Computed tomography radiomic analysis of paraspinal muscles in the prognosis of advanced head and neck cancers

Background & aims

Sarcopenia is a well-recognized risk factor for adverse outcomes in neoplastic diseases, and body composition assessment using computed tomography is a standard method for its evaluation. Radiomics, an automated and quantitative image-analysis approach that has demonstrated prognostic value in various clinical contexts, has not yet been applied to the assessment of axial musculature for outcome prediction in head and neck cancers. The primary aim of this study was to investigate whether radiomic analysis of the paravertebral muscles on computed tomography imaging could improve survival prediction in patients with locally advanced head and neck cancer.

Methods

We retrospectively included 71 patients with locally advanced head and neck cancer who received induction chemotherapy at our institution. Radiomic features were extracted following manual segmentation of the paravertebral muscles at the L1 level on computed tomography scan. Only features that were unaffected by the timing of contrast injection and demonstrated high intra-observer reproducibility were retained for analysis. Associations between these radiomic features and survival were assessed using univariate and multivariate Cox proportional hazards regression. Relationships with treatment toxicity and therapeutic response were evaluated using either Student's t-test or the Mann–Whitney test, as appropriate, and multivariate logistic regression.

Results

A total of 21 radiomic parameters were retained for analysis. In the multivariate analysis, none of these parameters were significantly associated with survival. However, the ability to maintain oral feeding at diagnosis and one histogram-based radiomic feature - the sum of Hounsfield unit values after discretization - emerged as the most promising predictors. After binarization of this histogram feature, both variables were significantly associated with survival, stratifying the cohort into four groups with distinct survival outcomes (p < 0.001). None of the radiomic parameters demonstrated a significant association with treatment-related toxicity in the multivariate analysis. Nevertheless, the CT subcutaneous fat index and the second-order radiomic feature GLRLM SRE exhibited a trend toward being risk factors for toxicity.

Conclusions

No parameter, including radiomic features, was able to statistically and independently demonstrate prognostic value for locally advanced head and neck cancers. However, a radiomic feature, the sum of Hounsfield unit value after discretization, when used in conjunction with Computed Tomography body composition parameters and clinico biological nutritional parameters could help predict survival.