Autonomous tracking and motion compensation of patients in 3D space
Robotically aligned OCT automatically tracks and compensates for patient motion in near real‑time utilizing tracking cameras combined with open and closed loop control of the robot tool end effector and electronically controlled opto-mechanics (see Methods) (12). Face-tracking cameras and software identified the location of the patient’s face and tracked the target eye of the patient. Using an open loop control system, these estimated coordinates were used to direct the robot to automatically align to the patient’s eye and compensate for bulk patient motion during imaging. Once grossly aligned, pupil-tracking cameras tracked the target ocular pupil of the patient. Using a closed loop control system, these residual position estimates were used to position the robot, sample arm 2D scanning mirror, and voice‑coil reference arm. This negated the requirement for patient stabilization, such as chin and forehead rests, during image acquisition.
Figure 1 demonstrates the benefits of active motion compensation through estimated motion profiles and corrected OCT volumes in two different patients: one with a healthy retina and the other with both a diseased retina and tremors due to Parkinson’s disease. Tracking profiles illustrate motion estimation in three dimensions: Temporal-Nasal (T-N, fast OCT dimension), Superior-Inferior (S-I, slowOCT dimension), and Anterior-Posterior (A-P, depth OCT dimension). For both patients we show concurrent face-tracking and residual pupil-tracking position estimation and three consecutive OCT volumes acquired over a total of 4.14 seconds. Face-tracking of a healthy retina patient (Fig. 1A, undialated, 28 year old female; see Methods, Study Design) indicates a preliminary motion estimation of the right eye over a range of (1.5, 0.7, 2.1) mm [T-N, S-I, A-P]. Pupil-tracking (Fig. 1B) estimated a position residual range of (1.5, 0.4, 0.9) mm. In contrast, for the patient with tremors due to Parkinson’s disease (Figs. 1G&H, 81 year old female), face-tracking estimated a range of (2.5, 1.0, 3.4) mm and pupil-tracking estimated a position residual of (1.5, 0.9, 1.8) mm for their left eye. Our face-tracking algorithm estimates the ocular globe location and does not track the ocular pupil location. Pupil-tracking residual values account for both eye motion that occurs during image acquisition and residual robotic position error due to open loop control.
For the healthy patient, Figs.1C-F show three RAOCT volumes and corresponding B-scans acquired during the motion profile time sequences shown in Figs. 1A&B. All three volumes provide a stable, clinically relevant field-of-view (FOV) with both the foveal pit and optic nerve head (ONH) visible. Retinal thickness (µm), here defined as the thickness between the inner limiting membrane (ILM; inner most layer of the retina before vitreous) and retinal pigment epithelium (RPE; hyper-reflective pigmented epithelial layer at outer boundary of neurosensory retina and before Bruch’s membrane and choroidal vasculature), was mapped to the top surface of the OCT volume. We automatically excluded the area surrounding the ONH as the RPE and the corresponding retinal tissue do not continue through the ONH and represents the ‘blind-spot’ in an individual’s vision. The foveal pit thickness in this patient measured 280 µm (mapped to blue) while thickening toward ONH to 450 µm (mapped to red) as the nerve fiber layer increased in thickness.
For the patient with diseased retina and tremors, Figs. 1I-L show three RAOCT volumes and corresponding B-scans acquired during the motion profile time sequences shown in Figs. 1G&H. All three volumes show a stable, clinically relevant FOV however unlike the healthy subject, the foveal pit is not readily visible due to macular edema (mapped to red) and the presence of a macular hole (mapped to blue; detail in inset). Additionally, volume three (B-scan in Fig. 1L) shows a small nasal shift in FOV likely due to uncorrected changes in patient gaze. However, while there was millimeter scale motion of the patient during OCT volume acquisition, two-tiered tracking and compensation allow for visualization of the 50 µm walls of the macular hole processes (Fig. I, inset).
