IEEE TRANSACTIONS ON MEDICAL ROBOTICS AND BIONICS, VOL. 7, NO. 3, AUGUST 2025 975
Multi-Sensory System for Spatially
Aware Colonoscopy
Tuukka Panula , Bruno Rosa , Member, IEEE, Salzitsa Anastasova-Ivanova, Matti Kaisti ,
and Benny Lo , Senior Member, IEEE
Abstract—The traditional method of performing a colonoscopy
requires a trained physician operating the instrument manually.
There is no proper feedback to aid the user in the process and
typically the physician has to rely solely on the camera output
to guide the endoscope. This method can lead to discomfort or
even colon perforation due to the elongated endoscope scratching
or tearing the colon tissue. To address this issue and enable
spatial awareness to the instrument, we propose a modular multi-
sensor system that integrates bending, pressure, and motion
sensing units into the endoscope. All sensors are manufactured
using inexpensive off-the-shelf components. The proposed sensor
system was characterized on a robotic test bench and validated
in a colon phantom study. The results demonstrate the feasibility
and robustness of the proposed sensor fusion approach in
colonoscopy which has the potential for safer and more effective
inspection of the bowels. The introduction of these sensing
modalities to an endoscope paves the way for AI-assisted and
possibly autonomous colonoscopy in the future.
Index Terms—Colonoscopy, sensor, pressure, strain, IMU.
I. INTRODUCTION
A. Motivation
COLONOSCOPY is the gold standard method for diag-nosing gastrointestinal disease, with more than one
million procedures performed each year in the UK alone [1].
Although generally safe, it requires a highly trained and
experienced clinician to carry out the procedure and could
still lead to complications. As traditional colonoscope does
not have any sensing capability, endoscopists have to rely
on the camera images to navigate inside the colon and their
feeling of friction when inserting the colonoscope into the
lumen to avoid damages to the colon tissues. However, the
elongated colonoscope, which is bent and twisted while being
inserted into the large intestine, can easily scratch or tear the
Manuscript received 19 November 2024; revised 14 February 2025;
accepted 7 April 2025. Date of publication 23 May 2025; date of current
version 21 August 2025. This article was recommended for publication
by Associate Editor F. Ficuciello and Editor P. Dario upon evaluation
of the reviewers’ comments. This work was supported in part by the
University of Turku Graduate School (UTUGS) Grant; in part by the
Instrumentarium Science Foundation, Finland; and in part by the Engineering
and Physical Sciences Research Council, U.K., under Grant EP/P012779.
(Corresponding author: Tuukka Panula).
Tuukka Panula is with the Hamlyn Centre, Imperial College London,
SW7 2AZ London, U.K., and also with the Department of Computing,
University of Turku, 20500 Turku, Finland (e-mail: tuukka.j.panula@utu.fi).
Bruno Rosa, Salzitsa Anastasova-Ivanova, and Benny Lo are with the
Hamlyn Centre, Imperial College London, SW7 2AZ London, U.K.
Matti Kaisti is with the Department of computing, University of Turku,
20500 Turku, Finland.
Digital Object Identifier 10.1109/TMRB.2025.3573053
colon tissue, causing pain and discomfort for the patient, and
could even rupture the colon wall [2]. Pain experienced during
colonoscopy is usually caused by one of three reasons [3].
i) When the operator pumps CO2 gas to insufflate the colon
to allow the colonoscope to pass through the lumen and
to obtain a clear view of the surrounding tissue, a portion
of the colon could get overly inflated which causes pain.
ii) Excessive force applied to the colon wall by the colono-
scope can cause pain and, in the worst case, perforation
- rupturing the colon wall. iii) In many cases, the pain
experienced is due to looping, which means the colonoscope
is being forced into certain shapes while inserting into the
colon. Typical loop formations include the N and alpha
loops [4], with names that stem from their respective shapes.
It has been reported that looping occurs in 90 percent of
all colonoscopies [3]. Both excessive pressure and looping
can be very difficult to prevent without having any other
feedback than the scope’s camera view. Assistive systems have
been introduced to assist endoscopic gastrointestinal surgery,
especially in the form of robotic-driven platforms [5].
B. Related Work
There are devices available commercially to minimize
pain and track the movement of the endoscope during the
procedure. The simplest is an abdominal compression vest
(Colowrap, North Carolina) that is used to prevent the colon
from moving and thus the instrument from looping. However,
its efficacy has been questioned [6]. ShapeLock (USGI
Medical, California) is a sleeve that when placed around the
colonoscope, prevents looping by restricting the curving of the
instrument by locking itself to place [7]. Magnetic endoscope
imaging (MEI) is a powerful technique where a set of coils
are attached to the endoscope shaft at certain intervals and
a detection coil is placed near the patient [8]. This way, a
three-dimensional image of the instrument can be constructed
to be used as an aid to the examiner in locating the endoscope
tip. One such system on the market is ScopeGuide (Olympus,
Japan) [9]. Magnetic manipulation of the scope tip has also
been proposed in [10]. Relating academic endeavors have
also been reported, such as a soft robotic sleeve that has
inflatable pockets along with optical waveguides to measure
contact pressure [11]. The air pockets can be inflated when
excessive force is directed to the endoscope. The forces exerted
in colonoscopy using sensor technology have been studied
before. By placing a handle equipped with force sensors on
the scope shaft, researchers were able to record the push and
c© 2025 The Authors. This work is licensed under a Creative Commons Attribution 4.0 License.
