Endoscopic Expansion Techniques

• Endoscopic expansion is a suite of techniques that enlarge device capabilities or generate synthetic images to improve performance in confined environments.
• Physical expansion methods, such as bellows-driven scissor-struts and toroidal chambers, optimize contact mechanics and traction for enhanced maneuverability.
• Synthetic expansion via diffusion-based models like Polyp-Gen addresses limited annotated data by producing realistic and diverse endoscopic imagery for CAD systems.

Searching arXiv for the papers on arXiv and closely related work to ground the article in current literature. Endoscopic expansion denotes a family of techniques that enlarge, adapt, or augment endoscopic capability under constrained anatomical or data conditions. In recent literature, the term spans two technically distinct operations: physical expansion of endoscopic devices to regulate contact, traction, field of view, or steering within a lumen, and synthetic expansion of endoscopic image datasets to support CAD and ADS training when real annotated data are scarce or privacy-sensitive (Liu et al., 28 Jan 2025, Du et al., 2024, Consumi et al., 2022, McCandless et al., 3 Nov 2025). Across these settings, expansion is not merely geometric scaling; it is coupled to control, contact mechanics, image realism, and downstream task performance.

1. Conceptual scope and operating regimes

Physical endoscopic expansion is used to match local anatomy, stabilize propulsion, increase normal force against surrounding tissue, or enlarge the optical window. In the size-adaptable robotic endoscope, expansion is achieved by four bellows-driven scissor-strut units that push tracks radially outward, changing tip diameter and frictional drive conditions (Du et al., 2024). In SoftSCREEN, two inflatable toroidal chambers displace six surrounding tracks radially outward, modulating contact and traction while maintaining a 360° soft ring contact (Consumi et al., 2022). In the steerable balloon cardioscope, axial wall-thickness variation causes low-pressure optical-window expansion to precede higher-pressure bending, yielding a deliberate “expansion-then-bend” sequence (McCandless et al., 3 Nov 2025).

Synthetic endoscopic expansion addresses a different bottleneck: the limited availability of annotated endoscopic imagery for ADS development. Polyp-Gen formulates dataset expansion as full-automatic diffusion-based image generation, producing both polyp and normal images without requiring expert-drawn masks at sampling time (Liu et al., 28 Jan 2025).

System	Expansion variable	Representative result
Polyp-Gen (Liu et al., 28 Jan 2025)	Dataset size via synthetic endoscopic images	FID $=13.817$, IS $=3.454$ on LDPolypVideo
Adaptable robotic endoscope (Du et al., 2024)	Tip diameter via bellows and scissor-struts	$53\%$ expansion rate; $F_{\max}=3.89\,\rm N$
SoftSCREEN (Consumi et al., 2022)	Capsule diameter via toroidal chamber inflation	Diameter change from $65$ mm to $94$ mm at $P=16$ kPa
Steerable balloon endoscope (McCandless et al., 3 Nov 2025)	Optical diameter and bending angle via one inflation input	$D$ from $4.63$ mm to $11$ mm; $=3.454$0 up to $=3.454$1

A plausible implication is that “expansion” functions as a systems-level design principle rather than a single mechanism: it can target morphology, contact mechanics, visual workspace, or training data distribution depending on the endoscopic task.

2. Diffusion-based expansion of endoscopic datasets

Polyp-Gen is presented as the first full-automatic diffusion-based endoscopic image generation framework, with the explicit goal of synthesizing high-quality, diverse endoscopic images for enlarging limited, privacy-sensitive GI datasets used in CAD and ADS pipelines (Liu et al., 28 Jan 2025). Its backbone is a text- and image-conditioned latent diffusion model based on Stable Diffusion, factorized into a VAE encoder/decoder pair $=3.454$2 and a U-Net denoiser $=3.454$3 operating in latent space. The latent encoding and DDPM-style noising process are written as

