ESTRO 2023 - Abstract Book

S807

Monday 15 May 2023

ESTRO 2023

Materials and Methods The optical system includes four cameras and two pattern projectors mounted on the nozzle (Fig. 1-a) allowing for pupil tracking via stereo-triangulation and three-dimensional fringe pattern profilometry of the visible surface of the eye. Fringe pattern analysis, based on the Fourier transform method, and ad-hoc calibration procedures were implemented to map the pattern deformation into physical coordinates that are consistent with the pupil localisation.

For comparing the optical-based tumour position estimates with X-ray measurements, the MR-based eye model is aligned to the clips position, through point-based registration, thus defining a reference tumour position according to the clinical standard. Afterwards, the corneal topography and pupil position (Fig. 2) were used in a constrained iterative closest point registration to obtain a corresponding clipless estimate of the eye position. The two measurements were compared to evaluate the performance of the optical system in measuring the tumour centre of mass position.

Results Seventeen measurements of the eye position of three eye tumour patients acquired during the course of the treatment were included in the analysis. On average (standard deviation) the discrepancy in tumour position estimate was -0.21mm (1.10mm), 0.17mm (0.67mm), 0.29mm (0.84mm) and 1.33mm (0.79mm) in the X, Y, Z directions and in 3D, respectively. The data show that corneal topography of the entire visible ocular region is not always possible due to the absence of reflection of the projected pattern by the pupil (see Fig. 1). A moderate positive correlation between discrepancies and fringe pattern visibility within the eye was found (Pearson’s rho=0.51, p-value<0.05). Of note, when the fringe pattern was recognizable on more than 80% of the visible eye region, the mean discrepancy on 3D target localisation went down to 0.76 mm (0.20mm). Conclusion Sub-millimetric agreement in tumour position localisation was reported for the optical system compared to X-ray clips imaging, provided the projected fringe pattern was discernible in more than 80% of the region of interest (Fig. 1-b,c). In principle, an improvement of the reflection properties of the eye surface in the pupillary region could be achieved through the instillation of fluorescein. Corneal topography could prove useful not only to non-invasively verify treatment geometry but also to minimise range uncertainties, especially for patients treated through the eyelid (Fig. 2-c). MO-0959 A novel multi-task hybrid deep neural network (DNN) predicts tumor progression during MRgRT J.M. Bryant 1 , P.G. Saghand 2 , K. Latifi 1 , J. Frakes 1 , S. Hoffe 1 , E. Moros 1 , K. Mittauer 3 , R. Kotecha 3 , I. El Naqa 2 , S. Rosenberg 1 1 H. Lee Moffitt Cancer Center & Research Institute, Department of Radiation Oncology, Tampa, USA; 2 H. Lee Moffitt Cancer Center & Research Institute, Department of Machine Learning, Tampa, USA; 3 Miami Cancer Institute, Department of Radiation Oncology, Miami, USA