Automatic Segmentation of Acute Stroke Lesions Using Convolutional Neural Networks and Histograms of Oriented Gradients

Abstract

In this work, research was focused on the recognition of stroke lesions in FLAIR MRI images. The main tools used in the study were convolution neural networks (CNN) and histograms of oriented gradients (HOG). To train the neural network, 706 FLAIR MRI images of real patients were collected and labeled. During the testing of the program implementing the algorithm, it was found out that the histogram of oriented gradients method was ineffective in clarifying the edges of the lesion. To replace the HOG technology, we proposed a method on calculating the average pixel intensity in the lesion area. The result is a program that shows an average value of 0.554 by Dice score. The developed program can be used as a digital assistant for physicians who identify the lesions of stroke on MRI images and for training physicians in radiology and neurology. With the help of the developed technology it is possible to reduce the time of detection of the stroke lesions, reduce variability of results, increase their reproducibility and process a huge amount of data.