$=3.454$4

$=3.454$5

with reverse objective

$=3.454$6

The central training contribution is a spatial-aware diffusion scheme designed to enhance the structural context of polyp boundary regions. Instead of relying on expensive pixel-wise masks, Polyp-Gen uses bounding boxes together with a boundary-enhanced pseudo mask (BPM): for each real polyp bounding box $=3.454$7, a random inscribed convex polygon is generated within $=3.454$8 to serve as training mask $=3.454$9. For the “Normal” prompt, either a random mask on a non-polyp frame or a random mask outside $53\%$0 on a polyp frame is used, so the model learns both polyp$53\%$1normal and normal$53\%$2polyp translations. Given the partially masked image $53\%$3, its latent $53\%$4, and resized mask $53\%$5, the mask-conditioned objective is

$53\%$6

A lesion-guided loss further weights masked pixels,

$53\%$7

and the final objective is

$53\%$8

At sampling time, Polyp-Gen removes the need for expert-specified mask placement through hierarchical retrieval-based sampling. A non-polyp reference database is first constructed by converting each real polyp frame $53\%$9 with ground-truth mask $F_{\max}=3.89\,\rm N$0 into a “clean” normal image,

$F_{\max}=3.89\,\rm N$1

and storing $F_{\max}=3.89\,\rm N$2 pairs. For a query non-polyp frame $F_{\max}=3.89\,\rm N$3, DINOv2 global pooled features $F_{\max}=3.89\,\rm N$4 and dense patch features $F_{\max}=3.89\,\rm N$5 are extracted. Stage 1 retrieves top-$F_{\max}=3.89\,\rm N$6 nearest references via

$F_{\max}=3.89\,\rm N$7

Stage 2 performs local nearest-neighbor matching inside reference mask regions, clusters the resulting matched patches via DBSCAN with radius $F_{\max}=3.89\,\rm N$8, and uses the largest cluster’s bounding rectangle as the proposed mask $F_{\max}=3.89\,\rm N$9. The final synthetic polyp image is then generated by

$65$0

Quantitatively, Polyp-Gen reports state-of-the-art generation quality on both in-domain and zero-shot settings. On LDPolypVideo it achieves FID $65$1 and IS $65$2, compared with $65$3 for ControlPolypNet, $65$4 for Blended Latent Diffusion, $65$5 for CondPolypDiff, and $65$6 for Polyp-DDPM. On Kvasir-Seg in zero-shot evaluation it achieves FID $65$7 and IS $65$8, again outperforming the listed baselines (Liu et al., 28 Jan 2025). For downstream detection, adding $65$9 synthetic images to $94$0 real frames increases CenterNet performance from AP $94$1, F1 $94$2 to AP $94$3, F1 $94$4. A reader study with two endoscopists found that $94$5–$94$6 of Polyp-Gen outputs were mistaken for real polyps, compared to $94$7–$94$8 for prior diffusion/GAN methods. The stated limitations are higher zero-shot FID than in-domain FID, occasional mask-proposal mismatches when anatomy differs drastically from the database, and the possibility of integrating anatomical priors or an end-to-end mask proposer in future work.

3. Pneumatic-mechanical tip expansion for shape-conforming propulsion

The adaptable robotic endoscope implements expansion as a local shape-conforming propulsion mechanism for colonoscopy (Du et al., 2024). Its expandable tip is a cylindrical assembly of four flexible tracks driven by a single worm gear. Four identical expansion units, each comprising a flexible bellows and scissor-struts, are placed symmetrically $94$9 apart around the tip frame; each unit pushes one track radially outward so that the tracks bear on the lumen wall.

The flexible bellows are fabricated from dual-shore silicone: a stiff blue half, Shore A 30, Smooth-On Mould-Star, and a soft yellow half, Shore 00-30, Ecoflex 00-30. Fabrication follows a two-stage molding process in which blue silicone is poured into molds A and B, cured, sliced, reassembled, and then filled with yellow silicone. The prototype is built at $P=16$0 scale with nominal unpressurized tip diameter $P=16$1 mm, and the strut length and bellows stroke are selected to achieve up to $P=16$2 mm radial excursion.

