Enhancing YOLOv11 with Large Kernel Attention and Multi-Scale Fusion for Accurate Small and Multi-Lesion Bone Tumor...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Oncology

Background:

• Primary bone tumors like osteosarcoma and chondrosarcoma are rare, aggressive cancers requiring prompt diagnosis.
• X-ray radiography is accessible but struggles with detecting small or multifocal bone lesions.
• Current deep learning models lack generalizability due to small, single-center datasets.

Purpose of the Study:

• To develop an advanced deep learning model for accurate primary bone tumor detection in radiographs.
• To enhance the detection capabilities for small and multifocal lesions, addressing limitations of existing methods.

Main Methods:

• Introduction of YOLOv11-MTB, an enhanced YOLOv11 model incorporating multi-scale Transformer-based attention, boundary-aware feature fusion, and receptive field augmentation.
• Training and evaluation on two multi-center datasets, including the BTXRD dataset with annotated bone lesion data.
• Focus on improving detection of small and multi-focal lesions.

Main Results:

• YOLOv11-MTB achieved state-of-the-art performance in bone tumor detection.
• The model attained a mean average precision (mAP) of 79.6% on the BTXRD dataset.
• Clinically relevant results include mAP of 55.8% for small lesions and 63.2% for multiple lesions.

Conclusions:

• The YOLOv11-MTB framework demonstrates strong generalization and accuracy for primary bone tumor detection.
• The model's effectiveness in identifying small and multiple lesions indicates significant potential for clinical application in radiology.