Automatic Head and Neck Tumor segmentation and outcome prediction relying on FDG-PET/CT images: Findings from the second edition of the HECKTOR challenge

• DOI: 10.1016/j.media.2023.102972
• Publication Date: 2023-09-18T16:25:04Z
• Authors: Vincent Andrearczyk; Valentin Oreiller; Sarah Boughdad; Catherine Cheze L...; Olena Tankyevych; Hesham Elhalawani; Mario Jreige; John O. Prior; Martin Vallières; Dimitris Visvikis; Mathieu Hatt; Adrien Depeursinge
• Keywords: Leverage (statistics)

By focusing on metabolic and morphological tissue properties respectively, FluoroDeoxyGlucose (FDG)-Positron Emission Tomography (PET) Computed (CT) modalities include complementary synergistic information for cancerous lesion delineation characterization (e.g. outcome prediction), in addition to usual clinical variables. This is especially true Head Neck Cancer (HNC). The goal of the HEad neCK TumOR segmentation prediction (HECKTOR) challenge was develop compare modern image analysis methods best extract leverage this automatically. We present here post-analysis HECKTOR 2nd edition, at 24th International Conference Medical Image Computing Computer-Assisted Intervention (MICCAI) 2021. scope substantially expanded compared first by providing a larger population (adding patients from new center) proposing an additional task challengers, namely Progression-Free Survival (PFS). To end, participants were given access training set 224 cases 5 different centers, each with pre-treatment FDG-PET/CT scan Their subsequently evaluated held-out test 101 two centers. For (Task 1), ranking based Borda counting their ranks according metrics: mean Dice Similarity Coefficient (DSC) median Hausdorff Distance 95th percentile (HD95). PFS task, challengers could use tumor contours provided experts 3) or rely own 2). obtained Concordance index (C-index) calculated predicted risk scores. A total 103 teams registered challenge, 448 submissions 29 papers. method average DSC 0.759, predictions C-index 0.717 (without relying contours) 0.698 (using expert contours). An interesting finding that reached DL approaches (with without explicit segmentation, 4 out ranked) standard radiomics using handcrafted features extracted delineated tumors, exploiting alternative (automated and/or volumes encompassing surrounding tissues) rather than contours. second edition confirmed promising performance fully automated primary PET/CT images HNC patients, although there still margin improvement some difficult cases. time, also addressed relatively good (C-index above 0.7). Both results constitute another step forward toward large-scale studies HNC.