Pith. sign in

REVIEW 4 major objections 5 minor 1 references

Improving Deformable Image Registration Accuracy through a Hybrid Similarity Metric and CycleGAN Based Auto-Segmentation

T0 review · 4 major / 5 minor · reviewed 2026-08-12 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read Auto-segmented contours from CycleGAN-corrected synthetic CT markedly improve CT-to-CBCT deformable registration, nearly matching expert contours.

desk verdict A legitimate integration-and-evaluation study for CBCT-guided pelvic DIR, with prostate improvements supported by fiducials, but the all-organ claim is undermined by a scoring asymmetry and a bladder result that goes the wrong way. read the letter →

arxiv 2411.16992 v1 pith:4X3UUKSF submitted 2024-11-25 physics.med-ph cs.CV

Keyur D. Shah , James A. Shackleford , Nagarajan Kandasamy , Gregory C. Sharp This is my paper · ORCID
classification physics.med-phcs.CV
keywords DeformableimageregistrationAdaptiveradiationtherapyCycleGANAuto-segmentationHybridsimilaritymetricPoint-to-distanceCT-CBCTU-Net
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

Deformable image registration between planning CT and daily cone-beam CT struggles in low-contrast soft tissue because pure intensity matching has no reliable landmarks there. This paper argues that adding a point-to-distance penalty on organ surfaces to the registration cost fixes most of that loss, and that the organ contours feeding that penalty can be produced automatically rather than by hand. The pipeline translates CBCT slices into synthetic CT with a CycleGAN, segments prostate, bladder, and rectum with a 3D U-Net, and uses those contours in a hybrid metric during B-spline registration. On a 14-case pelvic validation set the auto-contour workflow raised the prostate overlap score from 0.61 with intensity-only registration to 0.82, versus 0.89 with expert contours, and cut fiducial marker separation from 8.95 mm to 4.07 mm. If it holds, adaptive radiotherapy can get structure-guided registration without the manual-contouring step that currently slows on-line treatment adaptation.

What carries the argument

The load-bearing object is the point-to-distance (PD) penalty added to the B-spline registration cost. For each structure the penalty sums, over contour points $\pi_n$ on the fixed image, the value of an unsigned distance map at the transformed point in the moving image, weighted by $\eta_n$: $$C = \sum_{T(\$\theta$)\in\$\Omega$}\Psi(f,m) + \$\lambda$ S(\nu) + \sum_{n=1}^{N}\eta_n \sum_{\pi_n\in\Pi_n}|d_{\text{map}}(\pi'_n)|.$$ Minimizing this cost with L-BFGS-B forces the moving contour onto the fixed contour even where intensities give no gradient. The paper's contribution is to feed this term automatically: a 2D CycleGAN generates synthetic CT with better soft-tissue contrast from CBCT, and a 3D U-Net segments the three pelvic organs on those sCT images, so the PD term no longer depends on manual contouring. The cycle-consistency and identity losses in the CycleGAN are what keep the translated images anatomically faithful enough for segmentation to be meaningful.

What would settle it

Take a planning CT with expert contours, apply a known deformation to produce a synthetic CBCT, and run all three workflows; if the CycleGAN pipeline aligns true anatomy, its registration error against the known displacement should track the Expert PD result, and any large divergence would show the auto-contours are pulling the metric toward synthetic boundaries.

Watch

Extended reading notes

Core claim

The paper's central claim is that the hybrid similarity metric—intensity similarity plus a weighted point-to-distance (PD) term computed from organ boundaries—is what lifts CT-CBCT registration accuracy, and that the boundaries can be machine-drawn rather than expert-drawn. In the reported prostate results, adding CycleGAN-based auto-contours improved DSC from 0.61 ± 0.18 to 0.82 ± 0.13, reduced 95% Hausdorff distance from 11.75 mm to 4.86 mm, and reduced fiducial separation from 8.95 mm to 4.07 mm; expert manual contours yielded 0.89 ± 0.05, 3.27 mm, and 4.11 mm, respectively. The bladder and rectum also improved on DSC, though the bladder's 95% Hausdorff distance with auto-contours was not statistically different from intensity-only registration. The authors conclude that AI-based synthetic CT plus auto-segmentation can replace manual contours in the PD metric and that the hybrid metric is the main source of the accuracy gain.

