MARC SAGHIAH

AI tools for cross-sectional evaluation of post-RT brain tissues modifications from non-harmonized radiological protocols

No routine clinically-acquired post-operative and multimodal MR brain images of glioblastoma are freely available. This lack of annotated data is a major technical barrier, given that deep learning methods, that are the best AI technique for automated image segmentation in the medical domain and predictions of clinical endpoints, require large number of representative training samples to be efficient. Moreover, the heterogeneity of the voxel intensities of MR brain images that come from different non-harmonized radiological protocols makes it even more challenging to develop advanced spatio-temporal models and innovative AI tools for data processing. This lack of ground truth labels will prevent from exploring supervised machine learning techniques to extract spatio-temporal signatures of image-based biomarkers and characterize specific WMH. The inter-patients variability in terms of initiation and spatio-temporal progression of WMH will also make it potentially difficult to compare the extracted signatures from one subject to another and to build a meaningful population model.

Many sources of variations in MR images acquisition are inherent to the longitudinal follow-up of patients in clinical routine. My work consists of exploring advanced quality control and harmonization techniques to pool heterogeneous MRI data. This task aims at identifying one (or several) efficient generic method(s) to segment heterogeneous MR brain images acquired from several patients at different time points after surgery from non-harmonized radiology protocols. This work will be useful for the the development of a fully automated pipeline for the cross-sectional extraction and quantification of a panel of clinically relevant brain MR biomarkers for the follow-up of patients enduring brain surgery and chemoradiotherapy for glioblastoma.

Disciplines scientifiques