# Efficient attention-based Ghost-ResNet for brain tumor classification in magnetic resonance imaging (MRI)

## TL;DR

This paper introduces a lightweight deep learning model for brain tumor classification in MRI scans that achieves high accuracy while using fewer computational resources.

## Contribution

The novel contribution is an efficient deep learning architecture combining Ghost modules and Efficient Channel Attention (ECA) blocks for brain tumor classification.

## Key findings

- The proposed model achieved 97.85% overall classification accuracy on a brain tumor MRI dataset.
- It outperformed the DenseNet121 baseline by 1.65% in accuracy while maintaining a low false-positive rate.
- The model demonstrates high precision, recall, and specificity exceeding 97.8% for all tumor categories.

## Abstract

Brain tumor classification from magnetic resonance imaging remains a challenging task in medical image analysis, particularly when high diagnostic performance must be achieved under limited computational resources. Effective models are therefore required to balance classification accuracy with efficiency to support practical clinical deployment.

This study addresses this challenge by proposing an efficiency-oriented deep learning architecture that integrates Ghost modules into a ResNet-50 backbone and enhances feature learning through Efficient Channel Attention (ECA) blocks. The proposed design aims to improve discriminative capability while reducing feature redundancy and computational overhead.

The model was evaluated on the Bangladesh Brain Cancer MRI Dataset, which contains 6,056 MRI images representing three tumor categories: glioma, meningioma, and pituitary tumors. Preprocessing included contrast normalization using Contrast Limited Adaptive Histogram Equalization (CLAHE). Data augmentation was selectively applied to improve generalization while avoiding excessive artificial amplification of feature representations.

Experimental results demonstrate the effectiveness of the proposed attention-assisted lightweight architecture. The model achieved an overall classification accuracy of 97.85%, while macro-averaged precision, recall (sensitivity), and specificity all exceeded 97.8% (as defined in the Methods section). This corresponds to a 1.65% absolute improvement in accuracy compared with the strongest baseline model, DenseNet121, while maintaining a low false-positive rate. These findings suggest that competitive performance can be achieved without increasing architectural complexity.

The results highlight the potential of pursuing efficiency-driven architectural designs as an alternative to increasingly complex deep learning models. In particular, channel-attention-assisted feature generation appears to preserve high diagnostic accuracy while reducing representational and computational overhead, supporting its suitability for resource-constrained medical imaging applications.