This document summarizes a presentation on developing an automated patient immobilization system for head and neck cancer radiotherapy without the use of masks. It describes current radiotherapy techniques, issues with positioning errors, and an initial study using a soft robot system with an air bladder and visual feedback to automatically control a mannequin's head motion in one degree of freedom. Nonparametric analysis was used to model the system dynamics and design a PID controller to track position trajectories within 14 seconds, demonstrating a potential solution for accurate in-treatment positioning and motion compensation.
Towards fine-precision automated immobilization in maskless radiosurgery
1. HEAD AND NECK CANCER RADIOTHERAPY
TOWARDS FINE-PRECISION AUTOMATED IMMOBILIZATION IN
MASKLESS RADIOSURGERY
Presented by /Olalekan Ogunmolu @patmeansnoble
3. Head and neck (H&N) cancers are among the most fatal of
major cancers in the United States
2014: 35% of all pharynx and oral cavity cancers developed
led to fatility [Siegal, R. et. al]
Cancer kills almost 600,000 people each year in the U.S.
alone.
Source: NVIDIA Foundation Award.
4. SO HOW ARE H&N CANCERS TYPICALLY
TREATED?
Surgery
Pros: Oldest technique and most successful
Cons: Only useful when cancer is localized (highly improbable in most cases)
Chemos
Pros: Kills cancer cells throughout the body
Cons: Highly toxic; can kill healthy cells; highly carcinogenic
5. Radiation Therapy
Pros: Good procedure for distributed cancer cells
Pros: Palliative treatment when eliminating cancer tumors is impossible
Pros: Helpful to shrink cancer tumors pre-surgery or tumor leftovers post-surgery
Pros: Minimal exposure of patient to radiation (treatment less than 15 mins typically)
Often involves a combination of drugs and chemos
Body cancer radiotherapy (RT) typically use IMRT & IGRT
6. WAIT, WHAT IS IMRT/IGRT?
IMRT: Intensity Modulated Radiation Therapy
Deals with modulating the dosage and shaping of the radiation beam to precise size of tumor cells
IMRT improves accuracy of carefully targeted radiation thereby minimizing exposure of healthy
organs
Deviations still occur between planned dose and delivered dose of radiation
Enter IGRT
7. IGRT: Image Guidance Radiation Therapy
Deals with precise and accurate patient positioning on a treatment table to avoid dose deviations
from planned targets
The uncertainty in dose measures to malignant tissues necessitated use of IGRT before treatment
Goal was to assure precise localization of the beam onto the target tumor cell
Source: Prostrate Cancer Center
9. WHAT IT ENTAILS
Accurate markers are placed inside a patient's body after
consultation with a medic
Few days afterwards, a radiation-based scan (CT) of the
markers is performed to localize the exact position of the
markers in the gland
The scan provides the size and shape of the cancer cells for
computerized treatment planning calculations
10. Current IGRT radiation-based systems include
[Jennifer De Los Santos et.al., 2012]
Electronic Portal imaging detectors
e.g. IGRT and MV imaging; 1 - 2 mm accuracy; does not acquire 3D volumetric info
Cone-beam CT
retractable conventional x-ray tube and amorphous silicon x-ray detectors mounted either
orthogonal to the treatment beam axis; used in lung/throat/liver, brain, head and neck cancer
Fan-beam CT:
in-room gantry-moving CT linac system to move across the patient instead of couch moving
patient into the scanner as in conventional CT designs
11. Stereoscopic imaging
used in CyberKnife; 2D imaging system; accuracy < 1mm
Combination alignment systems: optical imaging and 2-D
kV orthogonal imaging
Facilitates localization of rigid and mobile targets which may be volumetrically aligned with CBCT
12. MOTIVATION
Clinical studies have shown that small perturbations cause
high sensitivity to IMRT treatment dose [L. Xing, 2000.]
6D couch motion compensaion system is not time-optimal
in treatments
Treatment is often stopped for a medical physicist to
recalibrate patient set-up when there is a deviation from
target pose
Evidence of treatment discomfort and severe pain from
long hours of minimally invasive surgery [Takakura, T., et
al., 2010]
13. RELATED WORK
Frameless and Maskless Cranial SRS [Cervino et. al. 2010]
Idea was to verify accuracy of IGRT systems without rigid
frames on face
Employed deformable masks of the following sort:
14. PROS
hropomorphic head phantoms employed in checking the accuracy of a 3D surface imagin
em (AlignRT Vision System)
mpared results from an infra-red optical tracking system with the AlignRT vision software
em
different couch angles, the difference between phantom positions recorded by the two
ems were within 1mm displacement and 1° rotation
ent motion due to couch motion was less than 0.2mm
15. CONS
6DOF positioning systems model the human body as a rigidly
No accounting for flexibility/curvature of neck
Limited positioning of patient can reduce effectiveness
Patient motion due to couch motion was less than 0.2mm
If patient moves , therapy must be stopped, patient repositioned, costs time and mone
16.
