Einhaus, J., Gaudilliere, D. K., Hedou, J., Feyaerts, D., Ozawa, M. G., Sato, M., ... & Han, X. (2023). Spatial subsetting enables integrative modeling of oral squamous cell carcinoma multiplex imaging data. Iscience, 26(12)

2023
[34]

& Kislinger, T

Principe, S., Mejia-Guerrero, S., Ignatchenko, V., Sinha, A., Ignatchenko, A., Shi, W., ... & Kislinger, T. (2018). Proteomic analysis of cancer -associated fibroblasts reveals a paracrine role for MFAP5 in human oral tongue squamous cell carcinoma. Journal of proteome research, 17(6), 2045-2059

2018
[35]

& Huang, C

Tan, Y., Wang, Z., Xu, M., Li, B., Huang, Z., Qin, S., ... & Huang, C. (2023). Oral squamous cell carcinomas: state of the field and emerging directions. International journal of oral science, 15(1), 44

2023