For more information, see https://creativecommons.org/licenses/by/4.0/
976 IEEE TRANSACTIONS ON MEDICAL ROBOTICS AND BIONICS, VOL. 7, NO. 3, AUGUST 2025
Fig. 1. Instrument overview. a) Illustration of the proposed system in use. b) 3D rendition of the flexible endoscope tip with the sensors PCBs assembled.
Two different sized silicone tubes, outer and inner, are used. c) Pressure sensing module assembly. Four pressure sensors are attached in 90 degree intervals,
enabling a four directional pressure profile. The proposed instrument is a dedicated system rather than a conventional colonoscope. d) Cross-section figure
of a single pressure sensor module. e) Bending sensor module assembly. The sensor PCBs are attached to the substrate with the opposite sides connected
together via a flexible interconnect.
pull forces during the procedure [12]. In another study, a force
sensing sheath consisting of an array of conductive copper
tapes was wrapped around the endoscope. By measuring
the effect of applied force on their resistance, along with
simultaneous MEI, the researchers could reconstruct a 3D
force map of the colonoscope [13]. The use of bending sensors
to detect looping have also been proposed by [14]. In this
approach, the researchers attached a series of resistive bending
sensors to the probe shaft, giving the examiner feedback of
the probe angle. However, using only strain sensors, looping
detection is very limited and only works when the loop has
already formed. We believe, that by fusing sensor data from
multiple different sensor modalities we can collect sufficient
information to prevent looping.
Alternative directions of development in less intrusive
colonoscopy are the use of untethered devices and endorobots.
Capsule endoscopy is a technique where a small device
equipped with cameras is ingested by the patient. While
passing through the gastrointestinal system, the device sends
image data to the operator [15]. Similar technology is used
in a self-propelling variant [16]. In another study, a soft
robotic device uses pressurized balloons to navigate through
the colon [17]. Magnetic tracking technology has also been
adopted to capsule endoscopy in a instrumentation setup called
MagnetOFuse [18]
C. Overview
To minimize the risk of colon rupture and looping during
colonoscopy, we propose a modular sensor system equipped
with contact pressure and scope bending sensing, as well as
a 6-axis inertial measurement unit (IMU) to allow endoscope
sensing. These sensors can be used together to provide the
user real-time feedback of the forces exerted to the colon wall,
endoscope’s orientation, and bending of each joint. Traditional
endoscopes do not have any sensing modalities for receiving
input from its state. In our technique, contact pressure to the
lumen wall is measured via barometric pressure sensors that
are covered with air cushions, using a construction similar
to [19]. Bending sensing is achieved using strain gauges in
a half-bridge Wheatstone bridge connected to a instrument
amplifier setup [20]. The advancement of the scope is recorded
by measuring the 3D orientation of the scope tip using
integration and Kalman filtering, while being inserted to the
colon. A 3D rendition of the proposed system is shown in
Figure 1b. Consequently, the user can monitor both the tip
of the scope and the proximal portion of the shaft, thereby
enabling the detection of recurrent looping.
The sensor modules were built along with the data acquisi-
tion system and integrated into a proof-of-concept endoscopic
test bench that uses a wire-pulley system to operate the
endoscope tip, similar to that described in [21]. The pressure
PANULA et al.: MULTI-SENSORY SYSTEM FOR SPATIALLY AWARE COLONOSCOPY 977
Fig. 2. System block diagram. a) The bending sensor PCB. The strain gauges are connected to an instrumentation amplifier in a half-bridge configuration.
b) Inertial Measurement Unit (IMU) with a 3-axis accelerometer and a 3-axis gyroscope. c) A closeup image of the pressure sensor with the cushion attached
on top d) The data acquisition (DAQ) card has a USB connector for connecting to the PC.
sensors were tested by placing a series of weights on them.
Operation of the bending sensor modules was verified by
programming the scope to maneuver in concentric loops and
reading the sensor output. The effect of external temperature
on the pressure and bending sensors was observed in a separate
experiment. Operation of the IMU was verified by inserting
the sensor through plastic tubing fixed in an alpha loop
formation. Finally, the instrument was tested using a custom
built colon phantom. Using a phantom instead of a live animal
or human subjects was the clear choice at this point of device
development. We can obtain sufficient data for a proof-of-
concept study without the risks and cost of in vivo trials. The
system is designed with off-the-shelf components with the aim
of providing a low-cost solution and support the development
of disposable colonoscope so as to minimize the risk of cross
infection.
The combination of the sensing modalities has not
previously been reported in colonoscopy. Furthermore, actual
pressure values exerted to the colon wall have not been
exhaustively reported before. In one study, the researches
measured the forces associated with the insertion and twisting
of the colonoscope, by adding strain sensors to the instruments
handle [12]. Another study assessed the intraluminal pressures
caused by insufflation. Here, air was pumped into the colon
and the pump pressure was recorded [22]. The study found that
a pressure of approximately 19 kPa could cause the sigmoid
colon to rupture.
II. METHODS
A. Sensor Selection
Our aim was to find compact, easy-to-use, and accurate
sensors to use on the instrument. For the IMU we considered
three of the largest manufacturers’ products: BMI323 (Bosch
Sensortec, Germany), ICM-42670-P (TDK InvenSense, Japan)
and LSM6DSO (STMicroelectronics, Switzerland). These all
have the same 12-lead footprint and are similarly priced and
include both a 3-axis accelerometer and a 3-axis gyroscope.
The power spectral densities for the accelerometer are 180,
100 and 70 μg/
√
Hz, respectively. The gyroscope angular
rate spectral noise densities are 7, 7 and 3.8 mdps/
√
Hz,
respectively. While not being the only factor accountable for
the quality of the sensor, we chose the LSM6DSO as the IMU.
