Nimg-07. Applying a Glioma-Trained Deep Learning Auto-Segmentation Tool on Bm Pre- And Post-Radiosurgery

Abstract

PURPOSE Stereotactic radiosurgery (SRS) has become the mainstay to treat BM. Follow-up MRI provides important information on lesion treatment response and guides future therapy planning. Volumetric measurements of BM have shown promise over traditional uni- and two-dimensional measurements in more accurate and repeatable assessment. However, routine clinical use has yet to be achieved because the workflow is laborious. In previous work, we developed a PACS-integrated deep learning algorithm for automatic high- and low-grade glioma 3D segmentation. In this work, we applied this U-Net to segment BM on pre- and post-Gamma Knife (GK) MRI and evaluated the performance. METHODS 10 pre- and post-GK studies were autosegmented in five randomly selected patients (melanoma n= 3, breast n= 2). The glioma trained algorithm segmented the “Whole Tumor” (tumor core+peritumoral edema on T2w-FLAIR) and “Tumor Core” (CE tumor core+necrosis on SPGR). The AI generated segmentation was then revised as needed by a board-certified neuroradiologist and the dice-similarity-coefficient (DSC) between the revised and automatic volumetric segmentations were calculated. RESULTS Four patients had multicentric (2-4 BM) lesions. The mean± SD DSC for Whole Tumor and Tumor Core were 0.92±0.06 and 0.46±0.30 for pretreatment, 0.84±0.09 and 0.41±0.25 for posttreatment BM, respectively. The tool detected lesions with a sensitivity of 45% (5/11) for pretreatment and 50% (3/6) for posttreatment lesions. Three pretreatment and all posttreatment lesions that were not detected by the autosegmentation tool showed a very faint hyperintense peritumoral edema in T2w-FLAIR. CONCLUSION Volumetric segmentation of edema on FLAIR using the glioma-trained segmentation algorithm on pre- and post-GK BM did not require major adjustment of segmentation if it detects the lesion. On the other hand, with low sensitivity of lesion detection and low DSC for enhancing component, dedicated training of the algorithm on annotated BM data will be needed.