Investigation of autosegmentation techniques on T2-weighted MRI for off-line dose reconstruction in MR-linac workflow for head and neck cancers

Abstract

Background: In order to accurately accumulate delivered dose for head and neck cancer patients treated with the Adapt to Position workflow on the 1.5T magnetic resonance imaging (MRI)-linear accelerator (MR-linac), the low-resolution T2-weighted MRIs used for daily setup must be segmented to enable reconstruction of the delivered dose at each fraction.

Purpose: In this pilot study, we evaluate various autosegmentation methods for head and neck organs at risk (OARs) on on-board setup MRIs from the MR-linac for off-line reconstruction of delivered dose.

Methods: Seven OARs (parotid glands, submandibular glands, mandible, spinal cord, and brainstem) were contoured on 43 images by seven observers each. Ground truth contours were generated using a simultaneous truth and performance level estimation (STAPLE) algorithm. Twenty total autosegmentation methods were evaluated in ADMIRE: 1-9) atlas-based autosegmentation using a population atlas library (PAL) of 5/10/15 patients with STAPLE, patch fusion (PF), random forest (RF) for label fusion; 10-19) autosegmentation using images from a patient's 1-4 prior fractions (individualized patient prior [IPP]) using STAPLE/PF/RF; 20) deep learning (DL) (3D ResUNet trained on 43 ground truth structure sets plus 45 contoured by one observer). Execution time was measured for each method. Autosegmented structures were compared to ground truth structures using the Dice similarity coefficient, mean surface distance (MSD), Hausdorff distance (HD), and Jaccard index (JI). For each metric and OAR, performance was compared to the inter-observer variability using Dunn's test with control. Methods were compared pairwise using the Steel-Dwass test for each metric pooled across all OARs. Further dosimetric analysis was performed on three high-performing autosegmentation methods (DL, IPP with RF and 4 fractions [IPP_RF_4], IPP with 1 fraction [IPP_1]), and one low-performing (PAL with STAPLE and 5 atlases [PAL_ST_5]). For five patients, delivered doses from clinical plans were recalculated on setup images with ground truth and autosegmented structure sets. Differences in maximum and mean dose to each structure between the ground truth and autosegmented structures were calculated and correlated with geometric metrics.

Results: DL and IPP methods performed best overall, all significantly outperforming inter-observer variability and with no significant difference between methods in pairwise comparison. PAL methods performed worst overall; most were not significantly different from the inter-observer variability or from each other. DL was the fastest method (33 s per case) and PAL methods the slowest (3.7-13.8 min per case). Execution time increased with a number of prior fractions/atlases for IPP and PAL. For DL, IPP_1, and IPP_RF_4, the majority (95%) of dose differences were within ± 250 cGy from ground truth, but outlier differences up to 785 cGy occurred. Dose differences were much higher for PAL_ST_5, with outlier differences up to 1920 cGy. Dose differences showed weak but significant correlations with all geometric metrics (R2 between 0.030 and 0.314).

Conclusions: The autosegmentation methods offering the best combination of performance and execution time are DL and IPP_1. Dose reconstruction on on-board T2-weighted MRIs is feasible with autosegmented structures with minimal dosimetric variation from ground truth, but contours should be visually inspected prior to dose reconstruction in an end-to-end dose accumulation workflow.

Keywords: MR-linac; autosegmentation; head and neck cancers.

Figure 1:

Execution time per case for each autosegmentation method. Data is represented as mean and standard deviation (error bars).

Figure 2:

Dice similarity coefficient (DSC) and mean surface distance (MSD) in mm for the autosegmentation methods compared to ground truth contours and the pair-wise comparison of inter-observer variability. Distributions are shown as box plots, with the five horizontal bars in each distribution representing the minimum, first quartile, median, third quartile, and maximum.

Figure 3:

Segmentations for six of the seven OARs (excluding the brainstem since it cannot be visualized in the same plane as all the other segmentations) for one example patient. Ground truth segmentations are in green, and each set of autosegmented contours is in red. The spinal cord contour is missing in the IPP_PF methods because the segmentation failed.

Figure 4:

Heat map of p-values for Dunn’s test with the inter-observer variability as a control. Reported values are Bonferroni adjusted p-values. Red boxes indicate non-significant results (p>0.05), and blue boxes indicate significant results (p<0.05). White represents missing values (IPP_PF methods failed for the spinal cord for every test case). (Abbreviations: DL = deep learning, IPP = individualized patient prior, PAL = population atlas library, ST = STAPLE, PF = patch fusion, RF = random forest. 1, 2, 3, and 4 represent the number of prior fractions in IPP. 5, 10, and 15 represent the number of atlases in PAL. Glnd_Submand = submandibular gland. L = left, R = right.)

Figure 5:

Heat map of p-values for the Steel-Dwass test for pair-wise comparison between autosegmentation methods pooled over all ROIs. Red boxes indicate non-significant results (p>0.05), and blue boxes indicate significant results (p<0.05). (Abbreviations: DL = deep learning, IPP = individualized patient prior, PAL = population atlas library, ST = STAPLE, PF = patch fusion, RF = random forest. 1, 2, 3, and 4 represent the number of prior fractions in IPP. 5, 10, and 15 represent the number of atlases in PAL.)

Figure 6:

Differences (cGy) in dosimetric performance criteria (ΔD_mean and ΔD_max) between plans based on ground truth contours and select autosegmentation methods for 5 patient cases. Positive/negative values mean that the dose was higher/lower in the autosegmented structure than in the ground truth structure, respectively. Treatment sites for the five patient cases are: case 1: larynx – left supraglottis; case 2: left hypopharynx; case 3: larynx – left glottis; case 4: oropharynx – right tonsil; case 5: oropharynx – right tonsil. (Abbreviations: D_mean = mean dose, D_max = maximum dose, L = left, R = right, Glnd_Submand = submandibular gland, DL = deep learning, IPP_1 = individualized patient prior with 1 prior fraction, IPP_RF_4 = individualized patient prior with 4 prior fractions and random forest, and PAL_ST_5 = population atlas library with 5 atlases and STAPLE.)