For the pressure sensor, we considered mainly three fac-
tors: resolution, pressure range, and package costruction.
Especially the upper limit for absolute measurement range
should be high enough to measure the pressures exerted
to the colon wall. Here, we also examined three sensors:
Bosch Sensortec BMP390, TDK InvenSense ICP-20100 and
STMicroelectronics LPS22H series. The LPS22H outweighs
the other two in both resolution (24-bit vs. 20-bit) and pressure
range with the maximum pressure of 126 kPa versus 125
kPa for BMP390 and 110 kPa for ICP-20100. However,
alternatively to the air cushion construction, we wanted to
retain the ability to submerge the sensor to a silicone bulb,
similar to [23], which the molded construction of the LPS22H
does not allow. Our previous experience with the previous
generations led us to choose the BMP390 [19].
The selection of the 350  strain gauge BF350-3AA
(generic) was done simply due to it’s wide availability and
low cost.
B. Sensor Module Fabrication
The sensors are housed on individual 0.2 mm thick flexible
polyimide (PI) printed circuit boards (PCB). Each PCB has
a flat flexible connector (FFC) to interface with the data
acquisition card (DAQ). The pressure and bending sensor
PCBs have meandering interconnects that connect the different
parts of the PCB as shown in Figure 1b.
Each pressure sensor PCB houses two BMP390 absolute
barometric pressure sensors. To allow the sensors to sense
contact pressure, an air cushion is attached on top of each
sensor. The PCB is shown in Fig. 2c. We had already used this
approach to measure blood pressure from the fingertip [19].
The airtight pressure cushion made from low-density polyethy-
lene (LDPE) is created by punching a hole to a silicone
substrate with adhesive on both sides, and then attaching the
air cushion to one side. By cutting a small opening to the
air cushion, air is allowed to flow in the sensor cavity. This
978 IEEE TRANSACTIONS ON MEDICAL ROBOTICS AND BIONICS, VOL. 7, NO. 3, AUGUST 2025
TABLE I
COLLECTION OF THE POSSIBLE SENSORS CONSIDERED FOR THE DEVICE. THE SELECTED SENSORS ARE HIGHLIGHTED
assembly is then attached to the sensor PCB. After assemble
to the PCB, the air volume inside should remain constant and
by applying force to the cushion, the internal pressure in the
cavity can be sensed by the atmospheric pressure sensor. Two
pressure sensor PCBs - a total of 4 sensors - are used on
each joint of the endoscope in order to get a pressure reading
from four directions. The scope is at its thickest (23 mm)
around the pressure sensors, but due to the high conformity
of the cushions, the diameter when inserted into the colon is
19 mm. The pressure sensor was also used to measure the
temperature with its on-chip thermometer with a resolution of
16-bit/0.005 ◦C/bit.
The bending sensor PCB has two 350  aluminium foil
strain gauges. One module consists of two PCBs, each housing
two strain gauges. The PCB is shown in Fig. 2a. The strain
gauges are connected in a half bridge configuration with both
strain gauges attached to the opposite sides of the endoscope.
When the endoscope is bent, the strain gauge on one side will
be stretched and the one on the other side will be compressed,
doubling the sensing range and ensuring that the setup is partly
protected from temperature variation. The strain gauges are
connected to an instrumentation amplifier MAX41400 (Maxim
Integrated) with a gain of 200 (G = 100 selectable). The reason
for selecting the MAX41400 was its very small size (9-ball
wafer level packaging) and suitability for medical equipment
manufacturing. In addition, one of its intended use cases is to
be used with a Wheatstone bridge strain gauge setup.
The IMU sensor board houses an LSM6DSO IMU. The
assemled PCB is shown in Fig. 2b. In the current design, we
only used one IMU placed at the tip of the endoscope.
C. Data Acquisition Electronics
A 4-layer DAQ card was designed to interface with the
individual sensor boards. Nordic Semiconductor’s nRF52840
based USB dongle is used as the main microcontroller unit
(MCU). It hosts a 32-bit ARM Cortex-M4 processing unit
and has built-in USB capabilities. The DAQ card has 4 or
6-pin FFC connectors for connecting the sensors - Two for
the pressure sensor boards, two for the bending sensor boards
and one for the IMU. The pressure sensors and the IMU
are interfaced via SPI (Serial Peripheral Interface) and I2C
(Inter-Integrated Circuit) connections. An analog multiplexer
(MAX4734, Maxim Integrated) is used to select between the
two bending sensors. The amplified strain gauge signal is read
via a 16-bit analog-to-digital converter (ADC). A low-dropout
regulator (NCV8114, Onsemi) is used to generate a 1.65 V
reference signal for the strain gauge instrumentation amplifier.
The DAQ card is shown in Fig. 2d.
D. Endoscope Test Bench
A test bench system was built for sensor characterization.
The system consists of a flexible endoscope shaft, housing
a 9 mm instrument channel, a motorized capstan unit and
control electronics. The control electronics is built based on
the Arduino Uno R3 MCU platform and has an analog joystick
and a pair of darlington motor controllers. The shaft is divided
into a proximal part housing four Bowden cables, and a
distal portion containing the sensing mechanics, which can be
maneuvered in four directions by tightening the corresponding
tendon via a pair of capstan drives. The capstan is driven
by a 28BYJ-48 (generic) stepper motor, that is controlled
with an analog joystick (KY-023, generic), enabling 2-DOF
control of the endoscope tip. Additionally, the instrument can
be programmed to trace a pre-defined trajectory. Figures 3a
and b show the assembled system.
