RPS 108 - Integration of advanced imaging techniques and AI in the head and neck

Author Block: H. Zhang, Y. Meng; Xuzhou/CN
Purpose: Compare image quality of image reconstructed using DLIR and IR algorithms for head and neck dual-energy CT angiography (DECTA)
Methods or Background: This prospective study comprised patients with head and neck DECTA. Images reconstructed by four algorithms (120-kVp-like with ASIR-V40%,50 keV with ASIR-V40%,50 keV with DLIR-M,50 keV with DLIR-H) were compared. CT attenuation, image noise, SNR, CNR, edge-rise distance (ERD) and edge-rise slope (ERS) were calculated. Subjective image quality scores were evaluated.
Results or Findings: CT attenuation of vessels in 120kVp-like images were lower than 3 sets of 50 keV images with significant difference. In 50 keV images, both sternocleidomastoid muscle and white matter had a minimum noise in DLIR-H, and a maximum in ASIR-V40% group with significant difference. SNR and CNR in 50 keV images of all vessels had the same results: highest in DLIR-H group and lowest in ASIR-V40% group with significant differences. The mean value of ERD showed no significant difference among four groups. While 120kVp-like images had the lowest ERS, which showed statistically significant difference with the other groups. In terms of overall image quality, sharpness, and artifacts, the scores of DLIR-M and DLIR-H at 50 keV were not statistically different, and were higher than ASIR-V40% at 50 keV images, and higher than ASIR-V40% at 120 kVp-like. The scores of DLIR-H at 50 keV were highest in terms of noise and average scores.
Conclusion: DLIR is a potential solution for DECTA reconstruction since it can greatly reduce image noise, improving image quality of head and neck DECTA at 50 keV.
Limitations: First, our population was relatively small. Secondly, our study only compared 40%ASIR-V and DLIR, without including higher levels of ASIR-V for comparison. Finally, patients with BMI out of the normal range were excluded.
Funding for this study: We acknowledge financial support from the Jiangsu Traditional Chinese Medicine Science and Technology Development Plan Project (MS2021100), and the Key Research and Development Program of Xuzhou Science and Technology Bureau (KC20159).
Has your study been approved by an ethics committee? Yes
Ethics committee - additional information: Ethical approval was obtained from the Affiliated Hospital of Xuzhou Medical University, (approval number XYFY2024-KL456-01)

Author Block: D. Giardino, A-I. Nica, P. Thoenissen, T. J. Vogl, I. Yel, R. Sader, C. Booz; Frankfurt/DE
Purpose: To investigate the relationship between primary tumor volume, the spatial distribution of cervical lymphnode metastases, and TNM staging in patients with oral and pharyngeal squamous cell carcinoma (HNSCC) using high-resolution MRI.
Methods or Background: This retrospective analysis evaluated 116 predominantly male (62.9%) patients (age 67.7 ± 11.5 years) with histologically confirmed HNSCC who underwent surgical resection with neck dissection or definitive chemoradiation. MRI-based volumetry and center-to-center distance measurements between primary tumor and cervical lymphnode metastases were performed using 3D postprocessing software.
Results or Findings: Tumor-to-lymphnode distances ranged from 11.7 to 117.3 mm (mean: 47.6 ± 19.4 mm). Tumor volume varied from 0.5 to 87.2 cm³ (median: 13.3 ± 13.2 cm³), and was higher in male patients. Significant correlation was observed between tumor volume and lymphnode distance and pT-stage (p<0.0001). The correlation between tumor volume and the distance to metastatic lymphnodes was stronger (Spearman’s r=0.4541, p<0.0001) than the correlation between the pT-stage (TNM classification) and nodal distance (Spearman’s r=0.2682, p=0.0036).
Conclusion: MRI-based assessment revealed that tumor volume is a stronger predictor of the spatial extent of lymph node metastases than T-stage alone. Larger tumors were associated with greater distances to metastatic lymph nodes. These findings highlight the value of quantitative 3D MRI-based volume and distance analyses in potentially improving surgical and radiotherapeutic planning.
Limitations: Limitations of the study include its retrospective nature.
Funding for this study: No funding was obtained for this study.
Has your study been approved by an ethics committee? Yes
Ethics committee - additional information: This study received approval from the Ethics Committee.

