ESTRO 2024 - Abstract Book

S4510

Physics - Machine learning models and clinical applications

ESTRO 2024

Keywords: SBRT, Complete response, Gynecological cancer

1963

Proffered Paper

Identifying spatial dose patterns predictive of haematological toxicity after prostate cancer RT

Carlos Sosa-Marrero 1 , Maddalena Pavarini 2 , Lisa Alborghetti 2 , Umang Pandey 3 , Javier Pascau 3 , Renaud de Crevoisier 1 , Tiziana Rancati 4 , Cesare Cozzarini 5 , Oscar Acosta 1 , Claudio Fiorino 2 1 Univ Rennes, CLCC Eugène Marquis, Inserm, LTSI-UMR 1099, F-35000, Rennes, France. 2 IRCCS San Raffaele Scientific Institute, Medical Physics, Milano, Italy. 3 Universidad Carlos III de Madrid, Bioengineering Department, Madrid, Spain. 4 Fondazione IRCCS Istituto Nazionale dei Tumori, Progetto Prostata, Milano, Italy. 5 IRCCS San Raffaele Scientific Institute, Radiotherapy, Milano, Italy

Purpose/Objective:

Haematological toxicity after prostate cancer radiotherapy including pelvic lymph node irradiation may affect the post-treatment quality of life of patients. Voxel-based methods [1, 2] allow to analyse the 3D planned dose distribution on the organs at risk and identify spatial patterns associated with short-and-long-term side effects. The objective of this study was thus to perform a voxel-based analysis to identify bone subregions where the planned dose is predictive of haematological toxicity in the context of prostate cancer radiotherapy (RT).

Material/Methods:

A cohort of 100 patients treated in the same Institute with external beam RT including pelvic lymph node irradiation was available for this study. Patients were prospectively followed for a minimum period of 24 months. Blood cell counts were collected at baseline, at mid and end-RT, at 6 months and every 6 months up to 5 years. The 25th percentile of the variations of the absolute lymphocyte count (ALC) and the absolute neutrophil count (ANC) with respect to the baseline at nadir were scored as endpoints. Lumbar spine, pelvic bone and ilium were manually segmented on the pre-treatment CTs. The experimental design consisted of the following steps. First, patient anatomies were aligned in a common coordinate system by means of a non-rigid registration based on structural descriptions of the bones. The obtained transformation was then applied to the 3D planned dose distributions. In order to consider registration uncertainties and compensate outliers, dose maps were smoothed using a Gaussian filter of σ = 1.5. Once 3D doses were spatially normalised, for the two haematological toxicity endpoints, a non-parametric permutation test was performed on every voxel of the bone region to derive a 3D map of statistical significance. Clusters of voxels with a p-value < 0.01 were identified. These anatomical subregions were then propagated back to each patient native space by means of a non-rigid registration also based on structural descriptions of the bones.