Within the working pressure range from $P=16$3 mbar to $P=16$4 mbar, the bellows exhibit nearly linear force and stroke behavior: $P=16$5 From the reported data, $P=16$6 and $P=16$7 at $P=16$8 mbar. The tip expansion rate is defined as

$P=16$9

that is, a $D$0 expansion rate. Static frictional drive force generated by $D$1 tracks in contact is modeled as

$D$2

where $D$3 is the track-wall friction coefficient, $D$4 is equivalent radial stiffness of one track, $D$5 is pipe inner diameter, and $D$6 is worm-gear diameter. This formalizes the basic rationale of the design: expansion increases normal force, thereby increasing available static friction and propulsion.

The external drive system uses a NEMA 17 stepper motor controlled by Arduino UNO and a TB6600 driver to deliver torque $D$7 and angular speed $D$8 to a rigid worm gear through a flexible shaft. The worm engages the toothed tracks and generates thrust $D$9 and tip speed $4.63$0. The paper gives force and torque balance expressions,

$4.63$1

with

$4.63$2

and a closed-form relation linking motor torque $4.63$3 to thrust $4.63$4. Motor speed is approximated by

$4.63$5

and tip speed by

$4.63$6

where $4.63$7 is the transmission ratio.

Experimental locomotion tests use straight acrylic tubes of inner diameters $4.63$8, $4.63$9, $11$0, and $11$1 mm, with lining surfaces approximating $11$2, $11$3, and $11$4. Propelling force is measured with a FUTEK LSb201 load cell, and linear speed is measured over a fixed distance for motor speeds of $11$5, $11$6, $11$7, $11$8, and $11$9 rpm. On smooth acrylic, propelling force rises from approximately $=3.454$00 N at $=3.454$01 rpm to approximately $=3.454$02 N at $=3.454$03 rpm, then decreases at higher angular velocity because of shaft slippage. On artificial bowel tissue, the peak propelling force is $=3.454$04 N at $=3.454$05 rpm, and the average maximum speed is $=3.454$06 mm/s at $=3.454$07 rpm on wet tissue. On foam, peak propelling force reaches $=3.454$08 N at $=3.454$09 rpm, and maximum linear speed at $=3.454$10 rpm is approximately $=3.454$11 mm/s. The reported interpretation is that the prototype realizes shape adaptation in order to obtain more propulsion, while the paper identifies traction-force relationships, structural optimization, and miniaturization as open problems (Du et al., 2024).

4. Toroidal chamber expansion in the SoftSCREEN capsule

SoftSCREEN approaches endoscopic expansion through a tethered soft shapeshifting capsule robot based on eversion navigation (Consumi et al., 2022). The device consists of three concentric layers: an inner rigid cylindrical chassis that houses a single brushless DC motor coupled to a $=3.454$12 gearbox and worm-gear drive; six elastic toothed tracks arranged uniformly around the chassis circumference; and two independently inflatable toroidal chambers made of DragonSkin 20 A silicone with SLIDE™ STD low-friction additive. The toroids wrap around the chassis and pass through the track loops.

The lateral-flange profile used to secure each toroid to the chassis was selected over a central-flange design because FEA showed $=3.454$13 higher axial shear stiffness once inflated. The uninflated capsule diameter is approximately $=3.454$14 mm, which would scale to approximately $=3.454$15 mm in vivo, and when both chambers are inflated to the maximum tested pressure of approximately $=3.454$16 kPa, the envelope expands to fit lumen diameters up to approximately $=3.454$17 mm in the $=3.454$18 demonstrator. The toroid cross-section is circular with mean radius $=3.454$19 mm and wall thickness $=3.454$20 mm.