Load-bearing premise

The load-bearing premise is that the CycleGAN's synthetic CT images show the patient's real anatomy rather than plausible-looking artifacts, because the auto-contours are only useful if they trace true organ boundaries; if the sCT fabricates or blurs structures, the point-to-distance penalty will align to synthetic shapes and the reported accuracy gains would not reflect real anatomy.

Editorial extensions

If this is right

  • Online adaptive radiotherapy could use the auto-contour pipeline to drive structure-guided registration without a manual contouring step.
  • The roughly 55% reduction in prostate fiducial separation (8.95 mm to 4.07 mm) supports the possibility of tighter planning margins and less healthy-tissue dose.
  • Because the PD term is structure-agnostic, the same hybrid cost should transfer to other organs and imaging modalities whenever a segmentation model is available.
  • Because the CycleGAN requires only unpaired CT and CBCT volumes, the sCT step can be reproduced at institutions that lack paired training data.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • Beyond the paper: the near-identical fiducial separation between CycleGAN PD (4.07 mm) and Expert PD (4.11 mm), despite a DSC/HD gap, suggests auto-contours are centered correctly but noisier in shape; a per-contour surface-error map would test this directly.
  • Beyond the paper: the bladder's non-significant 95% HD improvement hints that a single PD weight per structure may be too coarse for hollow, volume-changing organs; varying $\eta_n$ per organ could recover the expert-level boundary accuracy.
  • Beyond the paper: deliberately corrupting auto-contours (erosion, dilation, or random slice replacement) and measuring the resulting registration error would map how much contour quality the PD metric tolerates before it loses to intensity-only registration.
  • Beyond the paper: since the architecture is modality-agnostic, retraining on MR-CBCT or PET-CBCT pairs is a natural next step, with the same unpaired-translation advantage.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 5 minor

Summary. The paper proposes an improved deformable image registration (DIR) workflow for CT-to-CBCT registration in adaptive radiotherapy. A hybrid similarity metric combines intensity matching with a point-to-distance (PD) structural term. A CycleGAN generates synthetic CT (sCT) from CBCT to improve soft-tissue contrast, and a 3D U-Net auto-segments the prostate, bladder, and rectum on the sCT. Three workflows are compared: intensity-only DIR (No PD), DIR using U-Net contours on sCT (CycleGAN PD), and DIR using expert manual contours (Expert PD). On a validation cohort of 7 patients, the paper reports improved DSC and 95% HD for prostate and rectum, improved DSC for bladder, and reduced fiducial separation for the prostate, with p<0.05 for many pairwise comparisons.

Significance. The paper addresses a clinically relevant problem—improving DIR accuracy on low-contrast CBCT images—and combines image translation, auto-segmentation, and structure-guided registration in a plausible pipeline. The prostate fiducial results provide an independent and clinically meaningful metric, which is a notable strength. However, the central claim of accuracy improvement across all three organs is not fully supported: the bladder 95% HD worsens under CycleGAN PD in Table 2, and the evaluation protocol appears to use different contour sources across arms, confounding the DSC/HD comparisons. If the confound is resolved, the method could be valuable, but as presented the evidence for all-organ improvement is incomplete.

