Application research on YOLOv5 model based on Lightweight Atrous Attention Module in brain tumor MRI image segmentation

TL;DR

This paper introduces a new lightweight attention module to improve brain tumor MRI image segmentation accuracy and efficiency using a modified YOLOv5 model.

Contribution

The novel Lightweight Atrous Attention Module (LAAM) improves segmentation performance while reducing computational costs.

Findings

The YOLOv5s-LAAM model achieved 92.3% precision and 90.4% recall in brain tumor MRI segmentation.

The model's GFLOPs were reduced by 15% compared to the original YOLOv5s-ASPP baseline.

LAAM integration improved mAP@50 score to 0.925, demonstrating enhanced segmentation accuracy.

Abstract

To enhance the segmentation accuracy and computational efficiency of brain tumor magnetic resonance imaging (MRI) images, this study proposes a novel Lightweight Atrous Attention Module (LAAM) that integrates the Convolutional Block Attention Module (CBAM) with an Atrous Spatial Pyramid Pooling (ASPP) structure. The LAAM was integrated into the YOLOv5s model to enhance its performance, aiming to boost accuracy and recall while keeping computational efficiency.

This study utilized two publicly available meningioma and glioma MRI datasets from Kaggle. The LAAM incorporates depthwise separable convolutions, dual attention mechanisms, and residual connections to reduce computational complexity while enhancing feature extraction capabilities. The modified YOLOv5s model was trained and validated via five-fold cross-validation, with performance comparisons conducted against the original YOLOv5s architecture and other optimized models.

The enhanced YOLOv5s-LAAM model demonstrated superior performance, achieving a precision of 92.3 %, a recall rate of 90.4 %, and an mAP@50 score of 0.925. Concurrently, the model exhibited significantly reduced computational demands, with the GFLOPs reduced by 15 % compared to the original YOLOv5s-ASPP baseline.

The integration of the LAAM significantly enhances the YOLOv5s model's segmentation capabilities for brain tumor MRI images, making it a valuable tool for clinical diagnosis and treatment planning. The lightweight design ensures effective deployment in resource-constrained environments while maintaining high computational performance.

Figures