[Machine learning for resting state fMRI-based preoperative mapping: comparison with task-based fMRI and direct cortical stimulation]. / Primenenie mashinnogo obucheniya dlya predoperatsionnogo kartirovaniya po dannym funktsional'noi magnitno-rezonansnoi tomografii pokoya: sravnenie so stimul-zavisimoi funktsional'noi magnitno-rezonansnoi tomografiei i pryamoi elektricheskoi stimulyatsiei kory.

ABSTRACT

OBJECTIVE: To develop a system for preoperative prediction of individual activations of motor and speech areas in patients with brain gliomas using resting state fMRI (rsfMRI), task-based fMRI (tb-fMRI), direct cortical stimulation and machine learning methods. MATERIAL AND METHODS: Thirty-three patients with gliomas (19 females and 14 males aged 19 - 540) underwent DCS-assisted resection of tumor (19 ones with lesion of motor zones and 14 patients with lesions of speech areas). Awake craniotomy was performed in 14 cases. Preoperative mapping was performed according to special MRI protocol (T1, tb-fMRI, rs-fMRI).Machine learning system was built on open source data from The Human Connectome Project. MR data of 200 healthy subjects from this database were used for system pre-training. Further, this system was trained on the data of our patients with gliomas. RESULTS: In DCS, we obtained 332 stimulations including 173 with positive response. According to comparison of functional activations between rs-fMRI and tb-fMRI, there were more positive DCS responses predicted by rs-fMRI (132 vs 112). Non-response stimulation sites (negative) prevailed in tb-fMRI activations (69 vs 44). CONCLUSION: The developed method with machine learning based on resting state fMRI showed greater sensitivity compared to classical task-based fMRI after verification with DCS 0.72 versus 0.66 (p<0.05) for identifying the speech zones and 0.79 versus 0.62 (p<0.05) for motor areas.