major comments (4)
  1. [§2.5, Table 2] The manuscript does not specify which contours are used to compute DSC and 95% HD in each arm. In the CycleGAN PD arm, the only moving-image structures available are the U-Net contours on sCT, whereas in the No PD and Expert PD arms the moving structures are presumably expert contours on CBCT. If this is the case, the reported DSC/HD differences between CycleGAN PD and No PD are confounded by the contour source, not solely by the registration method. Please state the exact evaluation protocol, or recompute metrics using a common contour set (e.g., expert CBCT contours propagated by each DVF). This is load-bearing because the abstract's claim of improved accuracy for bladder and rectum rests on these DSC/HD numbers, and the bladder 95% HD in Table 2 actually worsens under CycleGAN PD (21.48 mm vs 19.59 mm, p > 0.05).
  2. [§2.2, Eq. (1)] The weights λ and η_n in Eq. (1) are never reported. These hyperparameters control the balance between intensity similarity, regularization, and structural guidance, and the behavior of the hybrid metric depends critically on them. Without their values or a sensitivity analysis, the experiments are not reproducible. Please report the values used for all experiments and describe how they were selected or tuned.
  3. [§2.3, §2.4] The CycleGAN is 2D and processes axial slices independently, yet no quantitative assessment of sCT quality or slice-to-slice consistency is provided; Figure 3 is only a visual example. Furthermore, Section 2.1 states that the sCT labels used to train the U-Net were 'contours transferred from CBCT,' meaning the segmentation accuracy on sCT is not validated against independent ground truth. Please provide quantitative sCT evaluation (e.g., mean absolute error against CT, and a 3D consistency metric) and, if possible, assess U-Net segmentation on sCT against expert contours drawn directly on CT.
  4. [§2.1, Table 2] The validation cohort consists of only 7 patients. With n=7, the reported p-values from ANOVA/Kruskal-Wallis and post-hoc tests are highly sensitive to outliers and have low statistical power. The paper should report effect sizes and confidence intervals, describe how the 7 validation patients were selected from the 14-case holdout, and discuss the limitation posed by the small validation cohort.
minor comments (5)
  1. [Abstract] The abstract states that 'Improvements were also observed for the bladder and rectum,' but Table 2 shows that the bladder 95% HD worsens under CycleGAN PD (21.48 mm vs 19.59 mm) and is not statistically significant. Please adjust the wording to reflect the mixed result.
  2. [§3.3] The sentence 'Overall, percentage improvements in DSC and 95% HD ranged from a 56.64% increase in DSC for the bladder to a 73.42% reduction in 95% HD' is unclear and appears inconsistent with Table 2. Please specify which comparisons yield these percentages and correct the values if needed.
  3. [Table 2] For the bladder 95% HD row, the p-value column lists both ANOVA (0.1414) and Kruskal-Wallis (0.005). This is confusing; specify which test was used and whether the pairwise comparison was adjusted for multiple testing.
  4. [§2.2, Eq. (1)] In the sentence following Eq. (1), the phrase 'π_n′ is the moving image' appears to be a typo; presumably π_n′ denotes a point set on the moving image boundary. Please correct.
  5. [§2.7] With n=7 per group, the Shapiro-Wilk normality test has very low power. Please consider reporting the raw distributions or using a more robust approach, and note this limitation in the statistical analysis.

Circularity Check

2 steps flagged · score 4.0 of 10

DSC/HD gains for CycleGAN PD and Expert PD are computed from the same contours that drive the PD term, making those structural metrics partly self-fulfilling; fiducial separation independently supports the prostate result, but bladder/rectum improvements rest on the circular metric and bladder HD actually worsens.

  1. self definitional [Sec. 2.4 (U-Net Auto-Segmentation), Sec. 2.5 (DIR Evaluation), Table 2]
    "The final trained model was used to automatically segment the prostate, bladder, and rectum in both sCT and CT images. These segmented structures were then employed to compute the PD metric for deformable image registration. ... DSC was used to quantify the overlap between segmented structures on the fixed and moving images."

    In the CycleGAN PD arm, the U-Net contours produced in Sec. 2.4 are the explicit structural input to the PD term in Eq. (1), and Sec. 2.5 defines DSC and 95% HD over 'segmented structures on the fixed and moving images' without introducing any independent contour set for this arm. Thus the reported overlap improvement is largely the same boundary-distance objective being optimized, rather than an independent anatomical check. For bladder and rectum, no fiducial or other external metric is available to break this tie; the bladder 95% HD actually worsens (19.59 to 21.48 mm, p > 0.05) while the abstract claims improvements for bladder and rectum.

  2. self definitional [Sec. 2.2 (Eq. 1), Sec. 2.5 (DIR Evaluation), Table 2]
    "The metric included a term that minimized the distance between points on a manually or automatically segmented structure boundary (e.g., prostate or bladder) on the fixed (CT) image, and the corresponding boundary in the moving (CBCT) image, using an unsigned distance map. ... Expert PD: Expert manual segmentations were used with the PD metric, serving as a benchmark for comparison with the automated methods."

