An advanced automated pipeline for brain tumour segmentation on magnetic resonance imaging for gamma knife radiosurgery

Prabhakar Ramachandran; Akshay C Dhargave; Zachery M Colbert; Akash Mehta; Catherine Jones; Ryan Lusk; Michael Jenkins; Heath Foley; Mihir Shanker; Michael Huo; Matthew Foote; Mark B Pinkham

doi:10.1016/j.phro.2026.101010

An advanced automated pipeline for brain tumour segmentation on magnetic resonance imaging for gamma knife radiosurgery

Phys Imaging Radiat Oncol. 2026 Jun 2:39:101010. doi: 10.1016/j.phro.2026.101010. eCollection 2026 May.

Authors

Prabhakar Ramachandran^{1

2

3}, Akshay C Dhargave¹, Zachery M Colbert¹, Akash Mehta¹, Catherine Jones¹, Ryan Lusk¹, Michael Jenkins¹, Heath Foley¹, Mihir Shanker¹, Michael Huo¹, Matthew Foote¹, Mark B Pinkham^{1

3

4}

Affiliations

¹ Department of Radiation Oncology, Princess Alexandra Hospital, Woolloongabba, Queensland 4102, Australia.
² Australian Institute for Bioengineering and Nanotechnology, The University of Queensland, St Lucia, Queensland 4072, Australia.
³ Queensland University of Technology, Brisbane, Queensland 4000, Australia.
⁴ Flinders University, Bedford Park, South Australia 5042, Australia.

Abstract

Background and purpose: Accurate delineation of intracranial tumours is crucial for stereotactic radiosurgery (SRS), where target definition directly influences treatment outcome. We developed and clinically integrated an automated multi-tumour segmentation pipeline using three-dimensional nnU-Net models for brain metastases, pituitary adenomas, vestibular schwannomas, and meningiomas.

Materials and methods: Four independent, tumour-specific models were trained on T1-weighted Magnetisation-Prepared RApid Gradient Echo Magnetic Resonance Imaging data using the nnU‑Net architecture. For model development, 100 cases per tumour-type (n = 400) were used, and to evaluate the clinical workflow, 25 additional cases per tumour-type (n = 100) were processed prospectively. The performance was assessed using the Dice Similarity Coefficient (DSC), the 95th-percentile Hausdorff Distance (HD95), and the Average Symmetric Surface Distance (ASSD). The pipeline continuously monitored incoming Digital Imaging and Communications in Medicine (DICOM) images using a listener and applied the appropriate tumour-specific segmentation model. It, then, automatically exported the DICOM images and the inferred Radiotherapy Structure-Set to the treatment planning system.

Results: Among all the tumour-types, vestibular schwannomas achieved the highest performance (DSC: 0.90 ± 0.03; HD95: 0.93 ± 0.34 mm; ASSD: 0.31 ± 0.09 mm) followed by brain metastases (DSC: 0.83 ± 0.08; HD95: 1.33 ± 0.55 mm; ASSD: 0.47 ± 0.19 mm), pituitary adenomas (DSC: 0.81 ± 0.09; HD95: 2.39 ± 1.14 mm; ASSD: 0.78 ± 0.32 mm) and meningiomas (DSC: 0.80 ± 0.11; HD95: 4.46 ± 3.64 mm; ASSD: 1.19 ± 0.80 mm). All tumour-types were segmented with consistent performance (SD_DSC < 0.11), and segmentation was completed within two to four minutes per case.

Conclusions: The auto-segmentation pipeline enabled consistent and rapid delineation of multiple intracranial tumours, achieving clinically acceptable performance metrics and efficiency suitable for SRS.

Keywords: Auto-segmentation; Brain tumours; Deep learning; Gamma Knife; MRI; Stereotactic radiosurgery; nnU‑Net.