JoVE - Journal of Visualized Experiments
JoVE Visualize
Contact Us

An adaptive and lightweight YOLOv5 detection model for lung tumor in PET/CT images

  • 0School of Computer Science and Engineering, North Minzu University, Yinchuan, 750021, China.
Scientific Reports +

|

December 27, 2024

|

December 27, 2024

View abstract on PubMed

Summary

This summary is machine-generated.

A novel Cross Modal YOLOv5 (CMYOLOv5) model enhances lung tumor detection by fusing multi-modal PET and CT images. This approach significantly improves the accuracy and feature representation of irregular tumors compared to existing methods.

Area Of Science

  • Medical Imaging
  • Artificial Intelligence
  • Oncology

Background

  • Multi-modal medical images are crucial for accurate tumor lesion detection.
  • Current models often rely on single imaging modalities, limiting the comprehensive understanding of lesion characteristics like shape, size, and contrast.
  • There's a need for models that can effectively integrate semantic information from multiple sources.

Purpose Of The Study

  • To propose a Cross Modal YOLOv5 (CMYOLOv5) model for improved multi-modal lung tumor detection.
  • To enhance the fusion of complementary information from PET and CT scans.
  • To improve the feature representation capabilities for irregular lung tumors.

Main Methods

  • Developed a dual-branch auxiliary network for PET and CT semantic extraction.
  • Implemented Cross-modal Features Fusion (CFF) and Self-Adaptive Attention Fusion (AAF) for integrating multi-modal data.
  • Utilized a Self-Adaptive Transformer (SAT) with deformable attention in the feature enhancement neck.
  • Incorporated Reparameter Residual Blocks (RRB) and Reparameter Convolution (RC) for richer feature learning.

Main Results

  • CMYOLOv5 achieved high performance metrics: 97.16% Precision, 96.41% Recall, 97.18% mAP, and 96.78% F1-score.
  • The model demonstrated superior detection accuracy for irregular lung tumors.
  • Achieved a processing speed (FPS) of 96.37 with a training time of 3912 seconds.

Conclusions

  • The proposed CMYOLOv5 model effectively leverages multi-modal PET/CT data for superior lung tumor detection.
  • The integration of advanced fusion techniques and Transformer architecture enhances the model's ability to capture complex lesion features.
  • CMYOLOv5 outperforms existing advanced methods in detecting irregular lung tumors.

Keywords: