PaperlayerView exact revision on arXiv

Source-linked AI summary

Prostate biopsy tracking with deformation estimation

Michael Baumann, Pierre Mozer, Vincent Daanen, Jocelyne Troccaz

arXiv:1107.1644v1cs.CVphysics.med-ph

TL;DR

Standard 2D ultrasound-guided prostate biopsy relies on systematic sampling, while clinicians may lack accurate guidance on target placement and core locations. The paper presents image-based 3D ultrasound tracking with rigid and elastic registration. Elastic registration improves target registration error across deciles, while real-time volume acquisition remains a constraint on convenient tracking updates.

  • Problem

    Under standard biopsy, clinicians may struggle to reach systematic targets accurately and core locations are only approximately known after the procedure.

    Source: §1.2 · p. 1

  • Method

    The system estimates rigid and residual elastic prostate motion using image registration, with a kinematic model of rectal probe motion and a local similarity measure.

    Source: §4.1 · p. 8, §4.4 · p. 9, §7 · p. 18

  • Results

    Elastic registration improves target registration error in all deciles by about at least one third, with a 1 mm to 2 mm average error decrease in the worst decile.

    Source: §6.3 · p. 17

  • Takeaways & Limitations

    The biopsy assistance system provides feedback on biopsy distribution and supports fusion with histology for diagnosis and therapy planning, as well as guidance toward MR-segmented targets.

    Source: §8 · p. 21

    Volume acquisition takes 0.5–5 s and registration takes 7–8 s, making conveniently fast tracking updates difficult because real-time TRUS acquisition remains a bottleneck.

    Source: §7.4 · p. 21

Abstract

from arXiv · show

Transrectal biopsies under 2D ultrasound (US) control are the current clinical standard for prostate cancer diagnosis. The isoechogenic nature of prostate carcinoma makes it necessary to sample the gland systematically, resulting in a low sensitivity. Also, it is difficult for the clinician to follow the sampling protocol accurately under 2D US control and the exact anatomical location of the biopsy cores is unknown after the intervention. Tracking systems for prostate biopsies make it possible to generate biopsy distribution maps for intra- and post-interventional quality control and 3D visualisation of histological results for diagnosis and treatment planning. They can also guide the clinician toward non-ultrasound targets. In this paper, a volume-swept 3D US based tracking system for fast and accurate estimation of prostate tissue motion is proposed. The entirely image-based system solves the patient motion problem with an a priori model of rectal probe kinematics. Prostate deformations are estimated with elastic registration to maximize accuracy. The system is robust with only 17 registration failures out of 786 (2%) biopsy volumes acquired from 47 patients during biopsy sessions. Accuracy was evaluated to 0.76$\pm$0.52mm using manually segmented fiducials on 687 registered volumes stemming from 40 patients. A clinical protocol for assisted biopsy acquisition was designed and implemented as a biopsy assistance system, which allows to overcome the draw-backs of the standard biopsy procedure.

1. Introduction

Standard 2D ultrasound-guided biopsies rely on systematic sampling because many prostate cancers are not visible on ultrasound, while clinicians have limited guidance on core placement and location. The paper motivates image-based 3D ultrasound tracking as a way to map biopsy locations and guide sampling, including toward non-ultrasound targets.

1.1. Prostate biopsies

  • Early- and mid-stage prostate carcinomas are mostly isoechogenic, so the standard protocol samples 10–12 cores systematically, accounting for the peripheral zone where about 70% of tumors develop.The needle guide aligns the trajectory with the ultrasound image plane, but the tumors may not be visible in that image.Source: §1.1 · p. 1

1.2. Clinical issues

  • The clinician may miss systematic targets without a constant visual reference, and acquired core locations are only approximately known after the intervention.Uncertainty about tumor location is described as a major reason prostate cancer therapy generally treats the entire gland.Source: §1.2 · pp. 1, 3

1.3. Proposed solutions

  • Reference-volume registration can project planned targets into control images and biopsy trajectories into the reference, supporting distribution checks and post-intervention analysis.The mapped trajectories can be combined with histology for therapy planning, and targets may include suspicious lesions identified in MR volumes.Source: §1.3 · p. 3
  • The proposed system uses purely ultrasound image registration to estimate prostate motion and deformation, with a probe-kinematics model rather than beam-tracking hardware.The authors also describe a clinical application for biopsy and cancer maps, non-ultrasound target guidance, and visualizing samples from previous sessions.Source: §1.3 · pp. 3–4

2. 3D ultrasound-based prostate tracking

The tracking approach uses coarse-to-fine registration, progressively increasing transformation complexity and image resolution. Its image-pyramid construction uses border extrapolation to preserve information as volumes are downsampled.

