Automatic detecting multiple bone metastases in breast cancer using deep learning based on low-resolution bone scan...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Oncology

Background:

• Whole-body bone scan (WBS) is crucial for diagnosing breast cancer bone metastases.
• Low resolution, small lesion size, and numerous lesions in WBS images pose challenges for deep learning detection.
• Existing deep learning models struggle with the unique characteristics of WBS for metastasis detection.

Purpose of the Study:

• To develop a unified deep learning framework for detecting multiple, densely distributed bone metastases in low-resolution WBS images.
• To enhance feature extraction capabilities for WBS images with low resolution and multiple small lesions.
• To improve the accuracy and efficiency of automated bone metastasis detection in breast cancer.

Main Methods:

• Proposed a novel unified detection framework incorporating a plug-and-play position auxiliary extraction module and a feature fusion module.
• Designed a self-attention transformer-based target detection head for accurate small metastasis identification.
• Retrospective study included 512 breast cancer patients; data split into 6:2:2 for training, validation, and testing.
• Evaluated against benchmarks like SSD, YOLOR, Faster_RCNN_R, and Scaled-YOLOv4 using Average Precision (AP) and Recall.

Main Results:

• The proposed method achieved an AP of 55.0 ± 6.4%, a significant improvement over the SSD baseline (9.8 ± 2%).
• Achieved an average recall of 54.3 ± 4.2%, substantially outperforming the SSD model (5.2 ± 12.7%).
• Ablation studies showed that adding the detection head and position auxiliary extraction modules increased AP by 14.03% and 19.3%, respectively.

Conclusions:

• The developed unified framework effectively addresses the challenges of detecting multiple small bone metastases in low-resolution WBS images.
• The method demonstrates superior performance and generalization capabilities on both private and public datasets.
• This framework represents a significant advancement in automated detection for breast cancer WBS, serving as a potential clinical decision support tool.