ESTRO 2023 - Abstract Book

S584

Monday 15 May 2023

ESTRO 2023

Conclusion A national auto-segmentation validation repository of 26 patients has been created. Areas of low image contrast present more contouring variance, and for OAR split into substructures, the transition causes a significantly increased variance. MO-0714 Statistical comparison between GTV and gold standard contour on AI-based registered histopathology A. Leroy 1,2,3 , A. Cafaro 1,2,3 , A. Champagnac 4 , M. Classe 2 , G. Gessain 5 , N. Benzerdjeb 6 , P. Gorphe 7 , P. Zrounba 8 , V. Lepetit 9 , N. Paragios 1 , E. Deutsch 2 , V. Grégoire 3 1 Therapanacea, Artificial Intelligence, Paris, France; 2 Gustave Roussy, Paris-Saclay University, Inserm 1030, Molecular Radiotherapy and Therapeutic Innovation, Villejuif, France; 3 Centre Léon Bérard, Radiation Oncology, Lyon, France; 4 Centre Léon Bérard, Pathology, Lyon, France; 5 Gustave Roussy, Paris-Saclay University, Pathology, Villejuif, France; 6 University Hospital of Lyon, Pathology, Lyon, France; 7 Gustave Roussy, Paris-Saclay University, Surgery, Villejuif, France; 8 Centre Léon Bérard, Surgery, Lyon, France; 9 Université Paris-Est, École des Ponts ParisTech, CNRS, Laboratoire d'Informatique Gaspard-Monge, Marne-la-Vallée, France Purpose or Objective Accurate delineation of GTV is crucial for proper dose prescription and treatment implementation. Manual contouring on radiologic imaging is the usual procedure but is time-consuming and suffers from interobserver variability because of the poor quality of acquisitions. Histopathology images from the resection of tumor specimens represent the gold standard for tumor characterization by providing cell-level information on disease invasion. The purpose of this study is two-fold: First, to assess variability in radiology-based GTVs from two expert centers. Second, to extensively compare these GTVs with tumor contours on the histological slides, after AI-based automatic registration of both modalities. The biological insights from histology can help practitioners better understand tumor environment in radiology towards reducing toxicity. Materials and Methods We collected a cohort from 77 patients (joint collaboration Gustave Roussy Institute - Centre Léon Bérard, France) on whom were acquired both pre-operative H&N CT scan and digitalized histopathology after total laryngectomy. For each center, junior and senior radiation oncologists delineated the GTV on CT, while senior pathologists contoured macroscopic tumor on histology. We applied a deep learning model to automatically register both images and deform the tumor masks consequently. We then computed various cross-center statistics to (i) highlight interobserver variability between radiation oncologists, (ii) validate the hypothesis of concordance between pathologists, and (iii) assess systematic errors on GTV delineation compared to histology, whether it be missed area or overestimation of disease extent. Results According to GTV interobserver variability, we report a Dice score of 0.63 between juniors and 0.68 between seniors. This highlights the limitations of CT images for precise delineation, even if a higher experience of observers improves agreement with a reduced mean tumor volume (-23%). Conversely, expert pathologists reach an excellent concordance with a Dice score of 0.96, proving the benefit of histological examination for better diagnosis. After co-registration of both modalities and the consensus contours, it appears that GTV is systematically overestimated compared to the histology-based contour, with a mean increased volume of 29% and a low Dice score of 0.59.