研究目的
To develop an automatic hip geometric features (HGFs) extraction technique from femur DXA images using regional random forest to improve hip fracture risk assessment.
研究成果
The proposed automatic HGFs extraction technique using regional random forest is effective, achieving high accuracies of 96.22% for phantom data and 95.87% for real human data. It reduces workload and improves the use of X-ray devices, with potential applications in hip fracture research and risk assessment guidelines. Future work may involve deep learning approaches for enhanced performance.
研究不足
The approach requires optimal supervised selection of features and may need manual correction for bad contours in some cases (4 out of 200 images). It relies on a specific data set and may not generalize without further validation.
1：Experimental Design and Method Selection:
The study uses a regional random forest (RRF) classifier to localize anatomical landmarks (ALMs) and extract HGFs from femur DXA images. The methodology includes image denoising with a non-local means filter (NLMF), hip segmentation using pixel label random forest (PLRF), contour extraction, hip anatomy localization, and RRF-based landmark detection.
2：Sample Selection and Data Sources:
Data includes phantom images from a femur aluminum bar phantom (Lunar 1043) and real human femur DXA images scanned with a DXA imaging system (OsteoPro MAX). A total of 400 images were used, with 250 for training and 150 for testing.
3：List of Experimental Equipment and Materials:
DXA imaging system (OsteoPro MAX, B. M. Tech Worldwide Co., Ltd), femur phantom (Lunar 1043), digital caliper (Digital Caliper Deluxe Model), and software like ImageJ for ground truth generation.
4：Experimental Procedures and Operational Workflow:
The workflow involves image acquisition, denoising, segmentation, contour processing, ROI selection, landmark localization using RRF, and HGFs extraction. Steps include generating BMD maps, applying PLRF for segmentation, using contour features for anatomy localization, and implementing RRF with hybrid features.
5：Data Analysis Methods:
Performance is evaluated by comparing extracted HGFs to ground truths using absolute differences, accuracy calculations, Jaccard index for anatomy localization, and receiver operating characteristic (ROC) curves. Statistical analysis includes mean and standard deviation calculations.
独家科研数据包，助您复现前沿成果，加速创新突破
获取完整内容
