2098 - Novel Application of a 3D U-Net Convolutional Neural Network for Detection, Segmentation, and Radiotherapy Response Assessment of Central Nervous System Lymphoma

02:30pm - 04:00pm PT

Hall F

Screen: 16

POSTER

Presenter(s)

Brianna Hostler, - University of California San Diego Health, San Diego, CA

B. Hostler¹, J. Rudie², R. Saluja³, E. Reid⁴, B. Heyman⁴, W. Pearse⁴, M. Choi⁴, J. A. Hattangadi-Gluth⁵, P. Sanghvi⁵, and K. R. Tringale⁵; ¹Department of Psychology, San Diego State University, San Diego, CA, ²Department of Radiology, University of California San Diego, La Jolla, CA, ³Cornell University, Ithaca, NY, ⁴Department of Medicine, University of California San Diego, La Jolla, CA, ⁵Department of Radiation Medicine and Applied Sciences, University of California San Diego, La Jolla, CA

Purpose/Objective(s):

The role of radiotherapy (RT) in central nervous system lymphoma (CNSL) is evolving with novel therapeutics. Understanding patterns of disease failure and response courses for primary (PCNSL) and secondary CNSL (SCNSL) is critical, yet deep learning (DL) based MRI techniques have only been explored in PCNSL. We hypothesized that a 3D U-net convolutional neural network (CNN) for brain metastases applied to CNSL can demonstrate post-RT response and relapse patterns in SCNSL vs PCNSL.

Materials/Methods:

RT-treated CNSL patients at a single institution were identified. Disease failure was characterized as local (involving initial site) or distant. Overall (OS) and CNS progression-free survival (CNS PFS) were assessed from RT start with Kaplan-Meier analysis. A 3D U-net CNN was applied to MRIs in those with parenchymal CNSL and =1 pre-RT T1-post-contrast MRI. Paired sample t-tests assessed CNSL volume change from pre- to post-RT MRI. A linear mixed effects model assessed tumor volume within 1yr of RT.

Results:

Of 30 patients (43% PCNSL, 57% SCNSL), most received RT for relapsed/refractory (23, 77%) vs definitive (6, 20%) or consolidative (1, 3%) intent. At time of RT, 1 (3%) had CSF, 13 (43%) had ocular, and 12 (40%) had extracranial disease. RT dose was higher for PCNSL (median 40.5 Gy vs 30 Gy for SCNSL, p=.03). RT targeted whole brain (WBRT; 17, 57% [6 + orbits]), orbits alone (10, 33%), focal (2, 7%), and spine (1, 3%). 17 patients relapsed: 9 (53%) distant, 5 (29%) local, 3 (18%; all SCNSL) synchronous local and distant. After WBRT, 2 had local-only (both SCNSL) and 2 had distant-only (both PCNSL) failures. Those with ocular disease received orbital +/- WBRT; most (8, 62%) had parenchymal relapse. Median OS was longer for PCNSL (19m [95%CI 0-42] vs 6m [95%CI 3-9]) for SCNSL; log-rank p=.05). Median CNS PFS was similar for PCNSL vs SCNSL (2.8m [95%CI 0-20] vs 4.4m [95%CI 1-7], respectively; log-rank p=.26).

20 patients (50% PCNSL, 50% SCNSL) had a median of 3 MRIs (IQR 1-5). 14 had a pre- and post-RT MRI with a median pre-RT DL-derived CNSL volume of 5.1 cm3 (IQR 3-14) at a median of 7d pre-RT (IQR 3-13). Pre-RT volume was similar between SCNSL and PCNSL (17.9 vs 10.5 cm3, respectively; p=.27). At a median of 4m post-RT (IQR 3-5), median CNSL volume was 1.6 cm3 (IQR 0.6-7.5), reduced from pre-RT (mean difference 4.3 cm3; p=.01) with similar degree of reduction between PCNSL and SCNSL (p=.81). CNSL volume decreased longitudinally post-RT (B=-0.5cm3/month, p=.09).

Conclusion:

Different relapse patterns were seen in SCNSL vs PCNSL receiving CNS-RT for mostly relapsed/refractory disease, suggesting differential approaches to RT targets. This is the first application of a DL-based MRI technique to characterize SCNSL and RT response, showing similar significant reduction in tumor volume post-RT for SCNSL and PCNSL. Work is underway to further optimize this algorithm for CNSL.