Radiomics in head and neck squamous cell carcinoma – a leap towards precision oncology

Abstract

Immunotherapy has revolutionized head and neck squamous cell carcinoma (HNSCC) treatment, with neoadjuvant chemoimmunotherapy showing promising pathological complete response rates (36–42%). Lin et al introduce a radiomics-clinical nomogram using MRI-derived intratumoral and peritumoral features to predict pCR, addressing a critical clinical gap. Their model, emphasizing the peritumoral region (within 3 mm), achieved high predictive accuracy area under curve (AUC) >0.8. While the multicenter design enhances generalizability, standardizing imaging protocols remains a challenge. Integrating radiomics with the Neck Imaging Reporting and Data System could refine post-treatment assessment. This study advances precision oncology in HNSCC, offering a non-invasive tool for personalized treatment strategies. Future directions include artificial intelligence-driven radiomics and radiogenomics to enhance treatment response prediction and patient selection.

The treatment of head and neck squamous cell carcinoma (HNSCC) has undergone remarkable advances over the past decade, with immunotherapy emerging as a cornerstone.¹ The KEYNOTE-048 trial established pembrolizumab, alone or with chemotherapy, as a first-line treatment for recurrent/metastatic disease, prompting interest in its neoadjuvant application.² Early studies of neoadjuvant chemoimmunotherapy (NACI) have shown promising pathologic complete response (pCR) rates of 36–42%,³ raising the possibility of treatment de-escalation and improved functional outcomes in locally advanced disease. Accurate interpretation of post-treatment imaging and early recognition of complications as well as recurrences are essential for optimal patient management.²

In this context, Lin et al present a compelling study in the Journal for ImmunoTherapy of Cancer, introducing a radiomics-clinical nomogram that leverages MRI-derived intratumoral and peritumoral features to predict pCR to NACI. Their work addresses a critical unmet need: while some patients respond well and may benefit from less aggressive therapy, others show limited response, risking unnecessary complications and delays in definitive treatment. Given the limited predictive value of existing biomarkers such as the combined positive score (CPS), this radiomics-based approach offers a promising alternative.⁴

The retrospective study offers valuable insights into the tumor microenvironment (TME). While tissue-based methods like multiplex immunofluorescence enable direct identification of specific immune cell populations within the TME,⁵ radiomics provides an indirect, non-invasive alternative. By extracting high-throughput quantitative imaging features—such as heterogeneity, texture, edge sharpness, and intensity gradients—radiomics captures spatial patterns that may reflect underlying biological processes, including immune infiltration, stromal remodeling, and angiogenesis. Although it cannot identify specific immune cell types, these imaging-derived patterns can act as surrogate markers of TME composition.⁶ Lin et al emphasize the significance of analyzing both intratumoral and peritumoral regions, particularly the immediate peritumoral zone (within 3 mm), which is increasingly recognized for its role in immune modulation. Their integration of radiomics features with CPS and clinical parameters (eg, age, gender, smoking history, histological type, clinical stage, and treatment cycle) into a predictive nomogram yielded robust performance, with AUC values exceeding 0.8 across cohorts. This supports the ability of radiomics to capture tumor and TME heterogeneity—a finding consistent with prior studies in gastric and esophageal cancers, where peritumoral radiomics improved predictive accuracy.^{7 8}

The prognostic value of peritumoral regions has similarly been demonstrated in breast and lung cancer, particularly in predicting response to neoadjuvant therapy.9,12 Lin et al’s use of combined intratumoral and peritumoral features parallels findings from gastric cancer studies that showed enhanced performance over intratumoral features alone.⁸ While CT remains the standard imaging modality in HNSCC staging, MRI was likely selected here for its superior soft tissue contrast and richer depiction of tumor microstructure, especially within the anatomically complex head and neck region.¹³ Nevertheless, the limited routine use of MRI in standard care presents a key barrier to clinical translation. Algorithms trained exclusively on MRI may face challenges in widespread adoption unless they are validated on CT-based data sets or incorporated into workflows that routinely use MRI.

The study’s multicenter design enhances robustness, yet variability in imaging protocols across institutions underscores the need for standardization to ensure reproducibility and facilitate clinical adoption. Additionally, the absence of positron emission tomography—a modality that provides complementary metabolic information—represents an opportunity for future integration to further strengthen predictive models.¹⁴ The application of standardized lexicons, such as the Neck Imaging Reporting and Data System (NI-RADS), has already shown value in assessing disease status post-treatment, with high predictive accuracy in both primary and nodal sites. Future research should explore how radiomics can be integrated with NI-RADS to refine imaging-based risk stratification in HNSCC.¹⁵

Looking ahead, this work lays the foundation for more personalized treatment approaches. Patients predicted to achieve pCR could be considered for de-escalation trials, potentially sparing them from treatment-related morbidity, while non-responders might be redirected toward alternative or intensified therapies. The non-invasive and repeatable nature of MRI-based radiomics is particularly suitable for longitudinal monitoring, adding further value in precision oncology.

The evolving role of artificial intelligence (AI) and deep learning in radiomics is promising. AI-driven radiomics models have shown utility in risk stratification, screening, diagnosis, treatment response prediction, and automated radiology reporting, minimizing report turnaround time in high-volume centers. The integration of AI-based image processing into radiomics workflows holds promise for improving model generalizability and clinical adoption.¹⁶ Moreover, integrating radiomics with emerging biomarkers, such as liquid biopsies and genomic profiling, could enhance predictive power. Radiogenomic analyses, for example, have linked imaging features with immune-related processes like lymphocyte infiltration, offering a biologically grounded rationale for integrative, multimodal approaches.⁷

In conclusion, Lin et al present a significant contribution to the advancement of radiomics in HNSCC, illustrating how advanced imaging can inform individualized therapy. While prospective validation remains essential, this study represents a critical step toward integrating imaging-based biomarkers into routine clinical decision-making in head and neck cancer.