TY - JOUR T1 - Distinction of lymphoma from sarcoidosis at FDG PET/CT - evaluation of radiomic-feature guided machine learning versus human reader performance JF - Journal of Nuclear Medicine JO - J Nucl Med DO - 10.2967/jnumed.121.263598 SP - jnumed.121.263598 AU - Pierre Lovinfosse AU - Marta Ferreira AU - Nadia Withofs AU - Alexandre Jadoul AU - Celine Derwael AU - Anne-Noelle Frix AU - Julien Guiot AU - Claire Bernard AU - Anh Nguyet Diep AU - Anne-Francoise Donneau AU - Marie Lejeune AU - Christophe Bonnet AU - Wim Vos AU - Patrick E. Meyer AU - Roland Hustinx Y1 - 2022/05/01 UR - http://jnm.snmjournals.org/content/early/2022/05/19/jnumed.121.263598.abstract N2 - Sarcoidosis and lymphoma often share common features on 18F-FDG PET/CT, such as intense hypermetabolic lesions of lymph nodes and multiple organs. We aimed at developing and validating radiomics signatures to differentiate sarcoidosis from Hodgkin (HL) and diffuse large B-cell (DLBCL) lymphoma. Methods: We retrospectively collected 420 patients (169 sarcoidosis, 140 HL and 111 DLBCL) who underwent a pretreatment 18F-FDG PET/CT at the University Hospital of Liege. The studies were randomly distributed to 4 physicians who gave their diagnostic suggestion between the 3 diseases. Individual and pooled performances of physicians were then calculated. The inter-observer variability was evaluated using a sample of 34 studies interpreted by all physicians. Volumes of interest (VOI) were delineated over the lesions and the liver using MIM software, and 215 radiomic features were extracted using Radiomics toolbox. Models were developed combining clinical data (age, gender and weight) and radiomics (original and tumor-to-liver TLR radiomics), with 7 different feature selection approaches and 4 different machine learning (ML) classifiers, to differentiate sarcoidosis and lymphomas on both lesion-based and patient-based approaches. Results: For identifying lymphoma vs. sarcoidosis, physicians’ pooled sensitivity, specificity, area under the curve (AUC) and accuracy were 0.99 (CI95%:0.97-1.00), 0.75 (CI95%: 0.68-0.81), 0.87 (CI95%: 0.84-0.90) and 89.3%, respectively, whereas for identifying HL in the tumor population, it was 0.58 (CI95%: 0.49-0.66), 0.82 (CI95%: 0.74-0.89), 0.70 (CI95%: 0.64-0.75) and 68.5%, respectively. A moderate agreement was found between observers for the diagnosis of lymphoma vs. sarcoidosis and HL vs. DLBCL with Fleiss kappa values of 0.66 (CI95%: 0.45-0.87) and 0.69 (CI95%: 0.45-0.93), respectively. The best ML models for identifying lymphoma vs. sarcoidosis showed AUC of 0.94 (CI95%: 0.93-0.95) and 0.85 (CI95%: 0.82-0.88) in lesion- and patient-based approaches, respectively, using TLR radiomics (+ age for the second). To differentiate HL and DLBCL, we obtained AUC of 0.95 (CI95%: 0.93-0.96) in lesion-based approach using TLR radiomics, and 0.86 (CI95%: 0.80-0.91) in patient-based using original radiomics and age. Conclusion: Characterization of sarcoidosis and lymphoma lesions is feasible using ML and radiomics, with very good to excellent performances, equivalent or better than those of doctors who showed significant interobserver variability in their assessment. ER -