E. Effect of External Temperature
The actual measurement environment where the instrument
is inserted into the human body differs from in vitro exper-
iments. One such aspect is temperature. The measurements
were done at room temperature, which is relatively low
compared to body temperature. We studied the effect of
external temperature on the sensors by placing the instrument
in a heating chamber. The temperature was linearly increased
from room temperature (26◦C) to body temperature (37◦C).
We compared a pressure sensor equipped with and without the
air cushion. In addition, the two bending sensors located in
the same joint were compared: stressed and unstressed.
F. Software and Data Analysis
The firmware for the nRF52 MCU was written in C using
Segger Embedded Studio. All sensors are sampled at 50 Hz.
The sensor data (4 x pressure, 2 x bending, 6-axis IMU,
temperature) along with the timestamp was sent via serial
port connection to a PC running MATLAB R2022b. Each
DAQ card is defined as a separate serial port object, so
the theoretical limit for individual joints on the scope is
256. The sensor data plots can be read real-time from the
GUI. The firmware for controlling the stepper motors for the
PANULA et al.: MULTI-SENSORY SYSTEM FOR SPATIALLY AWARE COLONOSCOPY 979
Fig. 3. Testbench design. a) Mechanics of the endoscope test bench. Four tendons are connected to a pair of capstans driven by stepper motors. The steel
tendons are routed to the tip of the endoscope via bowden cables. The motors are controlled by a custom control unit with an analog joystick. b) Photograph
of the test bench device with the sensor modules assembled.
endoscope test bench was built using the Arduino platform.
All post-processing and data analysis was carried out using
MATLAB R2022b. PCB design was done using Autodesk
EAGLE 9.6.2 electronic design automation (EDA) software
and mechanical design as well as finite element analysis (FEA)
was done on Autodesk Inventor 2022 computer aided design
(CAD) software. The 3D printing was done on a Formlabs
Form 3 stereolithography (SLA) printer.
G. IMU Data Processing
Orientation of the scope tip was calculated using the
gyroscope data fused with the accelerometer data using the
Kalman filter. The resulting quaternion vector was sampled
at 5 cm intervals. Since our device did not have pressure
sensor modules at the proximal part of the instrument, we
simulated the pressure sensor input by detecting the changes
in the accelerometer signal. The same algorithm is directly
applicable to the pressure sensors, since we know that the shaft
was advanced in exactly 5 cm intervals. The insertion intervals
were found from the accelerometer signal using a peak finding
algorithm. Since we know that all tip advancements were
exactly 5 cm, we can reconstruct every tip movement in a 3D
plane. Orientation at each point is defined by a quaternion q
and normalized (qn):
q = a + bi + cj + dk (1)
where i, j and k are the base vectors and a, b, c and d are real
numbers outputted by the sensor fusion algorithm. Quaternion
q and normalized (qn):
qn = q√
a2 + b2 + c2 + d2 (2)
We then define a vector u that points forward, e.g., to the
direction of the Y-axis, in the sensor coordinate system. Vector
u is converted to quaternion representation (uq).
u =
⎡
⎣
0
1
0
⎤
⎦ · D ⇔ uq = 0 + 0i + Dj + 0k (3)
where D is the distance the scope advances at each interval
(D = 5 cm). The scope advancement in 3D at each step
is calculated by performing a rotation using the Hamilton
product:
vq = quqq∗ (4)
where q∗ is the conjugate of quaternion q. Now, by appending
each resulting R3 vector to the previous vector’s endpoint, we
get the 3D scope trajectory in the tube[24].
H. Colon Phantom
A partial colon phantom was fabricated to test the proposed
sensor system in vitro. We used a 25.4 cm silicone bowel
model (Chamberlain group, Massachusetts) with an inner
diameter of 22 mm. The 3D printed supports were used to hold
the model in place. Three different setups were assembled.
The first assembly simulated a straight portion of the colon
(Figure 5c). The second formed an S-shaped curve simulating
the sigmoid colon. The sigmoid mesocolon was modeled
by attaching a spring loaded fixture to the silicone tubing,
simulating the resistance created by the mesentery (Figure 5d).
For the third assembly, we added a sphincter model to create
a narrowing in the tube (Figure 5e). This was achieved by
flexible wiring wrapped across a 3D printed frame, and
inserting the silicone tube through it. In order to protect the
sensor electronics and minimize friction, the scope tip was first
covered with a condom prior to being inserted into the colon
model. A water-based lubricant was used to prevent damage
to the silicone colon model.
We performed three types of experiments with the fabricated
colon phantom. These are depicted in Figure 5c through e. For
the first experiment, the silicone tube was set up as a straight
pipe (Figure 5c). The endoscope was inserted 15 cm deep in
the phantom. The scope tip was maneuvered to each four
orthogonal direction (Figure 5a) while recording the contact
pressure. Figure 5c shows the results of the experiment. For
the second experiment, the silicone tube was arranged in an
S-shape, simulating the sigmoid colon. It has been reported
that 77% of the pain associated with colonoscopy happens
when the scope tip is in the sigmoid colon [25]. The scope
was inserted through the colon model and its bending was
controlled with a joystick. Figure 5d shows the pressures from
980 IEEE TRANSACTIONS ON MEDICAL ROBOTICS AND BIONICS, VOL. 7, NO. 3, AUGUST 2025
Fig. 4. Sensor characterization. a) Pressure sensor sensitivity was tested by placing weights on the sensor cushion. b) Step response and sensor drift was
assessed by placing a weight on the sensor. c) The operation of the bending sensors was verified by programming the scope to move in concentric circles
(r=10 mm, r=25 mm, r=40 mm) with the radius increasing at each step. The figure shows the desired and measured trajectories. d) Finite element analysis
of the strain at the bending sensor plates at different bending directions. e) The endoscope tip was maneuvered in a circular trajectory, resulting in sinusoidal
outputs. f) Effect of external temperature on the pressure sensors (with and without an air cushion). g) Effect of external temperature on the bending sensors
in the same joint (x and y axes). Strain 1 is in a stressed state and strain 2 is unstressed, which causes the difference in their respective DC levels. Heating
was turned on at approx. 20 s.
the distal pressure module (top) and the recorded strain from
one of the bending sensors.
