Abstract
Gliomas represent 50% of primary brain tumors. Their least aggressive form, diffuse low-grade gliomas, classified as grade 1 and 2 by the World Health Organization (WHO), are characterized by a slow and continuous growth, preferentially along the tracts of the cerebral white matter. Low-grade gliomas always progress to an aggressive form, the high-grade glioma. With the desire to perform treatments and surgeries earlier and earlier, it becomes essential to be able to predict the progression of the tumor and its passage from a benign to a more severe state. MRI is a non-invasive, non-irradiating technique that has a higher sensitivity than other imaging techniques for the diagnosis of cerebral gliomas. MRI can assess anatomical aspects and exploring various functional parameters within a tumor. These include perfusion imaging which is capable of quantify cerebral blood volumes and flow rates. A lack of standardization in diagnostic methods has been identified, leading to variations in patient management and thus heterogeneity in prognosis of tumor growth. Studies have shown that the transition from low-grade to high-grade glioma can be predicted for a mean diameter growth rate greater than 8 mm/year. It has also been established that the cerebral blood volume at the tumor level can be an indicator of tumor growth and its next transformation. Indeed, a tumor development is always coupled to a chaotic development of vessels within it and thus to an abnormal increase of the cerebral blood volume at the level of the lesion. This vascular development can lead to a rupture of the blood-brain barrier and contrast is observed on T1-weighted MRI (hypersignal). This contrast enhancement is a sign of the low-grade tumor’s transformation to a high-grade glioma, it is called the anaplastic transformation. The observation of the T1 contrast enhancement is not specific of the moment of the malignant transformation. The multimodal approach in MRI (morphological and functional evaluation) provides a lot of diagnosis information. Their exploitation must be based on a set of information that can be obtained through the development of new automatic techniques for the processing of MRI datas. In the future, it will be necessary to define multimodal metrics (combining MRI hypersignal, perfusion, diameter, volume, etc.) in order to standardize diagnoses. Longitudinal evaluation is particularly important, since it is necessary in order to predict tumor growth.The final objective of this work is to be able to predict the anaplastic transformation or, at least, the negative evolution of diffuse low-grade glioma during its evolution.To achieve this goal, the first step was to set up an automatic segmentation of the tumor volume. A convolutional neural network algorithm was trained on a large database of low-grade gliomas. This database was collected in the neuroradiology department of the CHU Gui de Chauliac and the tumor areas were manually traced on each examination by neuroradiologists. The training of this algorithm was done exclusively on longitudinal follow-ups of low-grade gliomas because it was necessary to be able to segment the lesions on clinical routine (glioma with surgical interventions, different MRI acquisition machines...).To predict the evolution, it was necessary to characterize the tumor lesion during its longitudinal follow-up. This characterization was possible by calculating quantitative morphological parameters (volume, average diameter of the tumor ...)