[Paper Review] RepVGG-GELAN: Enhanced GELAN with VGG-STYLE ConvNets for Brain Tumour Detection

Thennarasi Balakrishnan, Sandeep Singh Sengar|arXiv (Cornell University)|May 6, 2024

Brain Tumor Detection and Classification4 citations

TL;DR

This paper proposes RepVGG-GELAN, a YOLO-based object detection model that integrates VGG-style reparameterized convolutions (RepVGG) with a Generalized Efficient Layer Aggregation Network (GELAN) for improved accuracy and efficiency in brain tumour detection. The model achieves 98.2% precision and 97.0% AP50 on a medical imaging dataset, outperforming RCS-YOLO and YOLOv8 while operating at 240.7 GFLOPs with only 25.4 million parameters.

ABSTRACT

Object detection algorithms particularly those based on YOLO have demonstrated remarkable efficiency in balancing speed and accuracy. However, their application in brain tumour detection remains underexplored. This study proposes RepVGG-GELAN, a novel YOLO architecture enhanced with RepVGG, a reparameterized convolutional approach for object detection tasks particularly focusing on brain tumour detection within medical images. RepVGG-GELAN leverages the RepVGG architecture to improve both speed and accuracy in detecting brain tumours. Integrating RepVGG into the YOLO framework aims to achieve a balance between computational efficiency and detection performance. This study includes a spatial pyramid pooling-based Generalized Efficient Layer Aggregation Network (GELAN) architecture which further enhances the capability of RepVGG. Experimental evaluation conducted on a brain tumour dataset demonstrates the effectiveness of RepVGG-GELAN surpassing existing RCS-YOLO in terms of precision and speed. Specifically, RepVGG-GELAN achieves an increased precision of 4.91% and an increased AP50 of 2.54% over the latest existing approach while operating at 240.7 GFLOPs. The proposed RepVGG-GELAN with GELAN architecture presents promising results establishing itself as a state-of-the-art solution for accurate and efficient brain tumour detection in medical images. The implementation code is publicly available at https://github.com/ThensiB/RepVGG-GELAN.

Motivation & Objective

• To address the gap in applying efficient YOLO-based object detectors to medical image analysis, particularly for brain tumour detection.
• To improve detection accuracy and computational efficiency in brain tumour detection using deep learning.
• To integrate RepVGG’s reparameterized convolutional architecture with GELAN for enhanced feature extraction and inference speed.
• To develop a lightweight, high-performance model suitable for deployment in clinical settings with limited computational resources.

Proposed method

• Adopts a YOLO framework enhanced with RepVGG, a reparameterized convolutional architecture that merges multiple convolutional paths into a single equivalent convolution during inference.
• Integrates the Generalized Efficient Layer Aggregation Network (GELAN) to improve feature fusion and hierarchical representation learning across multi-scale feature maps.
• Employs spatial pyramid pooling (SPP) within the GELAN module to enhance receptive field and robustness to scale variations.
• Uses a single-stage detection head with anchor-free prediction, applying sigmoid activation to class scores and regression heads for bounding box coordinates.
• Applies structural reparameterization to convert residual-like blocks into a single convolutional layer at inference time, reducing FLOPs and accelerating inference.
• Optimizes the model with a simplified down-sampling module and an optimized prediction head, eliminating the need for anchor boxes.

Experimental results

Research questions

• RQ1 Can the integration of RepVGG with GELAN improve detection accuracy and speed in brain tumour detection compared to existing YOLO-based models?
• RQ2 How does the RepVGG-GELAN architecture perform in terms of precision, recall, and mAP across different IoU thresholds on medical imaging data?
• RQ3 To what extent does the use of reparameterized convolutions reduce computational cost without sacrificing detection performance?
• RQ4 How does the model’s parameter count and FLOPs compare to state-of-the-art models like RCS-YOLO and YOLOv8 in the context of brain tumour detection?
• RQ5 What is the contribution of the RepVGG component to the overall performance gain in the RepVGG-GELAN model?

Key findings

• RepVGG-GELAN achieves a precision of 98.2%, representing a 4.91% improvement over the baseline RCS-YOLO model.
• The model attains an AP50 score of 97.0%, a 2.54% increase over the latest existing approach, demonstrating superior localization accuracy at IoU=0.5.
• Despite a slight drop in recall (89.0%) compared to RCS-YOLO (94.5%), RepVGG-GELAN maintains strong detection performance with a high mAP50:95 of 72.3%.
• The model operates at 240.7 GFLOPs with only 25.4 million parameters, indicating high computational efficiency and suitability for real-time deployment.
• Ablation studies confirm that the inclusion of RepVGG boosts precision from 96.4% (GELAN-only) to 98.2%, validating its contribution to reducing false positives.
• The model outperforms both RCS-YOLO and YOLOv8 in precision and AP50 while maintaining a smaller parameter count and lower FLOPs, establishing it as a state-of-the-art solution for brain tumour detection.