Author Block: A. h. Shah, S. Sen, A. Gehani, A. Chandra, P. Ghosh, S. Mukhopadhyay, A. Chatterjee, J. Khoda, A. Patra; Kolkata/IN
Purpose: Node-RADS was originally designed for staging untreated lymph nodes. Its performance in the post-treatment neck, where fibrosis and necrosis mimic recurrence, is unknown. We confirmed Node-RADS in treated head and neck cancer patients, with subsite-specific analysis across CT and MRI cohorts.
Methods or Background: We retrospectively reviewed 602 patients with head and neck squamous cell carcinoma treated between 2012 and 2024, yielding 1,098 post-treatment nodes. Imaging modality was decided by subsite: oral cavity and larynx underwent CT, while oropharynx and nasopharynx were assessed with MRI (no dual-modality cases). Two head and neck radiologists independently assigned Node-RADS categories (1-5). Reference standards included histopathology, MDT consensus, or ≥12-month imaging follow-up. Diagnostic performance and predictive values were calculated overall and by subsite; inter-observer agreement was assessed (κ).
Results or Findings: Node-RADS categories correlated strongly with recurrence risk (p<0.001). At a cutoff of Node-RADS ≥4, sensitivity was 81% and specificity 86%, with overall accuracy 84%. NPV exceeded 92% for Node-RADS 1-2, while Node-RADS 5 achieved PPV 89%. Subsite analysis showed comparable performance across CT-evaluated subsites (oral cavity, larynx) and MRI-evaluated subsites (oropharynx, nasopharynx). Interobserver agreement was substantial (κ = 0.74), aligning with prior untreated Node-RADS studies (κ ~0.6–0.7). Clinical concordance was high: 71% of Node-RADS 4–5 prompted MDT-directed biopsy or salvage, while most Node-RADS 1–2 remained under surveillance.
Conclusion: This is the first validation of Node-RADS in the treated neck. Structured nodal reporting retains strong predictive value post-therapy, with high reproducibility and clear MDT alignment. Node-RADS offers a practical lexicon for post-treatment surveillance, bridging radiology and clinical decision-making.
Limitations: Single-center retrospective design
Funding for this study: None
Has your study been approved by an ethics committee? Not applicable
Ethics committee - additional information:

Author Block: H. Steinberg¹, N. van Landeghem¹, D. Van Landeghem¹, B. M. Schaarschmidt¹, M. Forsting¹, I. Wanke², C. Deuschl¹, Y. Li ¹; ¹Essen/DE, ²Zürich/DE
Purpose: Head and neck paragangliomas are rare, hypervascular tumors for which preoperative embolization is often required to mitigate intraoperative blood loss.
This study compares intraoperative and histopathological outcomes of paragangliomas following preoperative embolization using particles, Onyx, or Glubran.
Methods or Background: Retrospective analysis of patients undergoing preoperative embolization for paragangliomas at a tertiary care center from 2010 to 2024. Outcomes included post-embolization angiographic devascularization, intraoperative bleeding, histopathological necrosis, inflammatory response, and resection margins.
Results or Findings: Twenty-two patients (15 female; mean age 52.6 years) were included. The mean interval between embolization and surgery was 5.1 days. Liquid embolic agents (Onyx: 54.5%, Glubran: 22.7%) were used in 77.3% of cases, with particles in the rest. Liquid embolics demonstrated superior tumor devascularization (median 90% vs. 50%, p=0.05) and necrosis (20% vs. 0%, p=0.039) compared to particles. However, no significant differences were observed in intraoperative blood loss (hemoglobin loss 1.4 g/dL for liquid embolics vs. 0.5 g/dL, p=0.676), procedure duration (134 vs. 96 minutes, p=0.085), or peritumoral inflammation. Complete resection (R0) was achieved in 81.25% of liquid embolic cases versus 100% with particles. Onyx and Glubran showed comparable efficacy.
Conclusion: Liquid embolic agents (Onyx/Glubran) achieved greater tumor devascularization and necrosis than particles, though surgical outcomes (blood loss, complete resection rates) were comparable. Agent selection may depend on procedural goals, balancing angiographic efficacy against surgical feasibility.
Limitations: The limitations of this study are its retrospective and single-center study design.
Funding for this study: No funding was received for this study.
Has your study been approved by an ethics committee? Yes
Ethics committee - additional information: The study was approved by the ethics committee of the University of Duisburg-Essen.

