RESUMO
Gliomas are the most common neuroepithelial brain tumors. The modern classification of tumors of central nervous system and treatment approaches are based on tissue and molecular features of a particular neoplasm. Today, histological and molecular genetic typing of tumors can only be carried out through invasive procedures. In this regard, non-invasive preoperative diagnosis in neurooncology is appreclated. One of the perspective areas is artificial intelligence applied for neuroimaging to identify significant patterns associated with histological and molecular profiles of tumors and not obvlous for a specialist. OBJECTIVE: To evaluate diagnostic accuracy of deep learning methods for glioma typing according to the 2007 WHO classification based on preoperative magnetic resonance imaging (MRI) data. MATERIAL AND METHODS: The study included MR scans of patients with glial tumors undergoing neurosurgical treatment at the Burdenko National Medical Research Center for Neurosurgery. All patients underwent preoperative contrast-enhanced MRI. 2D and 3D MR scans were used for learning of artificial neural networks with two architectures (Resnest200e and DenseNet, respectively) in classifying tumors into 4 categories (WHO grades I-IV). Learning was provided on 80% of random examinations. Classification quality metrics were evaluated in other 20% of examinations (validation and test samples). RESULTS: Analysis included 707 contrast-enhanced T1 welghted images. 3D classification based on DenseNet model showed the best result in predicting WHO tumor grade (accuracy 83%, AUC 0.95). Other authors reported similar results for other methods. CONCLUSION: The first results of our study confirmed the fundamental possibility of grading axial contrast-enhanced T1 images according to the 2007 WHO classes using deep learning models.
