GAM-SpCaNet: Gradient awareness minimization-based spinal convolution attention network for brain tumor classification

Chaosheng Tang, Bin Li, Junding Sun*, Shui Hua Wang, Yu Dong Zhang

*Corresponding author for this work

Research output: Contribution to journalArticlepeer-review

9 Citations (Scopus)


Brain tumor is one of the common diseases of the central nervous system, with high morbidity and mortality. Due to the wide range of brain tumor types and pathological types, the same type is divided into different subgrades. The imaging manifestations are complex, making clinical diagnosis and treatment difficult. In this paper, we construct SpCaNet (Spinal Convolution Attention Network) to effectively utilize the pathological features of brain tumors, consisting of a Positional Attention (PA) convolution block, Relative self-attention transformer block, and Intermittent fully connected (IFC) layer. Our method is more lightweight and efficient in recognition of brain tumors. Compared with the SOTA model, the number of parameters is reduced by more than three times. In addition, we propose the gradient awareness minimization (GAM) algorithm to solve the problem of insufficient generalization ability of the traditional Stochastic Gradient Descent (SGD) method and use it to train the SpCaNet model. Compared with SGD, GAM achieves better classification performance. According to the experimental results, our method has achieved the highest accuracy of 99.28%, and the proposed method performs well in classifying brain tumors.

Original languageEnglish
Pages (from-to)560-575
Number of pages16
JournalJournal of King Saud University - Computer and Information Sciences
Issue number2
Publication statusPublished - Feb 2023
Externally publishedYes


  • Brain tumor classification
  • Gradient awareness minimization
  • Intermittent fully connected layer
  • Positional attention convolution block
  • Relative self-attention transformer block


Dive into the research topics of 'GAM-SpCaNet: Gradient awareness minimization-based spinal convolution attention network for brain tumor classification'. Together they form a unique fingerprint.

Cite this