2239 - Deep Learning Based Automated Delineation of Vestibular Schwannomas: Geometric and Dosimetric Validation

Purpose/Objective(s): The accurate delineation of vestibular schwannoma (VS) targets is crucial to guide the treatment planning for stereotactic radiosurgery, and contouring automation can improve the efficiency of both treatment and monitoring. To this end, we developed a deep learning (DL)-based VS autosegmentation module and integrated it into a web-based platform for clinical use.

Materials/Methods: Our auto-segmentation module was developed using the nn-UNet framework and trained on publicly available VS patient data (paired post-contrast T1 MRI images and structure datasets) from The Cancer Imaging Archive. The trained VS auto-segmentation module was validated on a cohort of 76 patients who underwent radiosurgery treatment at our institution. We used a series of scripts to automate the import of the patient data into the platform, enabling the efficient evaluation of the model’s performance. By leveraging the clinically approved contours and dose distributions as benchmarks, we assessed the geometric accuracy of the auto-generated contours and their impact on dosimetry quality. The geometric accuracy was evaluated with DICE coefficient (DC), mean surface-to-surface distance (mSSD), 95 percentile Hausdorff distance 95 (HD95), and center-of-mass distance (COMD). The dosimetric impact was assessed by comparing auto-segmented volumes to the volumes covered by prescription isodose (RxIDL). In addition to DC and mSSD, we introduced the directional surface-to-surface distance (dSSD) metric to assess potential over- or under-segmentation, with positive values indicating under-segmentation and negative values indicating over-segmentation.

Results: Quantitative evaluations showed the model was able to accurately auto-delineate most VS lesions, with DC of 0.84±0.08, HD95 of 0.95±0.45 mm (~1-2 voxels on MR images), mSSD of 0.22±0.18 mm and COMD of 0.64±0.37 mm. The auto-segmented lesions showed a slightly greater difference from the prescription isodose line than the clinical contours, with a lower DC (0.81±0.1 vs 0.91±0.04) and a higher dSSD (1.00±2.25 vs 0.74±2.23). Segmentation accuracy was found to be lower for small (<0.5 cc) or cystic lesions, and within the auditory canal. The entire segmentation process for each case, from importing DICOM images to contouring and writing out the final DICOM RT structure, took approximately 2 minutes.

Conclusion: Our DL-based auto-segmentation module demonstrated high geometric accuracy and efficiency, suggesting its potential to effectively support treatment planning and monitoring workflows in vestibular schwannoma disease management.