Automatic tissue characterization from optical coherence tomography images for smart laser osteotomy

Fascinating experiments have proved that in the very near future, laser will completely replace mechanical tools in bone surgery or osteotomy. Laser osteotomy overcomes mechanical tools’ shortcomings, with less damage to surrounding tissue, lower risk of viral and bacterial infections, and faster wound healing. Furthermore, the current development of artificial intelligence has pushed the direction of research toward smart laser osteotomy. This thesis project aimed to advance smart laser osteotomy by introducing an image-based automatic tissue characterization or feedback system. The Optical Coherence Tomography (OCT) imaging system was selected because it could provide a high-resolution subsurface image slice over the laser ablation site.
Experiments were conducted and published to show the feasibility of the feedback system. In the first part of this thesis project, a deep-learning-based OCT image denoising method was demonstrated and yielded a faster processing time than classical denoising methods, while maintaining image quality comparable to a frame-averaged image. Next part, it was necessary to find the best deep-learning model for tissue type identification in the absence of laser ablation. The results showed that the DenseNet model is sufficient for detecting tissue types based on the OCT image patch. The model could differentiate five different tissue types (bone, bone marrow, fat, muscle, and skin tissues) with an accuracy of 94.85 %. The last part of this thesis project presents the result of applying the deep-learning-based OCT-guided laser osteotomy in real-time. The first trial experiment took place at the time of the writing of this thesis. The feedback system was evaluated based on its ability to stop bone cutting when bone marrow was detected. The results show that the deep-learning-based setup successfully stopped the ablation laser when bone marrow was detected. The average maximum depth of bone marrow perforation was only 216 μm.
This thesis project provides the basic framework for OCT-based smart laser osteotomy. It also shows that deep learning is a robust approach to achieving real-time application of OCT-guided laser osteotomy. Nevertheless, future research directions, such as a combination of depth control and tissue classification setup, and optimization of the ablation strategy, would make the use of OCT in laser osteotomy even more feasible.