17. RESEARCH GOALS
AIMS
Accurate and automatic patient positioning system (pre-treatment)
In-treatment automatic and accurate patient positioning with patient motion compensation
OBJECTIVES
Surface-image control of the cranial flexion/extension motion of a patient during simulated H&N RT
(pre-treatment)
Use radiation-transparent soft robot system for positioning/manipulation tasks
18. OVERVIEW
Initial study and experiment demonstrating a 1-DOF intra-cranial control of patient motion during
H&N Cancer RT
Testbed is a Mannequin head lying in a supine position on an inflatable air bladder (IAB)
Soft-robot consists of the IAB, two two-port SMC Pnematics Co. proportional valves, and silicone
tubes for conveying air from a pressurized air canister
A Kinect RGB-D camera is employed for head motion sensing and feedback to a classical control
network implemented on an NI myRIO hardware
Work in partnership with my advisor, Dr. Gans and Drs. Xuejun Gu and Steve Jiang of the Radiation
Oncology Department of UT Southwestern, Dallas, TX, USA
21. VISION-BASED HEAD POSITION ESTIMATION
Kinect RGB-D Camera employed for position-based visual
servoing
Better depth image and alignment; Skeleton tracking
Real-time Human Pose Recognition in Parts from Single
Depth Images. Jamie Shotton, et. al, CVPR 2011, Best
Paper Award
Real-time 3D face-tracking based on active appearance
model constrained by depth data. Nikolai Smolyanski et. al,
Image and Vision Computing, 2014, MS SDK v1.5.2
Generates 640 × 480 image at 30 fps with depth resolution
of 40 centimeters
Roughly +/5 mm of quantization error
22. 3D face detection and tracking using active appearance
model
AAM minimizes an energy function defined by distance
between 3D face model vertices and depth data coming
from a RGBD camera.
Residual errors are used to modify the model during run
time
Vertices correspond to facial features and can be identified
for consistent use
We use eyes, nose tip and edge of mouth for tracking in
image
We use bridge of nose for depth in control algorithm
23. Process Flow in Face Tracking Procedure
Find a face rectangle in a video frame using a face
detector
Find five points inside the face area – eye centers, mouth
corners, and tip of the nose
Precompute scale of tracked face from the five points
un-projected to 3D camera space and scale 3D camera
space appropriately.
Initialize next frame’s 2D face shape based on the
correspondences found by a robust local feature
matching between that frame and the previousframe.
24. RESULTS
With the depth constrained 2D+3D AAM fitting, we found
good position- estimation results on a human subject when
object is at a distance of 1 to 2.5m from the Kinect System
Generalization errors and hence incorrect position
estimation errors with respect to the mannequin head due
to inconsistency in depth data
25. Face Mesh and Tracker on (left) my face and (right) the manikin face.
26. MODELING PROCEDURE
Collect Data Set, of input-output signals, at each time
step for a total of samples
Objective: fit a continuous-time parametric model
structure similar to a one-step ahead predictor
Therefore, form the parameter vector
Z
N
k N
= − − ⋯ − y + +y
n
an−1 y
n−1
a0 bm u
m
+ ⋯ + ubm−1 u
m−1
b1 u˙ b0
θ = [ , ⋯ , , , ⋯ , ]an−1 a0 bm−1 b0
27. MODELING PROCEDURE CONT'D
and a memory-fading vector of past input-output data:
such that the estimated output can be written as
Choice of excitation signal important to reproduce desired
properties of the system in model and avoid wide crests as
much as we can
Identification Goal: identify the best model, χ, in the set
guided by the rigorous freq. distribution analysis
ϕ(t) = [− ⋯ − , −y, ⋯ u]y
n−1
y˙ u
m
(t|θ) = (t)θy^ ϕ
T
28.
29. Excitation signal is sawtooth waveform
Integral and differential of a sawtooth waveform preserves
the sawtooth waveform with only phase and amplitude
shifts
Spectrum contains both even and odd harmonics of the
fundamental frequency
:::to excite all frequency dynamics we want from model
Waveform amplitude = 165mA (max. operating current to
SMC valves)
Frequency was chosen to avoid aliasing [i.e. > 2 sampling
frequency (Nyquist Sampling theorem]
×
30. Excitation signal should produce corresponding
rise/decrease in head pitch motion
:::9,000 samples is not rich enough to avoid inherent noise which dwarfs data structure
Therefore, we remove means and linear trends
::to remove outliers, high frequency spikes etc
The rest of the modeling stages is straightforward
::prewhiten input signals, estimate impulse response (to examine degree of delay in data), examine
correlation functions (Wiener model)
Crosscorrelation from input to output should tell us about
the dynamics of system
::since it is proportional to the kronecker delta function (impulse response)
31. CROSS-CORRELATION ANALYSIS
The cross-correlation function provides an estimate of the
system impulse response and is defined as:
(τ ) =ψuy
[u(t − τ ) − ][y(t) − ]∑
t=τ +1
N
u¯ y¯
∑
t=1
N
[u(t) − ]u¯
2
− −−−−−−−−−−
√ ∑
t=1
N
[y(t) − ]y¯
2
− −−−−−−−−−−
√
where .τ = 0, ±1, ⋯ , ±(N − 1)
32. The cross-correlation function (CCF) is given by the
convolution of the system impulse response and the process
auto-correlation function (Wiener-Hopf equation)
(τ ) = ∫ h(ν)E[u(t)u(t + τ − ν)]dνψuy
= ∫ h(ν) (τ − ν)dνψuu
33. Wiener Block-Structured Model
CCF between the output and input is proportional to the system impulse response when the input is
white noise
Prewhiten input-output signals to change structure of signals
where is a zero mean white input sequence and is an autoregressive model filter
U (t) = (t)Uw
1
F ( )Z
−1
(t)Uw F ( )Z
−1
35. (Left) Cross-Correlation of input-output signals and (Right) input signal auto-correlation function
.
