# StarMA Net: A star-shape multi-scale attention network for medical imaging classification

## TL;DR

The paper introduces StarMA Net, a new neural network for medical image classification that improves accuracy by better capturing spatial and channel relationships.

## Contribution

StarMA Net introduces a novel star-shaped multi-scale attention mechanism to enhance feature representation and inter-class discrimination in medical imaging.

## Key findings

- StarMA Net outperforms existing methods in classification accuracy and robustness across five medical imaging datasets.
- The proposed attention mechanism improves the model's ability to handle intra-class variability and complex structures.
- Cross-spatial aggregation enhances the integration of multi-scale contextual information.

## Abstract

Medical image classification is crucial for clinical diagnosis. However, medical datasets often face challenges such as limited characterization capabilities, difficulties in category differentiation, and the presence of individual differences. Although attention mechanisms can enhance feature representation, existing methods often struggle to utilize spatial information effectively and lack modeling of inter-channel interactions. We propose star-shaped multi-scale attention (StarMA). (1) It retains spatial information along different orientations within each channel through axial decomposition, enhancing the model’s perception of complex structures. (2) A star-shaped structure is employed to project feature maps into a high-dimensional nonlinear space, strengthening inter-channel interactions and improving inter-class discriminability. (3) A cross-spatial aggregation learning strategy is introduced to integrate multi-scale contextual information, improving the model’s ability to handle intra-class variability. Based on StarMA, we developed StarMA Net and conducted comparative experiments on five different datasets. Compared to advanced algorithms, StarMA Net demonstrates better effectiveness and robustness.

•The attention module can generate more distinguishable feature representations•Integrating multi-scale information to enhance robustness to intra-class variations•Enhancement of channel modeling for improved inter-class discriminative capability•Propose a visual backbone based on the attention to medical image datasets

The attention module can generate more distinguishable feature representations

Integrating multi-scale information to enhance robustness to intra-class variations

Enhancement of channel modeling for improved inter-class discriminative capability

Propose a visual backbone based on the attention to medical image datasets

Medical imaging; Bioinformatics; Computer modeling