Semi-automated machine learning approach to segment and register tissue oxygenation maps onto clinical images of wounds

摘要

Near-infrared (NIR) spectroscopic imaging of wounds has been performed by past researchers to obtain tissueoxygenation at discrete point locations. We had developed a near-infrared optical scanner (NIROS) that performs noncontactNIR spectroscopic (NIRS) imaging to provide 2D tissue oxygenation maps of the entire wounds. Regions ofchanged oxygenation have to be demarcated and registered with respect to visual white light images of the wound.Herein, a semi-automatic image segmentation and co-registration approach using machine learning has been developedto differentiate regions of changed tissue oxygenation. A registration technique was applied using a transformationmatrix approach using specific markers across the white light image and the NIR images (or tissue oxygenation maps).This allowed for physiological changes observed from hemodynamic changes to be observed in the RGB white lightimage as well. Semi-automated segmentation techniques employing graph cuts algorithms was implemented todemarcate the 2D tissue oxygenation maps depicting regions of increased or decreased oxygenation and further coregisteredonto the white light images. The developed registration technique was validated via phantom studies (both flatand curved phantoms) and in-vivo studies on controls, demonstrating an accuracy >97%. The technique was furtherimplemented on wounds (here, diabetic foot ulcers) across weeks of treatment. Regions of decreased oxygenation weredemarcated, and its area estimated and co-registered in comparison to the clinically demarcated wound area. Future workinvolves the development of automated machine learning approaches of image analysis for clinicians to obtain real-timeco-registered clinical and subclinical assessments of the wound.