So we gained some intuition about system based on the
non-parametric analysis
notice data has an 18-sample delay ( 2 sec delay)≈
36. CORRELATION OF RESIDUALS
To determine the model structure of the system, we used
the original detrended data
We chose a linear, second-order grey-box model set whose
quality is measurable by a mean-square error (MSE) guided
by nonparametric estimate
Choice ensures cost of model is not too high in solving for
a high−order complex model is more difficult to use for simulation and control design. If it is not
marginally better than a simpler model, it may not be worth the higher price [Llung, 16.8]
θ^
N
§
37. COMPARING MODEL STRUCTURE
The confidence interval compares the estimate with the estimated standard deviation from the
validation dataset
A 99% confidence region (yellow bands) encloses the model response informing us we have a
reliable model [Llung (1999), §16.6]
Evaluation of different model structures and comparing quality of offered models
Best fit: a second-order process model with delay and a RHP zero
The model has an 87.35% fit to original data with a mean square error of 0.054982 and a final
prediction error of 1.672.
G(s) = ex .
−0.0006(s − 1.7137)
(s + 0.01)(s + 0.1028)
p
−2s
38. SUB-MODEL SELECTION & MODEL VALIDATION
A control system will perform well with an optimal linear
sub-model, tolerate disturbances and nonlinearities.
We pick the linear frequency range based on intuition
garnered from bode panalysis to represent the model.
Canonical correlation analysis of residuals from prediction,
to true , and estimated position by the auto-
regressive model we chose, i.e. the residual
(t)y^ y(t)
ϵ(t, ) = y(t) − (t| )θ^
N y^ θ^
N
39. (Left). Bode plot of detrended data; (Right) Residuals from input to output
Frequency response plot of residuals to output.
40. The prediction errors are computed as a frequency
response from the input to residuals
CONTROL DESIGN
Open Loop Step Response of Identified System
41. System is non-minimum phase with very slow transient
response.
We require a controller that will increase the response
time, guarantee cloloop stability whilst balancing
robustness and controller aggressiveness.
Approximating the delay with the second-order Pade
function,
H(s) = .
− 3s + 3s
2
+ 3s + 3s
2
42. and introduce the PI controller
nested within a PID controller:
Gc = 3.79 + .
0.0344
s
= 3.4993 + + 55.8988s,GP ID
0.054765
s
43. Block Diagram of Control Network
CL Step Response of Simulated System.
CL Step Reference Trajectory Tracking
44. CONCLUSIONS
Deviations from desired positions during H&N Cancer RT
cause dose variations and degenerate treatment efficacy
We have offered a proof-of-concept evaluation and trial of
the accurate control of the cranial flexion/extension
motion of a patient during maskless H&N RT
The soft robot system can track set trajectory within 14
seconds after start-up with the aid of a PID/PI cascaded
network
45. ONGOING WORK
Extend results to deformable motions of the upper torso,
and H&N.
Improved bladder control: Optimal LQG Control, Adaptive
Control, Dynamic Neural Network control
Incorporation of multi-bladders to accommodate multi-
axis positioning (Future Work)
Vicon Mocap System for online monitoring/feedback to
controller
46. BENEFITS
Comprehensive and accurate control of the patient’s
position
Elimination of anatomical deformations as a result of
positioning error
47. REFERENCES
Cervino, L. I., et al. Frame-less and mask-less cranial
stereotactic radiosurgery: a feasibility study. 2010, Physics
In Medicine And Biology 55(7): 1863-1873.
Jemal A, Siegel R, Xu J, Ward E. Cancer statistics, 2010. CA:
A Cancer Journal for Clinicians2010; 60(5):277–300.
L. Llung, System Identification Theory for the User, 2nd
Edition, Upper Saddle River, NJ, USA. Prentice Hall, 1999.
48. REFERENCES CONT'D
Xing, L. Dosimetric effects of patient displacement and
collimator and gantry angle misalignment on intensity
modulated radiation therapy. Radiotherapy & Oncology,
2000. 56(1): p. 97 - 108
49. HEY, CAN I HAVE YOUR SLIDES?
No pressure. Simply add "?print-pdf" as a query string to this
presentation weblink; then Save as PDF
1 of 49
Like this ? Why not share!