OS-DETR: End-to-end brain tumor detection framework based on orthogonal channel shuffle networks

Kaixin Deng ¹, Quan Wen ¹, Fan Yang ¹

⁰College of Computer Science and Cyber Security, Chengdu University of Technology, Chengdu, China.

Plos One +

|

May 13, 2025

View abstract on PubMed

Summary

This study introduces the Orthogonal Channel Shuffle Network (OSNet) integrated into the Detection Transformer (DETR) for improved brain tumor detection. The OS-DETR model enhances accuracy and reduces overfitting in medical imaging tasks.

Area Of Science

Artificial Intelligence
Medical Imaging
Computer Vision

Background

OrthoNets ensure filter orthogonality but lack internal filter constraints.
Overfitting is a risk in medical image analysis with limited data.
Existing object detection frameworks may not be optimal for precise brain tumor identification.

Purpose Of The Study

To introduce the Orthogonal Channel Shuffle Network (OSNet) for enhanced internal filter orthogonality.
To integrate OSNet into the Detection Transformer (DETR) for brain tumor detection (OS-DETR).
To improve model performance through deformable attention and an advanced Intersection over Union strategy.

Main Methods

Developed the Orthogonal Channel Shuffle Network (OSNet) ensuring internal filter orthogonality.
Integrated OSNet into the Detection Transformer (DETR) framework, creating OS-DETR.
Incorporated deformable attention mechanisms and a specialized Intersection over Union strategy.

Main Results

OS-DETR demonstrated significant advantages over mainstream object detection frameworks on the Br35H dataset.
Achieved high performance metrics: 95.0% Precision, 94.2% Recall, 95.7% mAP@50, and 74.2% mAP@50:95.
The enhanced network effectively reduces overfitting in medical image analysis.

Conclusions

OS-DETR offers a superior approach for brain tumor detection compared to existing methods.
The proposed OSNet architecture enhances robustness and accuracy in medical object detection.
The study highlights the potential of internally orthogonal filters for improving deep learning models in specialized domains.