Abstract: In this paper, we study the method of automatic recognition of acute ischemic stroke using non-contrast computed tomography images of the brain. We focus on the application of deep neural network model based on U-Net architecture and its modifications using the Attention mechanism. This approach makes it possible to achieve a high accuracy in the segmentation of the brain affected areas. However, such models have large demands on the computing power of devices, which complicates their practical use on computer tomography workstations. The purpose of this work is to study the application of compression methods for existing deep neural network models. In particular, the int8 quantization method is used. This method allows one to switch from floating-point calculations (float32) to integer calculations (int8), which provides a more compact representation of the model and the use of high-performance vectorized operations on many hardware platforms. The pruning is also used. It is shown that the compressed models (up to 70% pruning ratio) are still able to predict the affected areas of the brain with a sufficiently high degree of accuracy. © 2022 IEEE.

Cite: Mikhailapov D. , Tulupov A. , Alyamkin S. , Berikov V.
Compression of Deep Neural Network for Acute Ischemic Stroke Segmentation
In compilation 2022 Ural-Siberian Conference on Computational Technologies in Cognitive Science, Genomics and Biomedicine (CSGB), 7-8 July, Novosibirsk, Russia. 2022. – C.240-245. DOI: 10.1109/CSGB56354.2022.9865656 Scopus OpenAlex

Identifiers:

Scopus:	2-s2.0-85138471478
OpenAlex:	W4293812556

Citing: Пока нет цитирований

Altmetrics: