ESTRO 2021 Abstract Book

S520

ESTRO 2021

are surgically sutured near the tumour border on the scleral surface. During surgery, the clip-tumour distances are assessed and these clips and measurements are used for CTV marking and PBT planning. To support the localisation of the clips, target delineation and 3D modelling of the tumour, a dedicated MRI protocol was developed. Is this study we evaluate these MR-based clip-tumour measurements. Materials and Methods 23 consecutive UM patients planned for PBT were scanned on a 3T Philips MRI with a 47mm surface coil. The dedicated protocol was based on the earlier work of Ferreira [Cancers 11: 337, 2019], but modified to limit the susceptibility artefacts of the clips by localized shimming and an increased gradient strength of at least 23 mT/m. The protocol consisted of isotropic 3D T1 and T2 Spin Echo (SE) sequences, followed by multi-slice T2- weighted sequences, with a higher in-plane resolution, for each clip. To aid clip localisation, a T1-weighted gradient echo (GE) sequence was added. Differences in size of the clip artefact were evaluated in 3 patients. The clip-tumour distances were measured by a radiologist and compared to the peroperative measurements of the ophthalmologist. For clips distances with more than 1mm difference, the most likely origin of this discrepancy was discussed in a multidisciplinary setting by ophthalmologists, radiologists, radiation oncologist and MR-physicist. Results The GE sequences showed significantly larger signal-voids (5.1±0.4mm) compared to SE (3.0±0.4mm p<0.01). In 55% of the 87 evaluated clips, the MRI and peroperative measurements differed less than 1mm, fig 1. The largest differences, up to 14mm, were observed in flat UM (15% of the clips), where the MRI was generally considered less reliable. In 15% of the clips, the discrepancy between the MRI and peroperative measurement was attributed to a complex tumor geometry, fig 2AB. In 10% of the clips, the anterior localisation of the tumour casted a shadow on the eye wall, resulting in an overestimation of tumour dimensions during surgery, fig 2C. For these patients, the MR-imaging was considered to be more reliable.

Figure 1: Comparison between peroperative and MRI-based clip-tumour measurements.

Figure 2: A,B) 2 UM with a complex tumour-retina relation. C) Anterior UM are overestimated as the tumour casts a shadow.

Conclusion MRI and peroperative measurements are of equal value in most cases. For Flat UM, MRI underestimates tumor dimensions. For complex and anterior UM MRI-based clip-tumour distance is more reliable which could result in more accurate target definition, thereby possibly reducing toxicity and improving probability to retain vision. PH-0652 Synthetic CT from MRI with deep learning: Assessing the clinical impact of generated errors E. Alvarez Andres 1,2 , A. Gasnier 2,3 , C. Veres 2 , F. Dhermain 2,3 , S. Corbin 3 , F. Auville 3 , B. Biron 3 , A. Vatonne 3 , T. Henry 2,4 , T. Estienne 2,5 , M. Lerousseau 2,5 , L. Fidon 5 , E. Deutsch 2,3 , N. Paragios 1 , C. Robert 2,3 1 TheraPanacea, Radiotherapy, Paris, France; 2 Paris-Sud University - Gustave Roussy - Inserm - Paris-Saclay University, U1030 Molecular Radiotherapy and Innovative Therapeutics, Villejuif, France; 3 Gustave Roussy - Paris-Saclay University, Department of Radiotherapy, Villejuif, France; 4 Gustave Roussy - Paris-Saclay University, Department of Nuclear Medicine, Villejuif, France; 5 CentraleSupélec, Paris-Saclay University, MICS Laboratory, Gif-sur-Yvette, France Purpose or Objective The pseudo Computed Tomography (pCT) generation from Magnetic Resonance Imaging (MRI) using deep learning is a promising method toward improving Radiotherapy (RT). Most studies evaluate the pCT quality using the Mean Absolute Error (MAE). Yet, it is unclear how MAE relates to the dose calculation error. This