Retinal imaging with robotically aligned optical coherence tomography
To quantitatively evaluate RAOCT, we imaged patients with healthy and diseased retinas recruited from clinics at the Duke Eye Center and compared the RAOCT measurements with clincially available, technician acquired OCT. We imaged each subject in triplicate with RAOCT and a clinical spectral-domain OCT system, OCT Spectralis from Heidelberg Engineering (see Methods, Study Design). For each acquired RAOCT volume we generated a retinal thickness map (from ILM to RPE) which matches the tissue boundaries utilized by Spectralis to generate their thickness maps (15). Representative thickness maps from both devices in healthy and diseased retinas can be seen in Fig. 2. It should be noted that, RAOCT thickness maps were overlaid on summed volume projections from the RAOCT data utilized to generate the thickness maps while the Spectralis thickness maps were overlaid on separately acquired, slightly larger FOV scanning laser ophthalmoscopy (SLO) images. Autonomous robotic alignment and compensation allows for the patient to sit without chin or forehead rests as seen in Fig. 2A. Taking advantage of active compensation, we acquired, registered, and averaged up to 100 repeated B-scans at a single location enabling high resolution retinal images. A healthy retina, as seen in Fig. 2D, shows the fovea (the region of retina responisble for our highest acuity vision) and vasculature surrounding the optic nerve head. This particular B-scan was acquired during the photograph shown in Fig. 2A. When imaging diseased retina, fine details such as walls of cystoid macular edema (Fig. 2G) and the boundary of the posterior hyaloid membrane (Fig. 2M) can be seen.
To quantitatively compare the two devices, we calculated the mean retinal thickness in a 1 mm diameter area centered at the fovea for each acquired volume. This metric is important because changes in foveal thickness, and macular thickness overall, are relevent biomarkers for diabetic retinopathy (16, 17) and macular degeneration progression(18–20); specifically, retinal pathologies can cause retinal thickness increases (e.g., from edema) or decreases (e.g., from atrophy or degeneration). We imaged each eye in triplicate to investigate intra-session variation. In Table 1, we report mean foveal thickness across the three volumes and intra-session variation. Several patients with diseased retinas had one of their eyes excluded from the study due to comorbidities (i.e. cataract, tremors, enuculation) that prevented the technician from acquiring volumes with clinical OCT. For RAOCT, the mean foveal thicknessacross all eyes was 297.8 µm with a mean intra-session variation of ±1.5 µm and with a total population variation of ±61.7 µm. We then split the population into healthy and diseased retinas. For healthy retinas, the mean thickness was 282.9 µm, intra-session variation was ±0.8 µm, and with a total population variation of ±21.7µm. For diseased retinas, the mean thickness was 310.3 µm, intra-session variation was ±2.0 µm, and with a total population variation of ±79.1 µm. For clincial OCT, the mean foveal thicknessacross all eyes was 297.9 µm with a mean intra-session variation of ±7.0 µm and with a total population variation of ±59.8 µm. For healthy retinas, the mean thickness was 283.7 µm, intra-session variation was ±2.17 µm, and with a total population variation of ±19.8 µm. For diseased retinas, the mean thickness was 310.8 µm, intra-session variation was ±11.4 µm, and with a total population variation of ±78.3 µm. When comparing the pair-wise difference in measurements between devices (see Fig. 3A) there was no statistically significant difference in central foveal thickness measurements between RAOCT and current, technician acquired clinical OCT for neither healthy retinas (p = 0.73, ICC = 0.956 [0.894 – 0.982]) nor diseased retinas (p = 0.95, ICC = 0.994 [0.986 – 0.998]). These limited differences can be better illustrated in the Bland-Altman plot in Figure 3B where there was a mean thickness difference of -1 µm between devices across the entire imaged population.
To establish that RAOCT performs comparably to clinical OCT across a broad range of demographics, we combined all thickness measurements and re-evaluated measured foveal thickness based on gender (female and male), race (White and Black), and age (25 – 39 years old, 40 – 54, 55 – 69, and 70 and older). Note that because we were only able to image a single patient of Asian descent, this individual was excluded from device comparisons for the race categorizations only. When comparing devices based on gender: female, p = 0.86 ICC = 0.994 [0.987 – 0.997] and male, p = 0.74 ICC = 0.986 [0.960 – 0.995]. When comparing devices based on race: White, p = 0.89 ICC = 0.994 [0.988 – 0.997] and Black, p = 0.94 ICC = 0.960 [0.830 – 0.992]. When comparing devices based on age: 25-39 years old, p = 0.82 ICC = 0.943 [0.767 – 0.988], 40 – 54 years old, p = 0.97 ICC = 0.985 [0.944 – 0.996], 55 – 69 years old, p = 0.98 ICC = 0.985 [0.955 – 0.996] and 70 and older, p = 0.72 ICC = 0.996 [0.984 – 0.999]. Importantly, when comparing the pair-wise difference in measurements between devices, we found no statistically significant difference in any of the sub-catagorizations with all p-values greater than 0.7 and ICC values greater than 0.9.
Table 1: Individual central foveal thickness for RAOCT and Heidelberg Spectralis devices in patients with and without retinal disease. Thickness was calculated in the central 1 mm diameter region of the fovea for each eye imaged with both devices in triplicate. Standard deviation represents intra-subject imaging session variation.
