Machine Learning Methods Uncover Radiomorphologic Dose Patterns in Salivary Glands that Predict Xerostomia in Patients with Head and Neck Cancer

Abstract

Purpose: Patients with head-and-neck cancer (HNC) may experience xerostomia after radiation therapy (RT), which leads to compromised quality of life. The purpose of this study is to explore how the spatial pattern of radiation dose (radiomorphology) in the major salivary glands influences xerostomia in patients with HNC.

Methods and materials: A data-driven approach using spatially explicit dosimetric predictors, voxel dose (ie, actual radiation dose in voxels in parotid glands [PG] and submandibular glands [SMG]) was used to predict whether patients would develop xerostomia 3 months after RT. Using planned radiation dose data and other nondose covariates including baseline xerostomia grade of 427 patients with HNC in our database, the machine learning methods were used to investigate the influence of dose patterns across subvolumes in PG and SMG on xerostomia.

Results: Of the 3 supervised learning methods studied, ridge logistic regression yielded the best predictive performance. Ridge logistic regression was also preferred to evaluate the influence pattern of highly correlated dose on xerostomia, which showed a discriminative pattern of influence of doses in the PG and SMG on xerostomia. Moreover, the superior-anterior portion of the contralateral PG and medial portion of the ipsilateral PG were determined to be the most influential regions regarding dose effect on xerostomia. The area under the receiver operating characteristic curve from a 10-fold cross-validation was 0.70 ± 0.04.

Conclusions: Radiomorphology, combined with machine learning methods, is able to suggest patterns of dose in PG and SMG that are the most influential on xerostomia. The influence pattern identified by this data-driven approach and machine learning methods may help improve RT treatment planning and reduce xerostomia after treatment.

Figure 1

Distribution of radiation dose in parotid and submandibular glands across the patient cohort. (A) Mean voxel dose, (B) standard deviation of voxel dose, and (C) mean dose of patients who developed xerostomia, minus the mean dose of patients who did not.

Figure 2

Flowchart of key steps for this analysis. Abbreviations: ROI = region of interest.

Figure 3

Distribution of xerostomia grade at baseline and 3 months after radiation therapy.

Figure 4

Voxel importance patterns learned from the 3 machine learning algorithms where the color corresponds to the relative importance of each voxel.

Figure 5

(A) Voxel importance pattern from ridge logistic regression and (B) different visualization of the same voxel importance result where voxel importance values that are 1 standard deviation (SD) away from the mean were saturated to increase the resolution of voxel importance closer to the mean value of the voxel importance. The saturated plot was created as the set of voxel importance values that are greater than 1 SD of the mean voxel importance values to be 1 SD of the voxel importance values plus the mean value. As a result, the voxels of which the voxel importance value is greater than 1 SD of the mean voxel importance values are shown in red. Similarly, voxels of which the voxel importance value is greater than 1 SD of the mean voxel importance values are shown in blue. Voxel importance values within 1 SD of the mean value is color coded from blue to red.

Figure 6

Two-dimensional cross-sectional images from computed tomography scans of the reference patient, displaying the spatial location of the influential area and colored distribution of voxel-based dose feature importance. (A) Axial view of voxel importance pattern, (B) sagittal view of the voxel importance pattern on the contralateral side, (C) coronal view of the voxel importance pattern, and (D) sagittal view of the voxel importance pattern on the ipsilateral side.