The concept of a 'virtual patient': Transformation of multiple lung cancer patients to a common spatial configuration using non-rigid registration, a way to simplify large datasets and to generate new hypotheses

Janssens, Guillaume [UCL] Petit, Steven [Maastricht University Medical Center (MUMC), Netherlands] Orban de Xivry, Jonathan [UCL] Hanbeukers, Bianca [Maastricht University Medical Center (MUMC), Netherlands] Macq, Benoît [UCL] van Elmpt, Wouter [Maastricht University Medical Center (MUMC), Netherlands] De Ruysscher, Dirk [Maastricht University Medical Center (MUMC), Netherlands] Lambin , Philippe [Maastricht University Medical Center (MUMC), Netherlands] et al.[show all]

Purpose: To investigate the use of non-rigid registration for transforming and fusing data from multiple lung cancer patients into a common spatial reference ("Virtual Patient' ) in order to perform spatial-based statistics. This allows us to study the eflects ot radiotherapy by comparing the spatial distribution of data such as PET and dose distribution for multiple patients with or without a complication or a relapse. Furthermore, the reverse transformation can be applied on the reference contours in order to automatically segment the patient anatomy. Materials: CT imaging was performed for 6 lung cancer patients . One patient was taken as reference and the lungs and spinal cord were delineated. The 5 other patients were also delineated for validation purposes. An intensity based affine alignment followed by a log-domain phase-based non-rigid registration were applied on these data, producing for each patient a deformation field representing the transformation from the reference to the patient, and its inverse. For each patient, the tumor was delineated and ignored during the deformation field computation. The direct transformation was used to deform each patient CT towards the reference anatomy. while the inverse transformation was used to deform the reference contours towards the patient anatomy. These contours were compared to manual contours for each patient In order to validate the registration process. Results: CT, PET and dose distribution were successfully deformed towards the reference configuration ("Virtual Patient") using the deformation fields resulting from registration (see Figure 1). After registration, the DICE coefficient, used as a measurement of overlap between deformed reference contours and manual contours of the lungs was of 92 ± 3% (1SO). Conclusions: These preliminary results showed that non-rigid registration was able to match accurately the lung contours of the reference with the lung contours of the patients. The resulting transformations can then be used to deform the CT, PET and dose distribution accordingly, in order to transform the individual patient information into a common spatial configuration for performing spatial-based statistical analysis on two populations (e.g. with and without complications) inside two ("Virtual") reference patients summarizing respectively the PET images and dose distribution of patients with and without a complication. This approach allows to "simplify" large complex datasets or randomized trials and to generate new hypotheses.