Laser Journal, Volume. 45, Issue 2, 152(2024)

Improved knowledge distillation transformer for medical imaging classification of new coronary pneumonia

BAI Haotian1... GU Yu1,*, YANG Lidong1, ZHANG Baohua1, LI Jianjun1, LYU Xiaoqi1,2, TANG Siyuan1,3, ZHANG Xiangsong1,4, JIA Chengyi1,5 and HE Qun1 |Show fewer author(s)
Author Affiliations
  • 1[in Chinese]
  • 2[in Chinese]
  • 3[in Chinese]
  • 4[in Chinese]
  • 5[in Chinese]
  • show less

    To address the problem of high false negative rate of RT - PCR in screening patients with COVID- 19 ,this paper proposes a DRPL-ViT computer - aided diagnostic network. The knowledge distillation mechanism is first introduced on the basis of Vision Transformer which enables the Transformer structure to be trained on small data sets to achieve better fitting re- sults. Then ,the dependencies between tokens can be better captured by encoding the position information of patches in a relative position encoding way that is more suitable for vision tasks. In order to focus on more local features a traditional convolution module is introduced in the Transformer Encoder module to extract local features. The experiments achieved an average classification accura- cy of 92. 11% on the four classification test sets and 97. 85% for COVID- 19. The experimental results indicate that the proposed network has a high accuracy in classifying neo - coronary pneumonia and other lung lesions ,and has some clinical application value.

    Tools

    Get Citation

    Copy Citation Text

    BAI Haotian, GU Yu, YANG Lidong, ZHANG Baohua, LI Jianjun, LYU Xiaoqi, TANG Siyuan, ZHANG Xiangsong, JIA Chengyi, HE Qun. Improved knowledge distillation transformer for medical imaging classification of new coronary pneumonia[J]. Laser Journal, 2024, 45(2): 152

    Download Citation

    EndNote(RIS)BibTexPlain Text
    Save article for my favorites
    Paper Information

    Category:

    Received: Jun. 11, 2023

    Accepted: --

    Published Online: Oct. 15, 2024

    The Author Email: Yu GU (guyu2010023@imust.edu.cn)

    DOI:10.14016/j.cnki.jgzz.2024.2.152

    Topics