PT  - JOURNAL ARTICLE
AU  - Nicolo Capobianco
AU  - Michel Meignan
AU  - Anne Segolene Cottereau
AU  - Laetitia Vercellino
AU  - Ludovic Sibille
AU  - Bruce Spottiswoode
AU  - Sven Zuehlsdorff
AU  - Olivier Casasnovas
AU  - Catherine Thieblemont
AU  - Irene Buvat
TI  - Fully automated deep learning FDG uptake classification enables Total Metabolic Tumor Volume (MTV) estimation in diffuse large B-cell lymphoma with similar predictive value as expert MTV measurements
DP  - 2020 May 01
TA  - Journal of Nuclear Medicine
PG  - 504--504
VI  - 61
IP  - supplement 1
4099  - http://jnm.snmjournals.org/content/61/supplement_1/504.short
4100  - http://jnm.snmjournals.org/content/61/supplement_1/504.full
SO  - J Nucl Med2020 May 01; 61
AB  - 504Objectives: Total Metabolic Tumor Volume (MTV) calculated from 18F-FDG PET/CT baseline studies is a prognostic factor in diffuse large B-cell lymphoma (DLBCL). Yet, MTV measurement requires the segmentation of all malignant foci throughout the body. There is currently no consensus on the most accurate approach for such segmentation and all methods still require extensive manual input from an experienced reader. We determined whether a deep learning (DL) based method could estimate MTV with the same predictive value as MTV measured by experts. Methods: Baseline 18F-FDG PET/CT images of 280 DLBCL patients from the REMARC trial (NCT01122472) were retrospectively analyzed. A fully automated whole-body high uptake segmentation algorithm was used to identify all 3D regions (ROI) with tracer uptake significantly above the blood pool according to PERCIST recommendations1. The resulting ROIs were processed by a research prototype and classified as nonsuspicious or suspicious uptake using a convolutional neural network trained on an independent cohort2. DL-based MTV (DL-MTV) was estimated as the sum of the volumes of ROIs classified as suspicious uptake. Reference MTV (REF-MTV) was measured by two experienced readers using a supervised segmentation algorithm involving automated high uptake region detection using component trees and shape priors, region growing and final region delineation using 41% of SUVmax, followed by manual addition of missed regions and deletion of physiological regions. DL-MTV was compared to REF-MTV in terms of prognostic value for progression free survival (PFS) and overall survival (OS)3. Results: DL-MTV was significantly correlated with the REF-MTV (Spearman rho=0.76, P<.001). Using the DL-based approach, an average of 24 regions per subject with increased tracer uptake were identified, and an average of 20 regions per subject were correctly identified as nonsuspicious or suspicious compared to the REF-MTV region. Median Dice score between the DL-MTV region and the REF-MTV region was 0.73 (Interquartile Range 0.33-0.86). Both MTV measurements were predictive of PFS (P<.001, HR=2.4 and 2.6 for DL-MTV and REF-MTV respectively) and OS (P<.001, HR=2.8 and 3.7 for DL-MTV and REF-MTV). Conclusions: MTV derived with the fully automated DL algorithm was consistent with the MTV obtained by experts using an independent method and was predictive of PFS and OS. Classification of high uptake regions using deep learning for discarding physiological uptake may considerably simplify MTV estimation, reduce observer variability and facilitate the use of MTV as a predictive factor in DLBCL patients.