Patients with Healthy Retinas
|
RAOCT Foveal
Thickness (µm)
|
Spectralis Foveal
Thickness (µm)
|
Patient #
|
Gender
|
Race
|
Age
|
OD
|
OS
|
OD
|
OS
|
1
|
Female
|
Black
|
55
|
297 ± 0.6
|
286 ± 1.5
|
293 ± 1.0
|
279 ± 1.0
|
2
|
Female
|
White
|
28
|
250 ± 0.6
|
253 ± 1.2
|
255 ± 7.8
|
251 ± 0.6
|
3
|
Female
|
Black
|
49
|
273 ± 0.6
|
265 ± 1.0
|
275 ± 1.5
|
272 ± 0.6
|
4
|
Female
|
White
|
46
|
263 ± 0.6
|
262 ± 0.6
|
266 ± 2.1
|
268 ± 0.6
|
5
|
Male
|
White
|
57
|
287 ± 1.0
|
283 ± 1.2
|
287 ± 2.0
|
291 ± 3.5
|
6
|
Male
|
White
|
57
|
301 ± 1.0
|
308 ± 1.0
|
296 ± 2.1
|
305 ± 1.5
|
7
|
Male
|
White
|
36
|
313 ± 0.6
|
316 ± 0.6
|
309 ± 2.3
|
304 ± 4.6
|
8
|
Female
|
Asian
|
25
|
264 ± 1.0
|
262 ± 1.2
|
278 ± 5.5
|
271 ± 1.7
|
9
|
Male
|
White
|
83
|
312 ± 1.2
|
317 ± 0.6
|
317 ± 2.1
|
322 ± 1.7
|
10
|
Female
|
White
|
34
|
275 ± 0.0
|
269 ± 0.0
|
271 ± 0.0
|
264 ± 1.5
|
Patients with Diseased Retinas
|
RAOCT Foveal
Thickness (µm)
|
Spectralis Foveal
Thickness (µm)
|
11
|
Female
|
White
|
91
|
284 ± 1.2
|
-
|
285 ± 3.5
|
-
|
12
|
Female
|
White
|
49
|
293 ± 1.5
|
295 ± 2.1
|
286 ± 1.2
|
291 ± 3.5
|
13
|
Female
|
White
|
60
|
300 ± 1.5
|
-
|
302 ± 4.6
|
-
|
14
|
Female
|
White
|
81
|
260 ± 2.1
|
228 ± 4.6
|
265 ± 3.8
|
228 ± 8.2
|
15
|
Female
|
Black
|
82
|
222 ± 1.2
|
-
|
224 ± 4.0
|
-
|
16
|
Male
|
White
|
69
|
-
|
304 ± 1.2
|
-
|
322 ± 9.3
|
17
|
Male
|
White
|
66
|
-
|
421 ± 1.7
|
-
|
430 ± 90.0
|
18
|
Female
|
White
|
81
|
572 ± 2.3
|
-
|
561 ± 17.2
|
-
|
19
|
Male
|
White
|
66
|
321 ± 0.6
|
280 ± 0.6
|
324 ± 17.2
|
276 ± 1.0
|
20
|
Male
|
White
|
76
|
-
|
286 ± 1.2
|
-
|
295 ± 3.5
|
21
|
Female
|
White
|
66
|
272 ± 0.6
|
283 ± 1.5
|
269 ± 5.1
|
285 ± 2.1
|
22
|
Female
|
Black
|
51
|
252 ± 1.2
|
249 ± 0.6
|
246 ± 2.1
|
246 ± 3.6
|
23
|
Female
|
Black
|
76
|
-
|
305 ± 1.5
|
-
|
326 ± 19.6
|
24
|
Male
|
White
|
71
|
474 ± 8.9
|
276 ± 2.0
|
454 ± 14.7
|
277 ± 5.6
|
25
|
Female
|
White
|
43
|
274 ± 4.4
|
361 ± 1.2
|
272 ± 2.1
|
373 ± 28.7
|
Clinical diagnostic comparison
In addition to providing quantitative information, OCT images are utilized in the clinic for diagnostic purposes. To test the diagnostic utility of the images, we performed a secondary study using the acquired patient data wherein a retina clinician experienced with OCT interpretation labeled a given volume as either normal or abnormal and compared the results to the clinical diagnostic label (see Methods, Study Design). We found the sensitivity and specificity for detecting abnormal retinas to be 93% and 90%, respectively, for RAOCT and 87% and 60% for Spectralis. We found the positive and negative predictive values of RAOCT to be 93% and 90% while they were 76% and 75% for Spectralis.