2.1. Coarse-to-fine strategies

  • Coarse-to-fine registration successively increases transformation degrees of freedom and image resolution to balance robust outcomes with computational efficiency.The first two registration steps use very coarse Gaussian-pyramid levels, while elastic registration descends to finer levels to estimate local details such as needle-insertion deformation.Source: §2.1 · p. 4, §2.3 · p. 5

2.2. Coarse-to-fine transformation spaces

  • The pipeline increases transformation degrees of freedom in stages because high-dimensional registration needs a good initial estimate to avoid local minima.It systematically explores a 3-DOF probe-kinematics model, locally refines promising minima in a 6-DOF rigid space, then estimates elastic deformation.Source: §2.2 · pp. 4–5

2.3. Coarse-to-fine image resolutions

  • Incomplete neighborhoods at beam borders cause information-containing volume to shrink exponentially with each downsampling, limiting the number of usable pyramid levels.The beam footprint is a section of a torus, so its Cartesian-space borders require a mask to distinguish image information from background.Source: §2.3 · p. 5
  • Information volume remains almost constant across pyramid levels with border extrapolation, whereas ordinary downsampling loses information at coarse levels.The method extrapolates border intensities before downsampling, averaging estimates from available neighboring directions and discarding voxels when fewer than one third of directions can be computed.Source: §2.3 · pp. 5–6

3. Rigid registration

Rigid registration is organized as parametric image-distance optimization, preceded by a constrained search over plausible probe positions. The best candidate then undergoes local rigid refinement before elastic estimation in the next stage.

3.1. Parametric optimization framework

  • The parametric framework searches for transformation parameters that optimize an image-distance measure, using the Pearson correlation coefficient in this approach.The paper also calls this measure cross correlation or simply the correlation coefficient.Source: §3.1 · p. 6

3.2. Global rigid registration with endorectal probe kinematics

  • The global search uses rectal-wall contact and sphincter constraints to define plausible probe positions relative to an approximate prostate ellipsoid.The model assumes the probe axis passes through a fixed point and explores five insertion depths in 5 mm steps, tilt angles up to 30°, and arbitrary axial rotation.Source: §3.2 · p. 7

3.3. Local rigid registration

  • The five probe positions with optimal DCC initialize local optimization in a 6-DOF rigid space; the transformation with minimal distance is retained for rapid refinement.The search seeks local minima near the initial estimates, leaving finer details to the subsequent elastic estimation step.Source: §3.3 · p. 8

4. Elastic registration

The framework estimates prostate deformation by variational image registration, adding inverse-consistency and elastic priors to address local misregistration in noisy ultrasound. Its shift-correlation measure handles intensity variation, and an iterative solver combines the registration terms with staged parameter settings.

4.1. Framework for non-linear registration

  • The variational framework estimates a diffeomorphic deformation ϕ(x) = x + u(x) by minimizing an image-distance objective supplemented with inverse-consistency and elastic energies.Straight minimization can fail amid noise, occlusion, and image imperfections, motivating the added priors.Source: §4.1 · p. 8

4.2. Inverse consistency constraints

  • Inverse consistency couples forward and backward deformation estimates, penalizing inconsistent inverse transformations and reducing the risk of local misregistrations.The estimates are coupled through alternating iterative optimization.Source: §4.2 · p. 8

4.3. Elastic regularization

  • Linear elastic regularization prevents strong deformations driven by non-physical fractional forces; the authors note the gland rarely deforms more than 10%.They did not investigate fluid regularization followed by linear elastic regularization, citing the low target registration error of rigid registration.Source: §4.3 · p. 9

