ESTRO 2021 Abstract Book

S407

ESTRO 2021

The proposed DL algorithm can safely generate sCT images in the thoracic district starting from 0.35T MRI. The sCT generated offer an accuracy level sufficient to calculate dose distributions of IMRT plans in presence of large ED heterogeneity. If confirmed on a larger cohort of patients, this approach could allow of removing the CT acquisition from the MRgRT clinical practice, paving the way towards more efficient MR-only RT paradigms. OC-0522 Cell-Rad: Towards Histology-driven Radiation Oncology from Multi-Parametric MRI A. Leroy 1,4 , K. Shreshtha 2 , M. Lerousseau 3,4 , T. Henry 4 , T. Estienne 4 , M. Classe 5,4 , N. Paragios 2 , E. Deutsch 4 , V. Grégoire 6 1 Therapancea, Radiotherapy, Paris, France; 2 Therapanacea, Radiotherapy, Paris, France; 3 Paris-Saclay University, CentraleSupélec, Gif-Sur-Yvette, France; 4 Paris-Saclay University, Gustave Roussy, Inserm 1030, Molecular Radiotherapy and Therapeutic Innovation, Villejuif, France; 5 Gustave Roussy, Pathology Department, Villejuif, France; 6 Centre Léon Bérard, Radiation Oncology Department, Lyon, France Purpose or Objective The pathological analysis of biopsy specimens is essential to cancer diagnosis, treatment selection and prognosis. However, biopsies are only taken from part of the tumor and cannot assess the full cellular extension. Such information is essential to accurately delineate the tumor volume. The aim of our work is to provide alternative means to gather clinical information at the cellular through MR image translation towards virtual pathological content generation. The latter could directly be integrated to the treatment planning either for better delineation or to reinvent the irradiation paradigm with dose painting. Materials and Methods The training set consists of a TCIA publicly available cohort gathering 26 patients with in vivo prostate T2 MR volumes and histopathological slices from prostatectomy specimen, resulting in 83 pairs of images, split into 50/10/23 for train/validation/test sets, with an MR-defined resolution. Conventional approaches to address the radiology-pathology fusion aspect require paired data that is cumbersome to achieve due to tissue collapse, different resolution scales and lack of plane correspondences. Our work’s novelty lays on the use of weakly paired images on which unsupervised learning is performed. We build a synthesis pipeline made of two cycle Generative Adversarial Networks (cycleGANs), the first one generating aligned ground truth histology, the second one improving synthesis on registered pairs in a self-supervised setting. Results We quantitively compared the generated images to original ones with four metrics, displayed in Table 1. On the held-out test set, our model achieves a mean absolute error (MAE) of 5.9 ± 1.7 × 10-2, proving the high quality of histopathological inference. We benchmarked it with simpler architectures to highlight the substantial gain in performance when concatenating two cycleGANs. Finally, qualitative assessment of reconstruction quality is depicted in Figure 1. Once trained, the inference pathway is fast to run and only requires MR volumes with a limited memory footprint.

Figure 1: Two samples with (left to right) original MR, synthetic histology, ground-truth histology

Table 1: Results with two benchmarks and final model. Four metrics compare generated images to the ground- truth original ones.