Author Block: F. Yang, M. Lin, H. Zhang; Beijing/CN
Purpose: To evaluate the value of intravoxel incoherent motion (IVIM) and arterial spin labeling (ASL)-based habitat analysis of intra- and peri-tumoral regions for predicting overall survival (OS) and progression-free survival (PFS) in NPC.
Methods or Background: 106 patients were prospectively included. Primary tumors were delineated on T2-weighted imaging, and three peritumoral regions (3, 5, and 10 mm) were automatically dilated and manually corrected according to the clinical tumor volume outlining criteria. The pure diffusion coefficient (D) and mean kurtosis (MK) maps of IVIM, and blood flow (BF) map of ASL were used for habitat analysis. Volume fraction and histogram parameters (Mean, Kurtosis and Skewness) were extracted for each subregion. Univariate and multivariate Cox analyses were used to construct intra- and peri-tumoral, Clinic, and combined models, which were assessed by the C-index. Nomogram, calibration curves, and Kaplan-Meier curves were also plotted.
Results or Findings: The intra-tumoral subregion 4 was characterized by high D and low BF and MK value, representing the radiotherapy-resistant region and patients with treatment failure had higher proportion of subregion 4. The Intra-score independently predicted OS (hazard radio [HR]: 1.023, P = 0.002) and PFS (HR: 1.028, P = 0.001). Combined clinical factors with intra- and peri-tumoral habitat features showed the highest performance (C-index: 0.780 [OS] and 0.721 [PFS]) and outperformed the Clinic model (C-index: 0.691 [OS] and 0.666 [PFS], P ≤ 0.03). The post-hoc subgroup analysis confirmed robustness.
Conclusion: Peri-tumoral regions provide valuable prognostic information. Combining IVIM-ASL-based habitat analysis offers a non-invasive approach to predicting treatment outcomes.
Limitations: Small sample size and lack of histopathological confirmation information were the limitations.
Funding for this study: None
Has your study been approved by an ethics committee? Yes
Ethics committee - additional information: Chinese Academy of Medical Science

Author Block: A. Sajust De Bergues De Escalup, A. Lecler, E. O'Shaughnessy; Paris/FR
Purpose: Orbital MRI is essential for assessing optic nerve disorders but is limited by motion artifacts, low signal-to-noise ratio (SNR), and susceptibility effects. Deep learning–based reconstruction improves image quality and shortens acquisition time. Canon’s Advanced intelligent Clear-IQ Engine (AiCE) reduces noise while preserving detail. This study evaluated AiCE’s impact on image quality and diagnostic confidence in coronal T2- and post-contrast T1-weighted orbital MRI.
Methods or Background: This retrospective single-center study included 72 orbital MRI exams on a 3T MRI. Coronal T2-WI (n=71) and post-contrast T1-WI (n=25) were reconstructed with and without AiCE. Two blinded radiologists separately reviewed series, assessing optic nerve hyperintensity, atrophy, and qualitative features (optic nerve sharpness, brain sharpness, overall quality) on 5-point scales. Motion artifacts were rated separately. SNR and contrast-to-noise ratio (CNR) were measured using. Wilcoxon and McNemar tests with Bonferroni correction were used. Agreement was assessed with Cohen’s κ and ICC.
Results or Findings: In coronal T2-WI, mean SNR rose from 25.94 ± 16.94 (non-AiCE) to 49.04 ± 26.01 (AiCE), and CNR from 14.29 ± 15.87 to 25.57 ± 22.48 (p < 0.001). In post-contrast T1-WI, SNR and CNR were unchanged (p = 1; p = 0.72). AiCE significantly improved brain and optic nerve sharpness and overall quality (p ≤ 0.001). Detection of optic nerve hyperintensity and atrophy was unchanged. Inter-reader κ for diagnostic features ranged 0.795–0.966; intra-reader κ for diagnostic features was 0.87–1; weighted κ for qualitative metrics was 0.209–0.803. ICCs ranged 0.234–0.624.
Conclusion: AiCE improved orbital MRI quality, especially in coronal T2-WI, without compromising diagnostic feature detection. These gains support its use for confident optic nerve assessment in routine practice.
Limitations: Quantitative gains were variable, with higher intra- than inter-reader agreement. Further studies should optimize its use across centers.
Funding for this study: No funding was provided for this study.
Has your study been approved by an ethics committee? Not applicable
Ethics committee - additional information:

