P13.09.B FULLY AUTOMATED BRAIN METASTASES SEGMENTATION USING T1-WEIGHTED CONTRAST-ENHANCED MR IMAGES BEFORE AND AFTER STEREOTACTIC RADIOSURGERY

Abstract

Abstract BACKGROUND Brain metastases (BM) represent the most common intracranial tumor in adults. An estimated 20% of all patients with cancer will develop BM. Stereotactic Radiosurgery (SRS) is a major treatment option for BM. For SRS treatment planning and outcome evaluation, magnetic resonance images (MRI) are acquired before and at multiple stages during the follow-up. Accurate segmentation of brain tumors on MRI is crucial for treatment planning and response evaluation. Detection and segmentation of BM which is a tedious and time-consuming task for many radiologists could be optimized with machine learning METHODS . The accuracy of the auto-segmentation, however, is influenced by the presence of false-positive and false-negative segmentation. There are studies which evaluated the segmentation performance of several deep learning algorithms, these were mainly focused on training and testing the models on either the pre-treatment or post-treatment images. The purpose of this study was to investigate a well-known deep learning approach (nnU-Net) for BM segmentation and to evaluate its performance on both pre-treatment and post-treatment images to assess if it could be a handy tool for the clinicians. METHODS Pre-treatment T1-weighted brain MRIs which were contrast-enhanced with triple-dose gadolinium were collected retrospectively for 266 patients with BM. Scans were made as part of clinical care at the Gamma Knife Center of the Elisabeth-TweeSteden Hospital (Tilburg, the Netherlands). All patients underwent Gamma Knife Radiosurgery, a form of SRS. This total of 266 patients were randomly split into 210 patients for model training/validation and 56 patients for testing. For these 56 patients used for testing, the post treatment follow-up T1-weighted brain MRI scans which were contrast-enhanced with single-dose gadolinium were also retrospectively collected. The nnU-Net model was trained with the pre-treatment training data, and then tested separately against the pre-treatment and follow-up data. RESULTS The model obtained a Dice score of 0.91 when tested with the pre-treatment images and a Dice score of 0.80 when tested with the follow-up after treatment T3 images. The False Negative Rate (FNR) when tested with the pre-treatment images was 0.07 and 0.24 when tested with post-treatment T3 images. CONCLUSION The model achieved a good performance score for pre-treatment images. The nnU-Net model can automatically detect and segment brain metastases with high sensitivity, and low FNR for treatment planning. This could be a beneficial tool for clinicians and assist SRS management for diagnosis and treatment planning. Though there is a decline in the Dice score and an increase in the FNR of the model for the post-treatment images, the performance still remained higher than in other similar studies.