ТRИЕТАПНИЙ ГП-ЗП АНСАМБЛЬ ЗГОRТКОВИХ НЕЙRОНИИХ МЕРЕЖ ДЛЯ СЕГМЕНТАЦЦ ЗЛОЯКІСННХ ПУХЛИН ГОЛОВНОГО МОЗКУ НА МРТ-ЗОБРАЖЕННЯХ.

In this paper, the problem of brain tumor binary semantic segmentation from MRI images is solved. The pixel-by-pixel determination of the anomaly region boundary is performed given the presence of noise in the training sample and input data. It is shown that in the case of using 2D models for solving 3D segmentation problems, spatial information between neighboring slices is not considered and not utilized. A new approach for optimizing the processing of 3D medical images using ensemble topologies in three stages is proposed. The first stage involves 2D ensemble processing of images in three dimensions to maximize the diversity criterion and accurately capture the region of interest (ROI). The second stage involves ensemble processing of 3D ROI regions extracted by neural networks with different 3D input block sizes to ensure diversity. In the third stage, the extracted abnormal regions (malignant tumors) from the first and second stages are aggregated by weighted summation and thresholding to obtain the final binary 3D mask of the brain tumor. The proposed approach was tested on the LGG Brain MRI Segmentation Dataset. It is shown that the segmentation accuracy is significantly improved in terms of dice score and mIoU, reducing the use of computationally expensive 3D networks. [ABSTRACT FROM AUTHOR]