4.4. Image distance measure

  • The shift-correlation measure tolerates low-frequency intensity shifts while requiring less data for statistical power than two-dimensional Pearson correlation.SSD performs poorly with frequent local intensity changes; the invariant frequency range depends on Gaussian-kernel standard deviation, while multiresolution filtering yields band-pass behavior.Source: §4.4 · pp. 9–10

4.5. Solver

  • The solver uses an iterative two-step fixed-point scheme, alternating forward and backward force updates and solving the resulting PDEs with Red-Black Gauss-Seidel relaxation and full multigrid.The full multigrid strategy is used to obtain linear computational complexity; iterations continue to convergence at each resolution level.Source: §4.5 · pp. 10–11

4.6. Parameter settings

  • Parameter settings disable compressibility initially and enable it in a second run, with ν changing from 0 to 0.48.The final displacement is limited to a norm below 1 to improve numerical stability; E and δt are set to 0.5.Source: §4.6 · p. 11

5. Clinical application

The clinical application builds on 3D TRUS tracking to support biopsy planning, acquisition, and post-biopsy mapping. Its workflow aims to add clinical accuracy and an intuitive interface without adding logistical or interventional overhead to standard practice.

5. Clinical application

  • The application aims to combine clinical accuracy with an intuitive interface and avoid added logistical or interventional overhead.It supports non-ultrasound targets, visualization of prior samples, and intra- and post-interventional biopsy maps.Source: §5 · pp. 11–12

5.1. Panoramic volume as anatomical reference

  • Three elastically registered ultrasound volumes are compounded into a panorama because one volume typically misses parts of the gland.The clinician defines the bounding ellipsoid after acquiring the first volume.Source: §5.1 · p. 12

5.2. Biopsy planning

  • Clinicians can project MR-defined targets into the ultrasound reference, while the viability of registering maps across repeated biopsy sessions remains unproven.The paper notes that time between sessions, organ changes, and imaging-hardware differences may complicate repeat-session registration.Source: §5.2 · p. 13

5.3. Biopsy acquisition

  • During acquisition, 2D ultrasound guides needle placement; triggered 3D volumes are registered to project the puncture path into the reference anatomy.After sampling, the needle is segmented when present and projected into the reference anatomy to show its position relative to previous biopsies.Source: §5.3 · p. 14

5.4. Multi-modal biopsy and cancer maps

  • Gleason scores are color-coded in biopsy maps, and accurate sample localization permits cancer-map projection onto an MR planning volume.The maps provide an overview of cancer distribution for diagnosis and treatment planning.Source: §5.4 · p. 14

6. Experiments and results

The experiments assess the kinematic model, registration accuracy, and elastic deformation estimation on biopsy-session ultrasound volumes. Elastic registration reduced target registration error relative to rigid registration, while the kinematic model efficiently initialized local registration.

6. Experiments and results

  • The evaluation used real patient images acquired during biopsy sessions, with interventional volumes captured after each biopsy shot while the needle remained inside.The protocol otherwise followed standard clinical practice; 687 registrations from 40 patients were used for the accuracy study.Source: §6 · p. 15, §6.2 · p. 16

6.1. Kinematic model

  • Of 786 biopsy volumes from 47 patients, 769 (97.8%) were classified as valid and 17 (2.2%) as failures; failures involved insufficient prostate information in the volume.The reported causes included inadequate ultrasound depth or probe pressure, partial rectal contact, lateral coverage, and an incomplete panorama; failures were not caused by patient movement.Source: §6.1 · p. 15
  • The kinematic model’s mean transducer-origin distance was 2.0±1.5 mm, while its mean angular distance was 9.7±5.1° from local rigid-registration estimates.Angular predictions were less accurate, in part because the rotational search used 15° steps and the fixed-point model of rectal-sphincter constraints was simplistic.Source: §6.1 · pp. 15–16

6.2. Target registration error

  • 0.8±0.5 mm TRE after deformation estimation versus 1.4±0.8 mm TRE after rigid registration; worst-decile error fell from 3.0 mm to 2.0 mm.Average computation time was about 7 s for deformation estimation versus about 2 s for rigid registration.Source: Table 1 · p. 17, §6.2 · p. 17

6.3. Rigid vs. elastic registration

  • Elastic registration improved TRE in every decile by at least one third, with stronger deformations observed in about 20% to 30% of analyzed volumes.The largest absolute benefit was in the worst decile, where average error decreased by 1 mm to 2 mm.Source: §6.3 · p. 17

7. Discussion

The discussion finds that elastic registration modestly improves tracking error, but MR–TRUS fusion and human trajectory reproduction remain important error sources. The system’s 3D approach is robust to some patient motion, yet acquisition and registration times constrain rapid target guidance.

