[Paper Review] BGF-YOLO: Enhanced YOLOv8 with Multiscale Attentional Feature Fusion for Brain Tumor Detection

Ming Kang, Chee‐Ming Ting+2 more|arXiv (Cornell University)|Sep 22, 2023

Brain Tumor Detection and Classification7 citations

TL;DR

BGF-YOLO augments YOLOv8 with Bi-level routing attention, generalized feature pyramids, and a fourth detection head to improve brain tumor detection, achieving a 4.7% absolute mAP50 gain over YOLOv8x on Br35H.

ABSTRACT

You Only Look Once (YOLO)-based object detectors have shown remarkable accuracy for automated brain tumor detection. In this paper, we develop a novel BGF-YOLO architecture by incorporating Bi-level routing attention, Generalized feature pyramid networks, and Fourth detecting head into YOLOv8. BGF-YOLO contains an attention mechanism to focus more on important features, and feature pyramid networks to enrich feature representation by merging high-level semantic features with spatial details. Furthermore, we investigate the effect of different attention mechanisms and feature fusions, detection head architectures on brain tumor detection accuracy. Experimental results show that BGF-YOLO gives a 4.7% absolute increase of mAP$_{50}$ compared to YOLOv8x, and achieves state-of-the-art on the brain tumor detection dataset Br35H. The code is available at https://github.com/mkang315/BGF-YOLO.

Motivation & Objective

• Motivate improved automatic brain tumor detection using advanced YOLO architectures.
• Investigate the impact of attention mechanisms and feature fusion on detection accuracy.
• Evaluate the benefits of a redesigned detection head within the YOLOv8 framework.

Proposed method

• Incorporates Bi-level routing attention to emphasize important features.
• Uses Generalized feature pyramid networks to merge high-level semantics with spatial details.
• Introduces a Fourth detecting head to augment detection capability.
• Evaluates different attention mechanisms and feature fusion strategies within YOLOv8-based architecture.
• Assesses detection performance on brain tumor datasets, notably Br35H.

Experimental results

Research questions

• RQ1 Does Bi-level routing attention improve brain tumor detection accuracy in YOLOv8-based architectures?
• RQ2 How do generalized feature pyramid networks affect feature representation for brain tumor detection?
• RQ3 What is the impact of adding a fourth detection head on detection performance?
• RQ4 Which attention mechanisms and feature fusion schemes yield the best mAP on Br35H?

Key findings

• BG F-YOLO achieves a 4.7% absolute increase of mAP50 over YOLOv8x.
• The architecture attains state-of-the-art performance on the Br35H brain tumor dataset.
• The study analyzes the effects of different attention mechanisms and feature fusions on detection accuracy.