An exploration into novel methods for real-time segmentation and tracking during computer-robotic assisted surgery
File(s)
Author(s)
Laws, Stephen
Type
Thesis
Abstract
One limitation of current computer aided surgical systems is their ability to localise themselves with respect to a patient's anatomy. A comprehensive review of technologies for non-invasive, intraoperative, system-integratable, tissue segmentation and tracking techniques was undertaken. The resultant review highlighted that computation time is of importance but the current research priority is classification accuracy.
Diffuse reflectance imaging based techniques demonstrated a balance between classification accuracies and computational time. A novel diffuse laser speckle reflectance imaging modality was discovered and investigated. Preliminary experiments validated that a red laser speckle can accurately classify cadaveric tissues using a convolutional neural network. A follow-up study demonstrated that a near-infrared laser was capable of classification accuracies over 90% when used to identify four cadaveric tissues.
These results influenced the design of a bespoke stereo imaging platform. Stereo matching was achieved using a semi-global block matching and pyramid stereo matching networks with root-mean-square accuracies of 0.96 and 1.04 mm respectively. Furthermore, the sensor system had sufficient resolution to, theoretically, classify a multitude of spots simultaneously.
To validate the sensor's classification capabilities the sensor system was incorporated into a clinically relevant surgical pipeline; the registration process for computer-assisted knee arthroplasty. The resultant classification algorithm robustly identified the distal femur in all test images. The cadaveric classification network achieved base accuracies of 80.1%, rising to 93.3% when the prediction threshold was increased to 85%. The sensor system could segment and reconstruct a depth scene in 0.4 s.
This thesis outlines a candidate new technology for simultaneous tissue classification and segmentation. Future research into this technology to further improve results could focus on; increased number of material types tested, software optimisations for improved computation time, incorporation of structured light techniques and extended wavelength studies.
Diffuse reflectance imaging based techniques demonstrated a balance between classification accuracies and computational time. A novel diffuse laser speckle reflectance imaging modality was discovered and investigated. Preliminary experiments validated that a red laser speckle can accurately classify cadaveric tissues using a convolutional neural network. A follow-up study demonstrated that a near-infrared laser was capable of classification accuracies over 90% when used to identify four cadaveric tissues.
These results influenced the design of a bespoke stereo imaging platform. Stereo matching was achieved using a semi-global block matching and pyramid stereo matching networks with root-mean-square accuracies of 0.96 and 1.04 mm respectively. Furthermore, the sensor system had sufficient resolution to, theoretically, classify a multitude of spots simultaneously.
To validate the sensor's classification capabilities the sensor system was incorporated into a clinically relevant surgical pipeline; the registration process for computer-assisted knee arthroplasty. The resultant classification algorithm robustly identified the distal femur in all test images. The cadaveric classification network achieved base accuracies of 80.1%, rising to 93.3% when the prediction threshold was increased to 85%. The sensor system could segment and reconstruct a depth scene in 0.4 s.
This thesis outlines a candidate new technology for simultaneous tissue classification and segmentation. Future research into this technology to further improve results could focus on; increased number of material types tested, software optimisations for improved computation time, incorporation of structured light techniques and extended wavelength studies.
Version
Open Access
Date Issued
2023-03-14
Date Awarded
01/01/2024
License URL
Advisor
Rodriguez y Baena, Ferdinando
Davies, Brian
Sponsor
Engineering and Physical Sciences Research Council Great Britain
Grant Number
EP/R513052/1
Publisher Department
Mechanical Engineering
Publisher Institution
Imperial College London
Qualification Level
Doctoral
Qualification Name
Doctor of Philosophy (PhD)