Clinical target volumes for glioma - Automated delineation to improve neuroanatomic consistency

Abstract

Background and purpose: Delineating clinical target volumes (CTVs) for glioma is challenging as consistency with the neuroanatomy needs to be carefully verified. We developed an automated approach that incorporates tumor infiltration pathways and anatomic barriers to improve the neuroanatomical consistency and efficiency of CTV delineation.

Materials and methods: A deep learning model for brain structure segmentation was developed based on manual delineations of hemispheres, brainstem, cerebellum, optic chiasm, optic nerves, ventricles, and midline on CT images of ninety-nine glioma patients. Brain structures predictions are integrated into a constrained distance transform that defines the CTV as a 15-mm expansion of the gross tumor volume. Connecting structures with white matter tracts allow for expansions across different structure boundaries, e.g., cerebellum and brainstem connecting at the cerebellar peduncles.

Results: Mean (±std) Dice Similarity Coefficient (DSC) for the hemispheres, brainstem, cerebel-lum, chiasm, optic nerves, midline and ventricles were (98.5 ± 0.8)%, (92.5 ± 2.8)%, (96.7 ± 2.2)% (63.9 ± 12.2)%, (83.8 ± 9.0)%, (81.2 ± 7.0) and (91.5 ± 3.9)%. Mean (±std) 95 % Hausdorff distance (HD95) were, in mm, 1.9 ± 2.5, 7.0 ± 5.4, 1.8 ± 1.2, 7.2 ± 3.2, 2.3 ± 1.0, 9.5 ± 10.5, and 3.8 ± 3.1, respectively. Auto-generated CTVs are compared against reference CTVs (15-mm expansion constrained by manually-contoured brain structures). The automatic CTVs showed excellent similarity to the reference CTVs with mean (±std) Surface DSC with 2 mm tolerance and HD95 scores of (95.6 ± 3.4)% and (1.4 ± 1.2) mm, respectively. A physician's quality assessment reported that the automated method would result in a substantial amount of time saved in 85 % of CTV delineations.

Conclusion: We have successfully incorporated expert knowledge to improve the neuroanatom-ical consistency of automatically-generated CTVs for glioma.

Keywords: Brain; Clinical target volume; Deep learning; Radiat. Oncol.

Graphical abstract

Fig. 1

Example of how anatomical barriers are accounted for in the delineation of the brain structures. (A) The brain hemispheres are anatomically separated by the falx cerebri, and the brain hemispheres are separated from the brainstem by the dura mater. (B) The brain hemispheres are separated from the cerebellum by the tentorium cerebelli. (C) The third ventricle connects to the lateral ventricles’ atrium towards the posterior horns.

Fig. 2

Example of brain structure delineations on the CT image of an example patient. The delineated structures are brain hemispheres, brainstem, cerebellum, chiasm, optic nerves, and ventricles. An auxiliary structure called the midline defines the separation between the two hemispheres. The arrows indicate locations where delineations were intentionally overlapped to ensure a seamless connection between structures (optic chiasm-to-brain, brain-to-brainstem and cerebellum-to-brainstem tracts).

Fig. 3

The CTVs produced by different methods are compared: the reference CTV (Ref CTV), the auto-generated CTV (Auto CTV), and the TotalSegmentator CTV (TS CTV). Figures A and B illustrate the expansion of the CTVs from the GTV inferiorly into the brainstem while avoiding the tentorium cerebelli. Figure C compares CTV expansion from the left hemisphere into the right hemisphere through the corpus callosum and into the septum pellucidum, while avoiding the falx cerebri and ventricles, respectively.

Fig. 4

Box- and violin plots showing the performance of the auto-generated CTVs (auto CTV) versus TotalSegmentator CTVs in terms of Dice Similarity Coefficient (DSC), Surface Dice Similarity Coefficient (SDSC) with 2 mm tolerance, and 95 % Hausdorff distance (HD95).