7. Discussion

  • The system combines motorized 3D TRUS acquisition, a rectal-probe kinematic model, and rigid plus residual elastic motion estimation.Its clinical application supports biopsy and cancer maps, MR–TRUS target projection, and guidance toward non-ultrasound targets.Source: §7 · p. 18

7.1. Target registration error

  • 3.6 mm total RMS error (83% target-hit probability) with elastic registration, versus 3.8 mm (81%) with rigid registration; further tracking-error reduction yields limited hit-rate gains.Reducing tracking error threefold to 0.25 mm RMS would increase the hit rate by only 2.5%; MR–TRUS fusion error is a priority for future improvement.Source: §7.1 · p. 20
  • For biopsy maps, the estimated total error is 1.34 mm RMS, while therapy-planning projection adds another error source.The discussion identifies MR–TRUS fusion and the operator’s ability to reproduce a suggested trajectory as important error sources to minimize.Source: §7.1 · p. 20

7.2. Rigid vs. elastic tracking

  • Elastic tracking reduces total TRE but takes three to four times longer; it is preferred for biopsy maps and cross-session panorama registration.Under the stated assumptions, it reduces the risk of missing an MR target by about ten percent, while other error sources predominate.Source: §7.2 · p. 20
  • Elastic registration reduces rigid-registration error by almost 50% in deciles 1–9 and 30% in the worst decile, without degrading rigid results.The authors describe the non-linear method as robust across all deciles.Source: §7.2 · p. 20

7.3. Patient motion

  • In the experiments, registration failures were attributed to insufficient prostate information in the tracking volume, not patient movement.The system can still be affected by movement during 3D acquisition, which distorts volumes, or during registration when guidance is underway.Source: §7.3 · pp. 20–21

7.4. 3D TRUS vs. 2D TRUS based tracking

  • 3D volume acquisition takes 0.5–5 seconds and registration takes 7–8 seconds, limiting convenient rapid target-position updates.The authors identify real-time TRUS volume acquisition as the current blocking issue for fast updates.Source: §7.4 · p. 21
  • Compared with 3D TRUS tracking, 2D TRUS tracking is less accurate and more sensitive to patient movement, with a smaller capture range.The discussion also notes lateral biopsies may be registered incorrectly when only a small part of the image contains the gland; real-patient reconstruction accuracy is unknown.Source: §7.4 · p. 21

8. Conclusion

The biopsy assistance system provides clinicians with biopsy-distribution maps that support histology review, treatment planning, and guidance to predefined targets. Its clinical use has extended beyond targeting, while real-time target guidance remains unavailable.

8. Conclusion

  • The clinical application gives immediate, precise feedback on biopsy distribution and supports histology fusion, prior-session map projection, and guidance to MR-segmented targets.It minimally alters the standard procedure and adds no significant logistical complexity.Source: §8 · p. 21
  • Cancer maps can support experimental focal-therapy strategies, including targeted follow-up biopsies around positive samples and projection into MR images to determine tumor shape.Supplementary targeted biopsies can also help confirm local absence of cancer.Source: §8 · p. 21
  • Guidance to predefined targets could increase biopsy sensitivity; MR-suspicious lesions can be projected into tracking images, and repeat sessions can avoid resampling examined tissue to improve gland coverage.The system cannot yet guide targets in real time.Source: §8 · p. 21
  • The industrial system has been used in more than 200 biopsy sessions and reportedly works reliably, while protocols for in-depth assessment of its clinical potential are being prepared.A preliminary study found that visual feedback about sample distribution led to a significant learning effect.Source: §8 · p. 21
  • The planned assessment includes the sensitivity and specificity of MR-targeted biopsies, particularly after negative initial biopsies when cancer suspicion remains.The authors also identify US-based tracking as a possible aid to accurate therapeutic-instrument positioning in prostate interventions.Source: §8 · p. 22
Loading 1107.1644v1…