Acute deep neck infections remain clinically challenging and often require multidisciplinary management and invasive procedures. Imaging plays a central role in diagnosis and treatment planning, particularly in assessing the extent of infection and identifying drainable abscesses. Magnetic resonance imaging offers high diagnostic accuracy in aute settings and enables detailed evaluation of soft tissue changes. Retropharyngeal oedema is a distinct reactive oedema pattern visible on fat-suppressed T2-weighted MRI and has previously been associated with intensive care unit admission. While earlier work treated retropharyngeal oedema as a binary finding, the potential clinical value of its volumetric extent has not been fully explored. An analysis of patients undergoing emergency neck MRI examined whether retropharyngeal oedema volume correlates with disease severity and whether deep learning could support automated segmentation and quantification.
Must Read: MR-Linac ADC Reproducibility in Head and Neck Cancer
Study Design and Quantitative Assessment
Patient data were retrospectively reviewed from a single academic tertiary care referral centre, including individuals with clinically confirmed acute neck infections who underwent emergency MRI between April 2013 and August 2021. The cohort comprised 479 patients, of whom 422 were adults and 57 were children. Clinical correlations were performed exclusively in adults to avoid bias related to anatomical scale differences.
Retropharyngeal oedema was manually segmented slice by slice on T2-weighted fat-suppressed Dixon images, followed by automated volume calculation in cubic millimetres. Clear anatomical landmarks defined the segmentation boundaries. Intraobserver agreement was excellent, with an intraclass correlation coefficient of 0.992, and interobserver agreement was good, with a coefficient of 0.827.
Among adult patients, 51% were retropharyngeal oedema positive. In this subgroup, the mean oedema volume was 5,028 mm³ and ranged from 234 to 31,490 mm³. Adults requiring intensive care had significantly higher oedema volumes than those not admitted to intensive care. Oedema volume also showed positive correlations with length of hospital stay, C-reactive protein levels and maximal abscess diameter. In contrast, average oedema volume did not differ significantly between adults with and without abscess.
In multivariable logistic regression predicting intensive care admission, maximal abscess diameter and C-reactive protein were statistically significant predictors, while oedema volume showed a trend towards significance. In classification analysis, oedema volume achieved an area under the receiver operating characteristic curve of 0.775, compared with 0.714 for binary oedema classification. A DeLong test demonstrated significantly better discrimination for volumetric assessment. The optimal cutoff identified for oedema volume was 1,389 mm³, with high sensitivity and moderate specificity.
Deep Learning Classification and Segmentation
To enable automated quantification, convolutional neural networks inspired by the U-Net architecture were developed. A slice-wise classification network first identified candidate regions across coronal, sagittal and axial views. Five-fold cross-validation was used, and training and testing were repeated to ensure separation between classification and segmentation tasks.
Classification performance was high across all views. Mean area under the receiver operating characteristic curve reached 0.982 for sagittal slices, 0.974 for coronal slices and 0.971 for axial slices. Mean accuracy ranged from 89% to 95%. Sagittal slices demonstrated significantly higher discrimination than coronal slices, while no significant differences were observed between sagittal and axial or coronal and axial views.
Segmentation was performed using modified U-Net models operating on 64 × 64 pixel inputs selected from predicted three-dimensional bounding boxes. This approach addressed the scarcity of positive voxels, which represented between 0.0023% and 0.31% of total image volume. Axial segmentation achieved the highest mean Dice similarity coefficient at 0.534, followed by the 2.5-dimensional approach at 0.518. Coronal and sagittal models showed lower mean Dice values.
Dice similarity coefficients demonstrated a low but statistically significant correlation with true oedema volume. All cases with a Dice value of zero in axial segmentation had true oedema volumes below 330 mm³. For patients with volumes of at least 3,000 mm³, axial segmentation achieved a mean Dice coefficient of 0.596. Predicted and true oedema volumes showed a strong correlation, with a Pearson coefficient of 0.775. The mean absolute error was 1,810 mm³ and the mean relative error was 48.9%.
Clinical Implications and Implementation Considerations
Volumetric quantification provided more granular information than binary classification and was associated with clinically relevant outcomes in adult patients. Larger oedema volumes correlated with intensive care admission, inflammatory markers and longer hospital stay. The identified volume threshold of 1,389 mm³ may support risk stratification, although further validation in larger and prospective cohorts is required before clinical integration.
Automated segmentation may reduce the time burden associated with manual delineation in emergency radiology. The bounding box approach enabled efficient localisation of oedema regions while limiting computational load. Training times for classification and segmentation models ranged from 3.6 to 129.5 minutes, depending on the configuration.
Limitations included retrospective single-centre data and the absence of an external test set. Only cross-validation was available to assess generalisability. Future work may involve multi-institutional datasets, advanced augmentation strategies or three-dimensional convolutional approaches to enhance segmentation accuracy and robustness across scanners and acquisition protocols.
Retropharyngeal oedema represents reactive oedema rather than a drainable collection, and its underlying pathophysiology in acute neck infections remains unclear. Nonetheless, consistent anatomical definitions and strong interobserver agreement support its reliability as an imaging feature.
Volumetric assessment of retropharyngeal oedema on MRI was associated with markers of severe illness and outperformed binary classification in predicting intensive care admission. A deep learning framework combining slice-wise classification and targeted segmentation enabled automated localisation and moderate-accuracy volume estimation. With further validation and technical refinement, retropharyngeal oedema volume may serve as a quantitative imaging biomarker supporting risk stratification in acute deep neck infections.
Source: European Radiology Experimental
Image Credit: iStock
