TY - JOUR T1 - Refining the stratification of diffuse large B-cell lymphoma patients based on Metabolic Tumor Volume (MTV) by automatically adapting the MTV cut-off value to the segmentation method<strong/> JF - Journal of Nuclear Medicine JO - J Nucl Med SP - 274 LP - 274 VL - 61 IS - supplement 1 AU - Fanny Orlhac AU - Nicolo Capobianco AU - Anne-Segolene Cottereau AU - Laetitia Vercellino AU - Sven Zuehlsdorff AU - Olivier Casasnovas AU - Catherine Thieblemont AU - Michel Meignan AU - Irene Buvat Y1 - 2020/05/01 UR - http://jnm.snmjournals.org/content/61/supplement_1/274.abstract N2 - 274Objectives: In diffuse large B-cell lymphoma (DLBCL), Metabolic Tumor Volume (MTV) calculated using 18F-FDG PET images has been shown to be a valuable prognostic factor. Different MTV segmentation methods yield correlated MTV values in the same patient cohort and show comparable prognostic value [1]. However, the cut-off value to identify high MTV subjects having increased risk of shorter progression free-survival (PFS) or overall survival (OS) depends on the segmentation method [2] used to calculate the MTV and there is no consensus on the best segmentation method [3]. As a result, each center has to optimize the cut-off value as a function of the employed segmentation method, which strongly limits the use of MTV as a stratification criterion in clinical routine. We propose and validate a method to automatically determine how the cut-off value should be shifted between two segmentation methods. Methods: A cohort of 280 patients with DLBCL from the REMARC trial (NCT01122472) was used. All patients underwent a baseline 18F-FDG PET exam according to a standard protocol. For each patient, MTV was obtained using two methods. Method 1 (M1) consisted in segmenting volumes of interest (VOIs) with high uptake using a supervised segmentation algorithm involving component trees and shape priors, region growing and final region delineation using 41% of SUVmax. Method 2 (M2) was fully-automated and started with the identification of VOIs with SUVpeak significantly higher than blood pool SUV according to PERCIST recommendations and delineated using a threshold of 42% of SUVmax. The resulting VOIs were then automatically classified as suspicious or non-suspicious using a convolutional neural network model [4] trained on a separate cohort. For M1 and M2, the optimal MTV cut-off values TM1 and TM2 were identified by maximizing the Youden Index (YI=Sensitivity+Specificity-1) for predicting the occurrence of a survival event, for PFS and OS. We then used the ComBat harmonization method [5] to determine how the optimal cut-off value found for M1, TM1, should be converted to M2 MTV. ComBat uses the M1 and M2 MTV distributions to identify the linear transformation relating M1 and M2 estimates and applies this transformation to TM1 in order to estimate the cut-off appropriate for M2 estimates, denoted ESTM2 here. Last, the accuracy of PFS and OS prediction from the M2 MTV values was characterized using YI in 3 situations: using TM1, using TM2, and using ESTM2. Results: Using M1 MTV, the optimal cut-off value TM1 was to 242 mL for PFS and 225 mL for OS with YI respectively equal to 0.26 (Se=67% ; Sp=58%) and 0.31 (Se=78% ; Sp=53%). Using M2 MTV, the optimal cut-off value TM2 was to 112 mL for PFS and 150 mL for OS with YI respectively equal to 0.23 (Se=66% ; Sp=57%) and 0.28 (Se=67% ; Sp=61%). When TM1 was applied to MTV values obtained with M2, YI dropped to 0.18 (Se=41% ; Sp=77%) for PFS and 0.20 (Se=47% ; Sp=73%) for OS. The linear transformation identified by ComBat was MTVM2=0.61 x MTVM1 -28.64. This transformation led to ESTM2 of 118 mL for PFS, and with that cut-off value, the YI was 0.22 (Se=64% ; Sp=58%) for PFS, close to the performance obtained with the optimal TM2 cut-off value (YI = 0.23). For OS, ESTM2 was 108 mL, corresponding to a YI of 0.25 (Se=71% ; Sp=54%), against 0.28 when using the TM2 cut-off value optimized from M2 data. Conclusions: We demonstrated that ComBat harmonization method makes it possible to determine a conversion from a MTV cut-off value optimized for one segmentation method to another segmentation method. ComBat is data-driven and does neither require any phantom acquisition nor calibration. ComBat conversion potentially enables harmonization of MTV cut-off value adapted to each center practice for the prognosis of diffuse large B-cell lymphoma patients. ER -