Predicting Xerostomia Using Diffusion Weighted Imaging in Head and Neck Cancer

Abstract

Diffusion weighted imaging is an MRI technique that is sensitive to cell density, and has been shown to be predictive of response and outcome in head and neck cancer (HNC). Its role in xerostomia prediction in HNC has been limited. ADC maps have been shown to be correlated to cellularity and have the potential to be used to better understand the physiological meaning of MR derived radiomics. We hypothesized that combining salivary gland ADC with MR image features significantly improves xerostomia prediction compared to MR features alone. Baseline images and xerostomia (NCI CTCAE v4.0) at 3 month follow-up were prospectively collected in HNC patients treated at our institution with radiotherapy (RT) from 2015-2018. Moderate to severe xerostomia was defined as grade ≥ 2. Prior to treatment, T1-weighted images with a turbo spin echo sequence post-Gd and multiple-b (b=0, 100, 600, 1000s/mm2) diffusion weighted images were acquired on a 1.5T MRI scanner. Ipsilateral/contralateral parotid and submandibular glands (iPG, cPG, iSG, cSG) contoured on the CT images were propagated onto the MR images. MR images were normalized using relative intensity based on the mean subcutaneous fat at the level of the salivary glands. 429 MR radiomic features were extracted. Features were pre-selected by examining the predictor-outcome association by removing predictors based on the p-value of the correlation (p>0.05). A shrinkage regression analysis of pre-selected features was performed using least absolute shrinkage and selection operator (LASSO). To address the imbalance of patients with xerostomia, we applied a synthetic minority over-sampling technique to create a balanced set. ADC maps were determined using a voxel-by-voxel basis by evaluating the signal loss after subsequent diffusion sensitizing gradients using the exponential fit S(b)=S(0)e(-ADC×b). Prediction modeling was performed using generalized linear model with ten-fold cross validation. Model performance was evaluated using the area under the receiver operating characteristic curve (AUC). Standardized beta coefficients were determined (β). Ninety-two HNC patients were evaluated, with (n=25, 61±11yr) and without severe xerostomia (n=67, 59±11yr). The AUC/sensitivity/specificity for the models containing T1-MR features, mean PG ADC, and T1-MR+PG-ADC, were 0.70/0.65/0.67, 0.65/0.74/0.58, and 0.76/0.65/0.67. In the T1-MR+ADC model, cPG ADC (β=4.16, p=0.004) significantly contributed to the model and there was a trend towards significance for iPG ADC (β=-1.27, p=0.06). These results suggest that increased diffusivity (suggesting decreased cellularity) of the cPG prior to RT is related to increased risk of xerostomia. In a small cohort of HNC patients, combining T1 features with ADC improved the xerostomia prediction model. This work suggests that diffusivity of the parotid glands may influence RT-induced xerostomia. Future work is required in larger cohort to generalize these findings.