Author Block: R. O. ÖNDER¹, T. Bekçi¹, A. Tosun²; ¹Giresun/TR, ²Trabzon/TR
Purpose: To evaluate the diagnostic performance of conventional radiological findings (CT and DWI) and radiomic features derived from ADC maps in the diagnosis of cholesteatoma, and to assess the added value of their artificial intelligence-based combination.
Methods or Background: In this retrospective study, 51 patients with suspected cholesteatoma (January 2020–August 2024) were analyzed. Non-contrast temporal bone CT and DWI were assessed, and radiomic features were extracted from ADC maps. Reproducibility was tested with intraclass correlation coefficients (ICC ≥0.75). Machine learning (ML) models, including a three-layer artificial neural network, were trained using radiology, radiomics, and combined datasets. An independent validation group (n=15) was used to confirm generalizability.
Results or Findings: Of 51 lesions, 23 (45.1%) were histopathologically confirmed as cholesteatoma. Cholesteatomatous lesions had higher CT HU values (p=0.004), more frequent diffusion restriction (p<0.001), and lower ADC values (p<0.001). Radiomic analysis identified 20 significant features. The combined ML model achieved the best performance (accuracy 98.0%, AUC 0.982), which remained robust in the validation cohort (accuracy 93.3%, AUC 0.951).
Conclusion: Artificial intelligence and ADC radiomics provide complementary diagnostic value to conventional radiology in the detection of cholesteatoma. The artificial intelligence-assisted integration of radiomics and radiology achieves superior performance, thereby supporting its potential role in non-invasive diagnosis and the clinical decision-making process.
Limitations: Single-center, retrospective design with a limited sample size; external validation is required.
Funding for this study: No funding was received for this study.
Has your study been approved by an ethics committee? Yes
Ethics committee - additional information: This retrospective study was approved by the Scientific Research Ethics Committee of Giresun Training and Research Hospital on 4 September 2024 (decision no. 17).

Author Block: H. K. Zhang¹, F. Chen¹, Y. Liao², X. Liu²; ¹Haikou/CN, ²Guangzhou/CN
Purpose: The Ki-67 status is often correlated with the clinical course of cancer. This study aimed to investigate the feasibility of predicting Ki-67 status in TSCC using histogram features of spectral images acquired with dual-layer spectral-detector CT(DLCT).
Methods or Background: A retrospective cohort of 69 patients with TSCC (50 Ki-67-positive and 19 Ki-67-negative, with a threshold of 20%) who underwent surgery with preoperative DLCT scans was collected. Conventional images and spectral images, including Virtual Monoenergetic Images (VMIs) at 40KeV, 70KeV, 100KeV, and 130KeV, iodine density maps, effective atomic number maps, virtual non-contrast (VNC) images, were generated based on venous-phase CT data. Tumor regions of interest (ROIs) were manually delineated on the VMI 40KeV images and then copied and pasted onto other images. A total of 12 histogram parameters were extracted from each image. Feature selection was performed using Spearman correlation analysis (threshold = 0.9) and backward stepwise regression. Logistic regression was used to construct predictive models. Model performance was assessed with receiver operating characteristic (ROC) analysis, and compared with the DeLong's test.
Results or Findings: In the discrimination of Ki-67 status based on individual features, the energy from VNC obtained the highest area under the curve (AUC) of 0.727 (95% CI: 0.578-0.876), followed by the energy from VIMs 70keV with an AUC of 0.719 (0.571-0.867). The model combining valuable features from all spectral images achieved the highest AUC of 0.863(0.765-0.962), which was significantly outperformed than the model corporating valuable features from conventional images with an AUC of 0.663(0.505-0.822), and a P<0.01.
Conclusion: The histogram features of spectral images exhibit promising performance in preoperatively predicting Ki-67 status in patients with TSCC.
Limitations: Retrospective, single-center study with a small sample size
Funding for this study: Not applicable
Has your study been approved by an ethics committee? Yes
Ethics committee - additional information: The Ethics Committee granted a waiver of informed consent for this retrospective study.