III. RESULTS AND DISCUSSION
A. Sensor Characterization
1) Pressure Sensor: We measured the sensitivity of the
pressure sensor modules by placing and removing very small
weights (75 mg, 150 mg and 225 mg) on the pressure sensor
cushion. The pressure output for the weights were 14.1 Pa,
30.8 Pa and 48.8 Pa respectively. The sensor was able to
detect even the smallest weight and its loading and unloading
cycle (Figure 4a). Because the weights used were so small, the
adjacent unloaded pressure sensor was used to compensate for
changes in atmospheric pressure. This was done by subtracting
the unloaded pressure from the device under test. A 90 s mea-
surement with the sensors unloaded resulted in atmospheric
pressure fluctuating with mean and standard deviation (SD) of
(102080 ± 3.4) Pa. Despite the compensation, there is some
hysteresis present in the measurements. The result for the
unloaded pressure showed a fluctuation with SD = 4.4 Pa. We
believe that the hysteresis and drift are likely due to the small
deformations of the cushion material between the loading and
unloading cycles. A larger weight (800 mg) was used to see
if there is any drift present. The initial overshoot was 6.8 Pa
higher than the final value. It took 9.1 s for the pressure signal
to settle to within 10% of the final value. As seen in Figure 4b
there is very minimal drift noticeable and the step response
is clearly sufficient in practice. It is notable, that the pressure
changes reported above are very small in comparison with the
pressures in the actual use case, where the expected pressures
are measured in kilopascals versus pascals. For this reason, the
amount of overshoot measured in the experiment is negligible.
Figure 4e shows, that when the heating is turned on, the
pressure signal with air cushion equipped experiences a fast
increase and stabilizes after that. The change in pressure is
approximately 10 Pa. This can be considered negligible in a
real use case. The bare pressure sensor does not experience the
same effect. The phenomenon can be explained by the ideal
gas law, since the increase in temperature causes the air inside
the cushion to expand, thus increasing the pressure. The bare
pressure sensor is not affected by this.
Characterization with actual pressure exerted in
colonoscopy were observed in the colon model experiments.
PANULA et al.: MULTI-SENSORY SYSTEM FOR SPATIALLY AWARE COLONOSCOPY 981
Fig. 5. Colon phantom experiments. a) Cross-section view of the scope with each tendon color coded. The plots in the figures correspond to the pressures
associated with each tendon being pulled. b) Image of the endoscope tip in relaxed (top) and bent (bottom) positions. c) Straight tube with the scope tip
maneuvered in four directions. d) The scope is advanced into the sigmoid colon model. The resistance of the sigmoid mesocolon is modeled via a spring-loaded
fixture. e) A sphincter model is used to create a narrowing in the tube, similar to, e.g., the anal sphincter. Figures I-III depict the direction of forces associated
with each experiment.
As can be seen from the pressure plot, P3 shows the greatest
increase when the scope is bent at approx. 30 s into the
measurement. This happens due to the scope tip being pushed
against the colon wall through the bend. After the bend, the
scope is again straightened at approximately 75 s. For the third
experiment, the silicone tube was again set up as a straight
pipe and a sphincter model was placed halfway through the
tube. When the scope was inserted through the sphincter
model, relatively even pressure was exerted to all four pressure
sensors. This is shown in Figure 5e. The pressure values in
the colon phantom experiments range from approximately
1 kPa to 20 kPa.
2) Bending Sensor: The bending module has four plastic
plates on which the strain gauges are placed. These are used
to transfer the movement of the central silicone tube to the
sensors. The plates are in a slightly stressed state when
the scope is straight, and either relax or bend further when
the scope is bent. Figure 4d shows the finite element analysis
(FEA) of an individual plate being stressed. Perpendicular
force exerted on the plate either increases or decreases the
maximal Von Mises stress seen on the plate. The gauges are
connected in a half-bridge configuration with the opposite
sides stretching in opposite directions when the probe is bent.
So each joint has two pairs of strain gauges in order to be
able to record motion in a two-dimensional plane. The bending
sensors had some variation in their output range, which can
be seen in Figure 4f. Here, the endoscope tip was maneuvered
in a circular trajectory, with the same tendon pulling ratio in
both x and y directions, resulting in sinusoidal outputs. The
same degree of bending resulted in slightly different changes
in resistance in the sensors. Because of this, each sensor output
had to be normalized. To get rid of the noise, a Butterworth
low-pass filter with a cutoff frequency of 1 Hz was used to
filter each signal. The variation could be partly caused by
982 IEEE TRANSACTIONS ON MEDICAL ROBOTICS AND BIONICS, VOL. 7, NO. 3, AUGUST 2025
Fig. 6. IMU experiment. a) Plastic tubing fixed in an alpha loop formation. b) Sensor fusion algorithm for acquiring the three-dimensional scope tip
trajectory. Angular velocity and acceleration are fused using the Kalman filter resulting in a quaternion describing the orientation of the scope tip. Insertion
depth information acquired from the pressure sensors is combined with the orientation data and used to reconstruct a 3D path of the scope trajectory.
manufacturing tolerances in the inexpensive foil strain gauges,
or the thickness of the glue used.
The operation of the calibrated bending sensors was verified
by programming the scope to maneuver in certain trajectories.
The tip was maneuvered in three concentric circular trajecto-
ries (r=10 mm, r=25 mm, r=40 mm) with the bending radius
increasing at each step. These trajectories equal approximately
the angles of 5◦, 30◦ and 60◦ in relation to the proximal
joint, respectively. Four full circles were done at each step.
Figure 4c shows the results along with the desired trajectories.
The root mean square error (RSME) of the measured and
desired output was 2.79 mm. The error is a combination of
both the inaccuracy of the sensors and the actuation itself.
Variation in tightness of the tendons and any play in the motors
can affect the actual movement of the scope.
Temperature variation affects the bending sensor readings
significantly. In case of the stressed bending sensor, a decrease
in the signal is caused by expansion of the gauges, which
leads to change in resistance. Some of this effect is canceled
due to the half-bridge configuration of the strain gauges, as
both elements expand approximately the same amount. In the
unstressed bending sensor, the temperature does not affect
the signal as much and the change is only 37% of the one
experienced by the stressed sensor.
In the colon model experiment, the tip of the scope was
steered into the sigmoid colon. The corresponding bending
sensor shows a decrease in strain gauge resistance as the
endoscope tip turns. The strain then increases when the tip is
straightened.
3) Inertial Measurement Unit: The estimation of position
from the IMU data requires double integration of the accelera-
tion signal, which leads to quadratic error propagation. Given
the fact that there is significant drift, it is not feasible to get
accurate 3D position solely from the inertial sensors [26].
Orientation detection, on the other hand, is much more accu-
rate, since it only requires a single integration step, and, e.g., a
Kalman filter to minimize gyroscope drift [27]. Insertion depth
of the endoscope can be estimated from the pressure sensor
outputs. As seen in Figure 5e, when the pressure module
passes the anal sphincter, all four sensors show a spike in
pressure. The sensor modules are placed on the scope shaft
on 5 cm intervals, so we know the insertion depth of the
scope shaft with a resolution of 5 cm. Since we can fairly
accurately estimate orientation and we know the insertion
depth of the endoscope, we can reconstruct the 3D trajectory
of the instrument. Since the proof-of-concept device only has
sensors in the scope tip, we demonstrated this by advancing
the endoscope in 5 cm intervals through a plastic tubing.
Figure 6a shows the setup with a 75 cm long tube with a
35 mm diameter representing an alpha loop formation.
IV. CONCLUSION
In this article, we proposed a multi-sensory system for
sensing contact pressure, instrument bending and orientation.
The system was designed and made using commercially
available sensors, and the sensors are configured and refined
for this specific application. A test bench system was built to
verify the design, and the sensors are characterized with test
bench. Finally, the design of the proposed sensing system was
validated with a colon phantom. The experiments prove that
the design is indeed viable for the use of colonoscopy. The
next steps include integrating the proposed sensor technology
into an colonoscope and validating the proposed with a more
advanced colon model and clinical trials. The goal is to
translate the proposed method for clinical use.
A. Challenges
Although the proposed system can provide the needed
sensing capability for the endoscope, there were some issues
with the design as well. Given the fact that the accelerometer
and gyroscope exhibit drift, it is not feasible to get accurate
3D position solely from the inertial sensors [26]. We used the
insertion depth data along with the orientation to reconstruct
the 3D trajectory of the scope. However, this approach assumes
that the tip is advanced the same amount each time, which is
not often the case in real colonoscopy. So the method only
provides a rough estimate of the trajectory of the scope. It is
also possible to add an actuator to control the advancement of
the scope as in [28]. This would enable high precision tracking
of the insertion depth. Another issue we found was the effect
of temperature variation to the bending sensor signal. Unlike
in the case of a colon phantom, the actual use case involves the
sensors to be exposed to substantial temperature changes, e.g.,
from room temperature of 20 ◦C to body temperature 37 ◦C.
Furthermore, the results show that the highest rate of change
happens right after the heating is turned on and the effect of
temperature is actually fairly small at body temperatures.
B. Translatability Into Clinical Practice
The current device is a proof-of-concept design, meant to
showcase the capabilities of such platform in an in vitro
PANULA et al.: MULTI-SENSORY SYSTEM FOR SPATIALLY AWARE COLONOSCOPY 983
setting. In order for the technology to be applicable for clinical
use, there are a couple of fundamental issues to be tackled.
Firstly, in the prototype device the sensors and tendons are
exposed in a way that would not be suitable for a clinical
version. A flexible silicone sleeve molded around the tip of the
instrument would provide protection for both the sensors and
the patient. Secondly, the tethering solution for the individual
sensor modules has to be iterated. Now all modules are
individually powered and accessed via FFC cables, which
results in a fairly large amount of cables. A possible solution
would include daisy-chaining of the modules into a single
bus with each module addressed individually with a unique
identifier. Moreover, the cables are routed externally, which is
not possible in a clinical version, which would require internal
routing.
C. Future Directions
Controlling the elongated flexible colonoscope to avoid
excessive contract pressure and looping is very challenging,
especially when the endoscopist can only rely only on cam-
era feedback. Each colon is different and even the most
experienced clinician can face difficulties in inserting the
colonoscope into the curved lumen. By introducing sensing
onto the scope will aid the user in navigating through the
colon and minimizing the risk on tearing and rupturing of
the colon during colonoscopy [3], [4]. In addition to reducing
the risk for life-threatening complications, we believe that
minimizing the pain experienced can lift the stigma associated
with colonoscopy, encouraging more people to get their
colonoscopy which is a proven method to minimizing the risk
of colorectal cancer [29], [30]. The proposed sensor system
could also be useful for training physicians, e.g., in a simulated
environment. Seeing which maneuvers produce excessive force
or are prone to leading to loop formation, can be valuable in
learning the best procedures. Altogether, the introduction of
additional sensors to colonoscopy opens up opportunities for
artificial intelligence (AI). AI has already shown remarkable
potential especially in the field of radiology, where it is used
as a complementary tool to aid the clinician. Currently in
colonoscopy AI is mostly used for detecting lesions from the
camera image [31], [32]. These approaches use convolutional
neural networks (CNN), which have been successfully applied
in image processing. However recent developments show
promise toward AI being used in navigating the instrument
and, e.g., assisting in biopsies [33]. A recent study proposed a
three-layer model of autonomy in colonoscopy, ranging from
manual robotic control to semi-autonomous operation, letting
the operator to focus on important tasks [10]. The increasing
levels of autonomy are an interesting direction of research
- yet they cannot be achieved with traditional endoscopes.
Accordingly, in the case of colonoscopy, we believe that rather
than fully automating the procedure, AI is best used as an
assistive tool by providing feedback to the user based on the
multi-sensory inputs. One possible tool could be a real-time
loop predictor that gives a warning when loop formation is
probable. This would give the endoscopist time to adjust the
manuever and prevent pain to the patient.
AUTHOR CONTRIBUTIONS
Conception and design: T.P., B.G., S.I, B.L., System devel-
opment: T.P. and B.G., Collection and assembly of data: T.P.,
Analysis and interpretation of the data: T.P., Drafting of the
article: T.P., M.K. and B.L., All authors read and approved the
final manuscript.
DATA AVAILABILITY
The data used and analysed during the current study is
available from the corresponding author on reasonable request.
CONFLICT OF INTEREST
The authors declare no conflict of interest.
REFERENCES
[1] S. Ravindran et al., “National census of U.K. endoscopy services in
2019,” Frontline Gastroenterol., vol. 12, no. 6, pp. 451–460, 2021.
[2] H. Kavin, F. Sinicrope, and A. H. Esker, “Management of perforation
of the colon at colonoscopy,” Amer. J. Gastroenterol., vol. 87, no. 2,
pp. 451–460, 1992.
[3] A. J. Loeve, P. Fockens, and P. Breedveld, “Mechanical analysis of
insertion problems and pain during colonoscopy: Why highly skill-
dependent colonoscopy routines are necessary in the first place· · · and
how they may be avoided,” Can. J. Gastroenterol., vol. 27, no. 5,
pp. 293–302, 2013.
[4] I. C. Roberts-Thomson and E. Teo, “Colonoscopy: Art or science?” J.
Gastroenterol. Hepatol., vol. 24, no. 2, pp. 180–184, 2009.
[5] A. Sivananthan, B. Glover, L. Ayaru, K. Patel, A. Darzi, and
N. Patel, “The evolution of lower gastrointestinal endoscopy: Where
are we now?” Ther Adv. Gastrointest. Endosc., vol. 13, Dec. 2020,
Art. no. 2631774520979591.
[6] S. D. Crockett, H. O. Cirri, R. Kelapure, J. A. Galanko, C. F. Martin, and
E. S. Dellon, “Use of an abdominal compression device in colonoscopy:
A randomized, sham-controlled trial,” Clin. Gastroenterol. Hepatol.,
vol. 14, no. 6, pp. 850–857, 2016.
[7] R. D. Pai, D. L. Carr-Locke, and C. C. Thompson, “Endoscopic evalu-
ation of the defunctionalized stomach by using ShapeLock technology
(with video),” Gastrointest. Endosc., vol. 66, no. 3, pp. 578–581, 2007.
[8] C. W. Teshima, S. Zepeda-Gómez, S. H. AlShankiti, and G. S. Sandha,
“Magnetic imaging-assisted colonoscopy vs conventional colonoscopy:
A randomized controlled trial,” World J. Gastroenterol., vol. 20, no. 36,
2014, Art. no. 13178.
[9] J. W. Leung et al., “Magnetic endoscope imaging (ScopeGuide) elu-
cidates the mechanism of action of the pain-alleviating impact of
water exchange colonoscopy-attenuation of loop formation,” J. Intervent.
Gastroenterol., vol. 2, no. 3, p. 142, 2012.
[10] J. W. Martin et al., “Enabling the future of colonoscopy with intelligent
and autonomous magnetic manipulation,” Nat. Mach. Intell., vol. 2,
no. 10, pp. 595–606, 2020.
[11] M. McCandless, A. Gerald, A. Carroll, H. Aihara, and S. Russo, “A soft
robotic sleeve for safer colonoscopy procedures,” IEEE Robot. Autom.
Lett., vol. 6, no. 3, pp. 5292–5299, Jul. 2021.
[12] M. N. Appleyard, C. A. Mosse, T. N. Mills, G. D. Bell, F. D. Castillo,
and C. P. Swain, “The measurement of forces exerted during
colonoscopy,” Gastrointest. Endosc., vol. 52, no. 2, pp. 237–240, 2000.