The chambers are modeled as a first-order Ogden hyperelastic material with strain-energy density

$=3.454$21

and, for $=3.454$22, the paper reports $=3.454$23 MPa and $=3.454$24. In toroidal geometry under internal pressure $=3.454$25, the pressure-stretch relation is treated through a generalized Laplace-type formula and numerically inverted in FEA. Under inflation to $=3.454$26 kPa, ANSYS™ 2019 simulation yields maximum radial expansion $=3.454$27 mm for each chamber and peak von Mises stress $=3.454$28 MPa, described as within safe limits for DragonSkin 20 A. Near working pressures, the experimental expansion law is reported as quasi-linear: $=3.454$29

Mechanically, inflation displaces the surrounding tracks radially outward. The normal force per track is approximated by

$=3.454$30

with $=3.454$31 N/mm, and the total traction force is

$=3.454$32

where $=3.454$33. The measured traction force rises from approximately $=3.454$34 N at $=3.454$35 kPa to approximately $=3.454$36 N at $=3.454$37 kPa against a silicone-wall phantom, confirming a nearly linear $=3.454$38 trend. In the prototype, inflation is open-loop via two FESTO VPPX regulators, with typical setpoints of $=3.454$39 kPa, $=3.454$40 kPa, and $=3.454$41 kPa.

Locomotion remains decoupled from inflation at the actuation level. Neglecting slip, the eversion kinematic relation is

$=3.454$42

with worm-thread pitch $=3.454$43 mm. Reported forward and reverse speeds in rigid pipes of diameters $=3.454$44, $=3.454$45, and $=3.454$46 mm are $=3.454$47–$=3.454$48 mm/s and $=3.454$49–$=3.454$50 mm/s, respectively. Inside a supported soft phantom of diameter approximately $=3.454$51 mm, maximum reverse speed is $=3.454$52 mm/s and maximum forward speed is $=3.454$53 mm/s, while in a collapsed phantom the reported values are approximately $=3.454$54 mm/s forward and $=3.454$55 mm/s reverse. The system is therefore presented as combining self-centering, adjustable normal load, and reliable propulsion in varying lumen diameters (Consumi et al., 2022).

5. Balloon expansion as a coupled optical and steering mechanism

The steerable balloon endoscope for robot-assisted transcatheter intracardiac procedures uses a single inflation input to independently control two outputs: balloon diameter, corresponding to field-of-view diameter, and balloon bending angle, enabling precise working-channel positioning (McCandless et al., 3 Nov 2025). The balloon body is made from Ecoflex 00-45 Near Clear, the proximal catheter tube is Pebax 5533 with $=3.454$56 mm wall thickness and $=3.454$57 mm inner diameter, and the working channel is silicone tubing with $=3.454$58 mm inner diameter and $=3.454$59 mm wall thickness.

The balloon wall is divided into six thickness segments $=3.454$60 and three section lengths $=3.454$61. Thin regions $=3.454$62 and $=3.454$63–$=3.454$64 mm form the optical face and distal neck and inflate first. Intermediate thickness $=3.454$65 mm and thick section $=3.454$66 mm form the steerable hinge, with $=3.454$67 slightly thinner than $=3.454$68 on the opposite side. Section lengths $=3.454$69 mm and $=3.454$70 mm set the lever arm and maximum bending angle $=3.454$71. The intended mechanical sequence is explicit: at low pressure only the thin optical face inflates; above a threshold pressure, the thinner hinge side expands more than the opposite side, generating net bending.

Under thin-walled membrane theory, local hoop stress is

$=3.454$72

and with linearized elastic modulus $=3.454$73, the strain and radius change satisfy $=3.454$74 and $=3.454$75. Because thickness varies axially and circumferentially, the thinner side undergoes greater expansion under the same pressure. The resulting bending moment is expressed as

$=3.454$76

and curvature $=3.454$77 yields tip deflection angle $=3.454$78. The authors note that the practical design was characterized empirically rather than through a full analytical solution.

Empirical characterization with saline infusion volumes from $=3.454$79 to $=3.454$80 mL shows the piecewise decoupling of diameter and steering. The field-of-view diameter $=3.454$81 grows from approximately $=3.454$82 mm in the collapsed state to $=3.454$83 mm by approximately $=3.454$84 mL, then plateaus at $=3.454$85 mm for approximately $=3.454$86–$=3.454$87 mL. Tip deflection $=3.454$88 remains approximately $=3.454$89 for $=3.454$90 mL, then rises roughly linearly to approximately $=3.454$91 at $=3.454$92 mL without a tool, or approximately $=3.454$93 with a $=3.454$94m wire inserted. This behavior confirms the intended expansion-then-bend sequence.