Author Block: E. Frodl, L. Golla, L. Gantner, M. Harth, J. Dietz, T. Vogl, P. Thoenissen, A. M. Bucher; Frankfurt/DE
Purpose: To identify independent body composition predictors of overall survival in head and neck squamous cell carcinoma patients using automated deep learning analysis of routine baseline CT at L3 and abdominal cavity levels.
Methods or Background: This retrospective study analysed 181 HNSCC patients (mean age 63.2±12.4 years; 50.8% female) using deep learning-based automated body composition analysis from baseline venous-phase contrast-enhanced CT (91.7%, 5mm slice thickness). The algorithm quantified tissue volumes and densities including intermuscular adipose tissue infiltration (IMAT), subcutaneous adipose tissue (SAT), muscle, and bone. Additionally, muscle-to-bone ratio (MBR) and IMAT/TAT ratio were calculated. Cox regression with multivariate adjustment for age, gender, T-stage, N-stage, and UICC stage identified independent predictors.
Results or Findings: Patients presented with advanced disease (UICC IVA: 31.5%, III: 22.6%, I: 21.6%) with 41.4% nodal involvement. Multiple body composition metrics showed independent prognostic value in multivariate analysis. IMAT/TAT ratios remained significant predictors at both abdominal cavity (volume: HR=1.490, p=0.0008, 95%CI: 1.180-1.882; attenuation: HR=1.465, p=0.0015, 95%CI: 1.158-1.854) and L3 levels (volume: HR=1.458, p=0.0011, 95%CI: 1.163-1.827; attenuation: HR=1.457, p=0.0007, 95%CI: 1.171-1.812). SAT attenuation independently predicted survival at both levels (abdominal: HR=1.469, p=0.0017, 95%CI: 1.156-1.867; L3: HR=1.486, p=0.0005, 95%CI: 1.189-1.857). Uniquely, muscle-bone volume ratio at abdominal cavity showed protective effects (HR=0.741, p=0.0477, 95%CI: 0.550-0.997), unavailable at single-slice L3. All metrics maintained significance alongside UICC stage (HR=2.435, p=0.0015) in multivariate models.
Conclusion: Deep learning-based body composition analysis identifies multiple independent predictors of overall survival in HNSCC. IMAT infiltration and SAT attenuation maintain prognostic value across both anatomical levels. The protective muscle-bone volume ratio, unique to abdominal cavity assessment, supports volumetric analysis. These automatically derived biomarkers enhance risk stratification beyond traditional staging without additional radiation exposure.
Limitations: The limitations of the study are the single-centre retrospective design and pending external validation.
Funding for this study: Funding was provided by the German Federal Ministry of Education and Research through the RACOON project (reference number 01KX2021).
Has your study been approved by an ethics committee? Yes
Ethics committee - additional information: The study was approved by University Medicine Frankfurt (UCT-9-2023).

Author Block: F. Albisinni, C. Carbone, M. Ravanelli, D. Farina, G. Corciulo; Brescia/IT
Purpose: This study aimed to evaluate the time-saving potential of deep learning (DL)-based reconstructions in T2-weighted head and neck MRI, while assessing their impact on image quality.
Methods or Background: Fifty-two patients underwent three TSE T2 sequences: (A) standard acquisition without DL (3 signal averages, 2'35"), (B) DL with intermediate strength (2 averages, 1'25"), and (C) DL with maximal strength (1 average, 43"). Images were blindly assessed at two anatomical levels (nasopharynx and oral cavity) by three radiologists of varying experience. Image quality was rated on a 3-point Likert scale across four categories: overall quality, artifacts, edge sharpness, and noise. Readers also attempted to identify the sequence type. Inter-rater agreement and image quality comparisons were analyzed statistically.
Results or Findings: A total of 636 images were evaluated. Acquisition time was reduced by 45% in sequence B and by over 70% in sequence C compared to standard imaging. Despite the substantial time savings, DL-based sequences maintained comparable image quality: overall quality scores for sequences A, B, and C were 2.63, 2.52, and 2.52, respectively (p = 0.055). Differences in artifacts and noise were not statistically significant. A minor reduction in edge sharpness was observed between sequences A and B (p = 0.02). Inter-rater agreement remained low, and the ability to identify the sequence type was limited (44% accuracy).
Conclusion: Deep learning-based reconstructions enable significant reductions in scan time—up to 70%—in head and neck MRI without compromising diagnostic image quality. These findings highlight DL’s potential to support faster, more efficient imaging protocols in clinical practice.
Limitations: This study is limited by low inter-rater agreement, evaluation restricted to T2-weighted images at two anatomical levels, and the absence of diagnostic accuracy analysis.
Funding for this study: This study received no external funding.
Has your study been approved by an ethics committee? Not applicable
Ethics committee - additional information: