Tissue classification using machine learning to aid in intraoperative registration: A pilot study

Modern handheld structured light scanners show potential in the filed of medical imaging and image-guided surgery. For the effective use of such scanners as a rapid registration tool, anatomical regions of interest must be identified. The purpose of this study is to investigate the use of machine learning to classify various anatomical tissues using the textural information collected from structured light scanners. We performed an ex vitro study using three fresh frozen knee specimens. Each specimen underwent multiple stages of dissection to reveal different anatomical tissues. At each stage of dissection, the specimens were scanned with a structured light scanner (Artec Spider, Artec Group, Palo Alto, USA). Using the texture information of the scanned model, a domain expert manually segmented four tissues of interest: muscle, tendon, cartilage, and bone. The RGB and HSL values of the data points in the manually segmented models were extracted for use in training and evaluating a random forest classifier. Our trained random forest classifier obtained a four-class classification accuracy of 77% and a three-class classification accuracy of 82%. The results of this study demonstrate the feasibility of a random forest to aid in semi-automatic or automatic segmentation of anatomical tissues using only textural information. Further experiments with in vivo tissues will need to be done to validate the application of such classifiers in an intraoperative environment.