[13] S. Dogramadzi, G. Virk, G. Bell, R. Rowland, and J. Hancock,
“Recording forces exerted on the bowel wall during colonoscopy: In
vitro evaluation,” Int. J. Med. Robot. Comput. Assist. Surg., vol. 1, no. 4,
pp. 89–97, 2005.
[14] J. Choi and D. Drozek, “Detection of looping during colonoscopy using
bending sensors,” Open Med. Devices J., vol. 5, no. 1, pp. 1–7, 2013.
[15] A. O. Tal, J. Vermehren, and J. G. Albert, “Colon capsule endoscopy:
Current status and future directions,” World J. Gastroenterol., WJG,
vol. 20, no. 44, 2014, Art. no. 16596.
[16] H. Liang, Y. Guan, Z. Xiao, C. Hu, and Z. Liu, “A screw propelling
capsule robot,” in Proc. IEEE Int. Conf. Inf. Autom., 2011, pp. 786–791.
[17] L. Manfredi, L. Yue, J. Zhang, and A. Cuschieri, “A 4 DOFs variable
stiffness soft module,” in Proc. IEEE Int. Conf. Soft Robot. (RoboSoft),
2018, pp. 94–99.
984 IEEE TRANSACTIONS ON MEDICAL ROBOTICS AND BIONICS, VOL. 7, NO. 3, AUGUST 2025
[18] S. S. Vedaei and K. A. Wahid, “MagnetOFuse: A hybrid tracking
algorithm for wireless capsule endoscopy within the GI track,” IEEE
Trans. Instrum. Meas., vol. 71, pp. 1–11, 2022.
[19] T. Panula, T. Koivisto, M. Pänkäälä, T. Niiranen, I. Kantola, and
M. Kaisti, “An instrument for measuring blood pressure and assessing
cardiovascular health from the fingertip,” Biosens. Bioelectron., vol. 167,
Nov. 2020, Art. no. 112483.
[20] B. Maundy and S. J. Gift, “Strain gauge amplifier circuits,” IEEE Trans.
Instrum. Meas., vol. 62, no. 4, pp. 693–700, Apr. 2013.
[21] W. Li, M. Shen, A. Gao, G.-Z. Yang, and B. Lo, “Towards a snake-like
flexible robot for endoscopic submucosal dissection,” IEEE Trans. Med.
Robot. Bion., vol. 3, no. 1, pp. 257–260, Feb. 2021.
[22] R. Kozarek, D. Earnest, M. Silverstein, and R. Smith, “Air-pressure-
induced colon injury during diagnostic colonoscopy,” Gastroenterology,
vol. 78, no. 1, pp. 7–14, 1980.
[23] M. Kaisti et al., “Clinical assessment of a non-invasive wearable MEMS
pressure sensor array for monitoring of arterial pulse waveform, heart
rate and detection of atrial fibrillation,” NPJ Digit. Med., vol. 2, no. 1,
p. 39, 2019.
[24] L. Vicci, “Quaternions and rotations in 3-space: The algebra and its
geometric interpretation,” Dept. Comput. Sci., Univ. North Carolina
Chapel Hill, Chapel Hill, NC, USA, Rep. TR01-014, 2001.
[25] S. Shah, J. Brooker, C. Thapar, C. Williams, and B. Saunders, “Patient
pain during colonoscopy: An analysis using real-time magnetic endo-
scope imaging,” Endoscopy, vol. 34, no. 6, pp. 435–440, 2002.
[26] H. J. Luinge and P. H. Veltink, “Measuring orientation of human body
segments using miniature gyroscopes and accelerometers,” Med. Biol.
Eng. Comput., vol. 43, pp. 273–282, Apr. 2005.
[27] C. Kownacki, “Optimization approach to adapt Kalman filters for the
real-time application of accelerometer and gyroscope signals’ filter-
ing,” Digit. Signal Process., vol. 21, no. 1, pp. 131–140, 2011.
[28] W. Li, Y.-Y. Tsai, G.-Z. Yang, and B. Lo, “A novel endoscope
design using spiral technique for robotic-assisted endoscopy inser-
tion,” in Proc. IEEE/RSJ Int. Conf. Intell. Robots Syst. (IROS), 2020,
pp. 3119–3124.
[29] D. A. Lieberman et al., “Use of colonoscopy to screen asymptomatic
adults for colorectal cancer,” New England J. Med., vol. 343, no. 3,
pp. 162–168, 2000.
[30] A. Sonnenberg, F. Delco, and J. M. Inadomi, “Cost-effectiveness of
colonoscopy in screening for colorectal cancer,” Ann. Internal Med.,
vol. 133, no. 8, pp. 573–584, 2000.
[31] D. Banik, K. Roy, D. Bhattacharjee, M. Nasipuri, and O. Krejcar,
“Polyp-Net: A multimodel fusion network for polyp segmenta-
tion,” IEEE Trans. Instrum. Meas., vol. 70, pp. 1–12, 2020.
[32] G. Yue, S. Li, R. Cong, T. Zhou, B. Lei, and T. Wang, “Attention-
guided pyramid context network for polyp segmentation in colonoscopy
images,” IEEE Trans. Instrum. Meas., vol. 72, pp. 1–13, 2023.
[33] L. Introzzi, J. Zonca, F. Cabitza, P. Cherubini, and C. Reverberi,
“Enhancing human-AI collaboration: The case of colonoscopy,” Digest.
Liver Dis., vol. 56, no. 7, pp. 1131–1139, 2023.