Context specific descriptors for tracking deforming tissue

In minimally invasive surgery, deployment of motion compensation, dynamic active constraints and adaptive intra-operative guidance require accurate estimation of deforming tissue in 3D. To this end, the use of vision-based techniques is advantageous in that it does not require the integration of additional hardware to the existing surgical settings. Deformation can be recovered by tracking features on the surface of the tissue. Existing methods are mostly based on ad hoc machine vision techniques that have generally been developed for rigid scenes or use pre-determined models with parameters fine tuned to specific surgical settings. In this work, we propose a novel tracking technique based on a context specific feature descriptor. The descriptor can adapt to its surroundings and identify the most discriminate information for tracking. The feature descriptor is represented as a decision tree and the tracking process is formulated as a classification problem for which log likelihood ratios are used to improve classifier training. A recursive tracking algorithm obtains examples of tissue deformation used to train the classifier. Additional training data is generated by geometric and appearance modelling. Experimental results have shown that the proposed context specific descriptor is robust to drift, occlusion, and changes in orientation and scale. The performance of the algorithm is compared with existing tracking algorithms and validated with both simulated and in vivo datasets.

An adaptive intra-operative tissue deformation tracking framework based on context specific descriptors incorporating both geometric and appearance information that is robust to drift, occlusion, artefacts and changes in orientation and scale.Figure optionsDownload high-quality image (56 K)Download as PowerPoint slideHighlights
► A tracking framework is proposed which uses context specific information.
► Tissue tracking is posed as a classification problem.
► Training data is generated online and with geometric and appearance models.
► Classifier training is improved using log likelihood ratios.
► It outperforms four existing trackers and is robust to deformation, rotation, scale and smoke.