Quantitative ultrasound augmented with machine learning to assess tissue microstructure during wound healing

Objective: Wound healing follows a complex process which is impaired in the case of chronic wounds. Therefore, there is a need to early identify healing deficiencies to adapt the treatments and reduce the significant burden of wound management on healthcare. Here, we propose an analysis method based on features extraction from B-mode ultrasound imaging, a non-invasive and clinically usable modality to quantitatively describe the wound healing process through its entire course. Methods: Fifteen C57BL/6 mice received a 1-cm full thickness wound on the dorsum and were imaged with a 15 MHz transducer during the first 3 weeks post-injury. Image features based on the texture and brightness of pixels were computed and compared to an automatic processing method. A machine learning algorithm was also applied to detect the healing status from images. Results: The histogram peak value from the surface region of interest (ROI) remained increased from day 8, while echogenicity of the underlying tissue was temporarily higher from day 2 to day 10. These two metrics respectively correlate with collagen and nuclei expressions from histology. An accuracy of 89.7 % was reached in determining the wound healing phases using a machine learning framework. Conclusion: Ultrasound imaging of open wounds provide time-dependent features which can predict the wound healing stage when coupled with machine learning. Significance: This study indicates that conventional ultrasound imaging can be a useful tool for an early identification of abnormalities during the wound healing process by processing B-mode images, which are accessible from any clinical ultrasound imaging system.