ИСТИНА |
Войти в систему Регистрация |
|
ИПМех РАН |
||
Background. Gliomas located near the motor cortex and pathways are quite common (S. Larjavaara2007). This type of tumors is still associated with a high rate of postoperative hemiparesis (temporary - 96% (Rossi2019), and persistent paresis - 47% (González-Darder2010). Postoperative deficit affects general survival (Rahman2017). Nowadays, there is no objective method for predicting motor worsening after surgery. Machine learning is supposed to be very perspective in this case. Objective. Development of convolution neuronal network (CNN) which can predict the postoperative motor worsening in patients with supratentorial gliomas using preoperative MR-data. Materials and methods. Preoperative MR-data (T1, T2, DWI, T2 FLAIR, T1+C) of 750 patients with supratentorial gliomas were collected from the Center's database. All included patients were surgically treated in the Center. Initially, four experienced doctors divided MR-data (T2 FLAIR) into slices with and without tumors. Data on pre- and postoperative motor status were taken from electronic medical records by using methods of the natural language text analysis. After that, all data were used for training CNN. Then the quality of CNN prediction was tested on 30 prospective patients with supratentorial gliomas. Results. At the test stage, CNN shows the accuracy 82%, sensitivity 87%, specificity 72% in predicting postoperative motor worsening (ROC AUC 82%, and F1 83%). Conclusions. The machine learning method allows us to predict the postoperative motor worsening in patients with supratentorial gliomas at the preoperative stage using MR-data. This provides objective information about the risk of developing neurological deficits, which is an essential aspect of informing patients at the preoperative stage. The research is funded by RFBR grant №19-29-01154 «Predicting of pyramidal symptoms and its reversibility in patients with supratentorial glial tumors located near the motor areas, using the knowledge transfer method and deep neural networks based on multifactor analysis of digital data of different modality»