Author: Zachery Colbert, Matthew Foote, Michael Huo, Mark Pinkham, Prabhakar Ramachandran, Mihir Shanker 👨🔬
Affiliation: Radiation Oncology, Princess Alexandra Hospital, Ipswich Road, Princess Alexandra Hospital 🌍
Purpose: The study aimed to develop and implement deep learning-based autosegmentation models for the autosegmentation of four key tumor types: brain metastasis, pituitary adenoma, vestibular schwannoma, and meningioma. This process was designed to enhance workflow and support clinicians in tumor volume delineation directly from MRI images.
Methods: Four nnU-Net deep learning autosegmentation models were developed, each trained on annotated MRI datasets corresponding to the specific tumor types. Each model was trained and validated on a cohort of 100 patients. Once MRI scans are acquired, the images are transferred to a dedicated deep learning computer equipped with all four models and a watchdog system. The watchdog system monitors all incoming DICOM scans, classifies images based on histology, and applies the appropriate segmentation model. The segmented tumor volumes are then transferred to the GammaPlan treatment planning system for review prior to integration into the radiotherapy planning workflow.
Results: Of all four deep learning models, vestibular schwannoma achieved the best validation DSC of over 0.91 followed by brain metastasis. The lowest mean DSC was observed with pituitary adenoma. The models' performances were influenced by image quality, tumor complexity and variations in tumor morphology.
Conclusion: The implementation of deep learning-based autosegmentation models for brain metastasis, pituitary adenoma, vestibular schwannoma, and meningioma enables automating tumor volume segmentation and aids clinicians in reviewing and treating Gamma Knife patients with a short turnaround time. The automated system will ensure consistent tumor volume segmentation, minimizing inter-observer variability and streamline the workflow reducing the bottleneck with manual segmentation. By streamlining the process, clinicians can spend more time on planning and decision-making rather than manual segmentation, potentially improving patient outcomes. Future work will focus on refining models' performances and developing models for other tumors.