Preclinical Performance Evaluation of a Robotic Endoscope for Non-Contact Laser Surgery

Abstract

Despite great efforts, transoral robotic laser surgery has not been established clinically. Patient benefits are yet to be proven to accept shortcomings of robotic systems. In particular, laryngeal reachability and transition from microscope to accurate endoscopic laser ablation have not been achieved. We have addressed those challenges with a highly integrated robotic endoscope for non-contact endolaryngeal laser surgery. The current performance status has been assessed in multi-level user studies. In addition, the system was deployed to an ex vivo porcine larynx. The robotic design comprises an extensible continuum manipulator with multifunctional tip. The latter features laser optics, stereo vision, and illumination. Vision-based performance assessment is derived from depth estimation and scene tracking. Novices and experts (n = 20) conducted teleoperated delineation tasks to mimic laser ablation of delicate anatomy. Delineation with motion-compensated and raw endoscopic visualisation was carried out on planar and non-planar nominal patterns. Root mean square tracing errors of less than 0.75 mm were feasible with task completion times below 45 s. Relevant anatomy in the porcine larynx was exposed successfully. Accuracy and usability of the integrated platform bear potential for dexterous laser manipulation in clinical settings. Cadaver and in vivo animal studies may translate ex vivo findings.

Keywords: Ablation; Continuum robot; Endoscopy; Haptics; Head; Motion compensation; Neck; Performance; TORS; Tracking.

Figure 1

(a) Schematic setup of robotic laser-assisted lesion delineation on vocal fold soft tissue. The beam is displaced from the posterior (1) to the anterior (3) section of the vocal fold while following the vibratory edge (2). The intervention is monitored within the field of view (FoV) of the endoscopic camera. Teleoperation, 3D visualisation, and haptic feedback are taken into consideration for robot control.

Figure 2

(a) Subject operating the slave robotic with the haptic master device. The user interface (UI) enables stereoscopic image rendering through stereoscopic monitor or head-mounted display. Side-by-side images are for demonstration purposes only. (b) Actuated endoscopic distal tip attached to flexible continuum segments with integrated stereo vision, illumination, and fix-focus laser optics. (c) Slave robot mounted to frame with drive unit, actuation unit, and tip. The overlay image shows the master device with corresponding reference frame of the stylus. (d) Magnified view of the endoscopic tip facing a target with nominal ablation patterns.

Figure 3

Sequence diagrams of visual measurements: (a) Image processing for nominal path detection, segmentation, extraction, and mapping from image to 3D space. (b) Laser spot segmentation and spatial mapping for consecutive measurements of the spot location during execution of the user task. Acronyms are defined as follows: false (F) and true (T).

Figure 4

Evaluation workflow of the path tracing study: (a) Discrete representation and geometric annotation for PTE metric determination and (b) example of processed trajectory in x-y plane of (CF)_L. Coloured features in (a) indicate the nominal path (black), measured laser spot positions (blue), orthogonal path errors (red), and projected positions (green). Circle marker in (b) denote a sparse set, i.e. only one-fifth of total data is depicted for visualisation purposes.

Figure 5

Anatomy of the porcine larynx: (a) dorsal view, (b) close-up of the anatomical differences, (c) view into the inner larynx. (d) Experimental setup and deployment of the endoscopic tip into the inner larynx. The specimen resides in a bespoke frame. (e) Visualization and tracing of the vocal folds. Anatomy annotation: 1. epiglottis, 2. paired arytenoid, 3. paired interarytenoid cartilages, 4. vocal fold, and 5. intralaryngeal fat pad.

Figure 6

Example of evaluation workflow and study results for subject P18 and trial M10. (a) Temporal sequence of teleoperated tracing motion on S-shaped nominal pattern. (b) Post-experimental representation of automatically segmented nominal centreline (red) and measured laser spot positions (blue). (c) Depth map-based mapping of data in (b) to 3D space with reference to (CF)_L. (d) Corresponding path tracing errors over length of nominal path.

Figure 7

Box plots of metrics path tracing error (PTE) and task completion time (TCT) for non-stabilised (NS) and active scene stabilisation (AS): (a), (b) Condition-related results. (c), (d) Trial-resolved results. (e), (f) Condition-resolved split to subject groups. Horizontal dashes in the box depict the median and cross symbols indicate outliers. Boxes define 0.25 and 0.75-quantiles of input data. Upper and lower whisker span all data within 1.5 interquartile range of the nearer quartile.

Figure 8

Scores of post-experimental user survey for path tracing study: (a) Results of individual statements and (b) grouped to categories. (c) Results of between-subject factor split to subject groups for individual statements and (d) grouped to categories. Category labels indicate learning curve (LC), stress (ST), system performance (PF), and system design (DS). Labels marked in bold indicate negative statements. Statements are linked by † to ST, ‡ to PF, * to LC, and ⋆ to DS.

Figure 9

Example of 3D path delineation: (a) Specimen reconstruction with non-planar surface relief and coated path. (b) Temporal sequence of motion-compensated left camera view. (c) Post-experimental results of segmented nominal path (red) and projected spot measurements (blue). (d) Post-experimental spatial results of segmented nominal path (SP) and measured spot path (MP).

Figure 10

Results of 3D delineation case study: (a) Local path errors e_PT and focal errors e_FP along the nominal path with assistance. (b) Box plots of path tracing error (PTE) and focal position error (FPE) metrics for trials with visuo-haptic assistance (VHA) and no assistance (NA). Dashes in the box define the median, diamonds mark the mean, and cross markers indicate outliers. Boxes define 0.25 and 0.75-quantiles of input data. Upper and lower whisker span all data within 1.5 interquartile range of the nearer quartile.