Generative approach for data augmentation for deep learning-based bone surface segmentation from ultrasound images.

Area of Science:

• Medical Imaging
• Deep Learning
• Surgical Navigation

Background:

• Precise localization of cystic bone lesions is vital for osteolytic bone tumor surgery.
• Ultrasound imaging is increasingly preferred over X-rays for intra-operative navigation due to its radiation-free and cost-effective nature.
• Accurate bone surface segmentation from ultrasound images is essential for reconstructing intra-operative bone models.

Purpose of the Study:

• To investigate the effectiveness of generative data augmentation for ultrasound bone surface segmentation.
• To identify standard data augmentation techniques for this application.
• To address the challenge of limited datasets in training deep learning models for bone segmentation.

Main Methods:

• Utilized a Pix2Pix image-to-image translation network for generative augmentation.
• Proposed a multiple-snapshot approach to mitigate uni-modal output limitations.
• Conducted experiments to identify and evaluate standard data augmentation methods.

Main Results:

• Generative augmentation improved accuracy in trained regions by +4.88% and untrained regions by +25.84% compared to standard augmentations alone.
• The generative approach, when added to standard methods, showed accuracy improvements of +8.74% in trained and +11.55% in untrained regions.
• Evaluated on 800 images from trained (humerus) and 1200 from untrained (tibia, femur) regions.

Conclusions:

• Generative approaches are highly beneficial for data augmentation in ultrasound bone segmentation, particularly with limited datasets.
• The proposed multiple-snapshot Pix2Pix method effectively generates multimodal images, significantly expanding dataset size.
• This technique holds potential for improving deep learning model performance in medical imaging applications.