Exploring the Impact of Noise and Image Quality on Deep Learning Performance in DXA Images

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Radiology

Background:

• Dual-Energy X-ray (DXA) imaging presents challenges for femur segmentation due to low contrast, noise, and anatomical variations.
• Accurate femur segmentation is crucial for reliable bone mineral density (BMD) assessment, essential for diagnosing conditions like osteoporosis.

Purpose of the Study:

• To investigate the impact of noise reduction techniques integrated with deep learning (DL) models on femur segmentation accuracy in DXA images.
• To evaluate the subsequent effect of improved segmentation on BMD calculation and overall diagnostic capability.

Main Methods:

• Convolutional neural network (CNN)-based models, specifically a fully convolutional neural network (FCNN), were developed and trained for femur segmentation.
• Various noise reduction filters were incorporated into the DL pipeline to assess their influence on segmentation performance and BMD accuracy.
• The FCNN approach was benchmarked against traditional noise reduction algorithms and manual segmentation.

Main Results:

• The FCNN model achieved a high segmentation accuracy of 98.84% for femurs in DXA images.
• BMD measurements derived from FCNN segmentation showed excellent correlation (0.9928) with established methods, indicating high precision.
• The FCNN demonstrated superior performance compared to standalone noise reduction algorithms and manual segmentation.

Conclusions:

• Integrating noise reduction with DL models, particularly FCNN, substantially enhances femur segmentation in DXA images.
• The FCNN approach offers a robust solution for improving BMD calculations, thereby aiding in the accurate clinical diagnosis of osteoporosis.
• This study underscores the potential of advanced DL techniques to overcome limitations in medical image analysis and improve patient diagnostics.