The system also incorporates image-based closed-loop control of bending angle. The on-board camera views the working-channel opening and red-dyed background; image processing segments the channel as a bright region, counts pixels inside and outside the channel contour, and computes

$=3.454$95

A fourth-order polynomial calibration $=3.454$96 with $=3.454$97 maps pixel ratio to tip angle. The control error $=3.454$98 drives a multi-threshold bang-bang controller commanding syringe-motor speed: $=3.454$99 Reported performance includes optical-window diameter from $53\%$00 mm to $53\%$01 mm, steering angle from $53\%$02 to approximately $53\%$03 without a tool or approximately $53\%$04 with a wire, average tip angular velocity of approximately $53\%$05s for a $53\%$06 step, settling time $53\%$07 s with overdamped response, steady-state error $53\%$08 during tool insertion and removal, no measurable overshoot for large steps, and sub-degree repeatability for $53\%$09 steps. The camera resolves Group 1 of the 1951 USAF chart, corresponding to $53\%$10 mm line pairs, and a $53\%$11 mm bull’seye target can be centered in the working channel under water.

6. Comparative principles, misconceptions, and unresolved problems

Across these studies, expansion is consistently used to shape the interaction between an endoscopic system and its operating environment, but the controlled variable differs sharply by application. In Polyp-Gen, expansion modifies the effective training distribution by generating realistic and diverse endoscopic images; the critical issue is preservation of polyp boundary structure and plausible lesion localization (Liu et al., 28 Jan 2025). In the adaptable robotic endoscope and SoftSCREEN, expansion regulates normal force and traction against the lumen wall, thereby affecting propulsion, stability, and shape adaptation (Du et al., 2024, Consumi et al., 2022). In the steerable balloon cardioscope, expansion first enlarges the optical workspace and then produces directional control through differential compliance, with closed-loop image feedback compensating for perturbations during tool manipulation (McCandless et al., 3 Nov 2025).

A common misconception is to treat endoscopic expansion as equivalent to simple diameter increase. The cited systems indicate otherwise. In the balloon cardioscope, the primary low-pressure output is field-of-view enlargement, while higher-pressure inflation yields bending rather than further substantial diameter growth (McCandless et al., 3 Nov 2025). In the adaptable robotic endoscope and SoftSCREEN, expansion is only useful insofar as it improves contact mechanics, because increased diameter must translate into increased normal force and frictional traction to support locomotion (Du et al., 2024, Consumi et al., 2022). In Polyp-Gen, expansion is not geometric at all; it refers to dataset enlargement under a realism-diversity constraint, where mask placement and lesion-boundary fidelity are central (Liu et al., 28 Jan 2025).

The principal limitations are also domain-specific. Polyp-Gen reports residual domain shift in zero-shot generation, with zero-shot FID higher than in-domain FID and occasional mismatches in retrieval-based mask proposal when anatomy differs drastically from the database (Liu et al., 28 Jan 2025). The adaptable robotic endoscope identifies the need for further work on the relationship between propelling force and traction force, structural optimization, and miniaturization (Du et al., 2024). SoftSCREEN highlights challenges in downsizing the motor, gearbox, worm drives, inflation tubing, and visualization package for a $53\%$12 clinical device (Consumi et al., 2022). The steerable balloon endoscope demonstrates precise angle control and adequate optical resolution, but its design is task-specific and relies on empirical characterization of the pressure-volume-angle relationship rather than a complete analytical model (McCandless et al., 3 Nov 2025).

Taken together, these results suggest that endoscopic expansion is best understood as a constrained design and control problem: one seeks to enlarge a useful operational quantity, whether data support, contact envelope, traction, field of view, or bending authority, without sacrificing anatomical plausibility, mechanical stability, or downstream performance.