    For the Expert PD arm, the same expert contours define the PD optimization target and supply the structures whose overlap and Hausdorff distance are scored. Reducing point-to-boundary distances on a contour pair and then measuring DSC/95% HD on that same contour pair makes the Expert PD gains over No PD largely by construction for the structural metrics. This is a benchmark arm, but the paper presents its large DSC/HD improvements as evidence for the hybrid metric's benefit; only the fiducial-separation metric provides independent confirmation for the prostate.

full rationale

The paper has two structural-evaluation arms with a self-referential evaluation design. In the CycleGAN PD arm, the U-Net contours generated in Sec. 2.4 are the explicit input to the PD term in Eq. (1), and Sec. 2.5 defines DSC and 95% HD over 'segmented structures on the fixed and moving images' without specifying any independent contour source for this arm. Optimizing a boundary-distance term on a contour pair and then measuring overlap/Hausdorff distance on the same pair makes the reported structural improvement partly the optimization objective itself. The bladder and rectum results have no fiducial or other external check to break this tie, and the bladder 95% HD actually worsens (19.59 to 21.48 mm, p > 0.05) despite the abstract claiming improvements for bladder and rectum. The Expert PD arm has the same structure: expert contours both define the PD target and supply the DSC/HD evaluation, so its large contour-metric gains versus No PD are partly built in. Not all claims are circular: prostate fiducial separation is an independent, externally meaningful metric and improves significantly under CycleGAN PD (8.95 to 4.07 mm, p = 0.014), so the central claim for the prostate retains independent content. The PD metric is self-cited (Shah et al. 2021), but it is fully defined in Eq. (1) and implemented in an open-source toolkit, so that citation is not load-bearing. Overall score reflects partial circularity in the structural metrics while acknowledging the independent fiducial support.

Assumptions & free parameters 2 free parameters · 5 assumptions · 0 invented entities

The method rests on two trained deep networks (CycleGAN and U-Net), a hybrid cost function with unstated weighting coefficients, and a 7-case validation cohort. The sCT fidelity and the accuracy of auto-segmentation are load-bearing, and the paper itself acknowledges that segmentation errors propagate. The independent fiducial metric reduces, but does not eliminate, the circularity of evaluating contour-guided registration with contour overlap.

free parameters (2)
  • PD metric regularization weight lambda
    Weight on the curvature regularizer S(nu) in Eq. (1). Not specified numerically; its value affects the balance between intensity, smoothness, and structure terms and could be tuned on the validation cohort.
  • PD metric per-structure weights eta_n
    Weights applied to the point-to-distance term for prostate, bladder, and rectum in Eq. (1). Not reported; the relative weighting directly controls how strongly contours influence the registration.
assumptions (5)
  • standard math B-spline free-form deformation with curvature regularization can represent the pelvic deformations between CT and CBCT.
    Section 2.2 and the Plastimatch toolkit: the DIR model is a B-spline FFD with lambda S(nu) curvature regularizer. This is a standard assumption in DIR but is not verified against a known deformation ground truth in this paper.
  • domain assumption The 2D CycleGAN, processing axial slices independently, preserves 3D anatomical consistency and does not introduce hallucinated structures in sCT.
    Section 2.3 processes axial slices independently; sCT quality is assessed only visually (Figure 3), not quantitatively or in 3D.
  • domain assumption U-Net segmentations on sCT are accurate enough to guide the PD metric.
    Section 2.4 uses U-Net contours to compute PD; Section 4 acknowledges that poor segmentation can propagate through the registration process.
  • domain assumption Expert manual contours and fiducial markers provide a reliable reference standard for evaluating DIR accuracy.
    Section 2.5 uses expert contours for DSC and HD and fiducial positions for independent validation; contour variability is not quantified.
  • domain assumption The 7-case validation cohort is representative of the broader pelvic ART population.
    Section 2.4 and 2.5 use 7 CT and 7 sCT images from PRD and MGH for DIR validation, with no external cohort.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Improving Deformable Image Registration Accuracy through a Hybrid Similarity Metric and CycleGAN Based Auto-Segmentation." pith.science (2026). https://pith.science/paper/4X3UUKSF

@misc{pith2026241116992,
  author       = {Pith},
  title        = {Pith review of: Improving Deformable Image Registration Accuracy through a Hybrid Similarity Metric and CycleGAN Based Auto-Segmentation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/4X3UUKSF}},
  note         = {Machine review of arXiv:2411.16992}
}
read the original abstract

Purpose: Deformable image registration (DIR) is critical in adaptive radiation therapy (ART) to account for anatomical changes. Conventional intensity-based DIR methods often fail when image intensities differ. This study evaluates a hybrid similarity metric combining intensity and structural information, leveraging CycleGAN-based intensity correction and auto-segmentation across three DIR workflows. Methods: A hybrid similarity metric combining a point-to-distance (PD) score and intensity similarity was implemented. Synthetic CT (sCT) images were generated using a 2D CycleGAN model trained on unpaired CT and CBCT images to enhance soft-tissue contrast. DIR workflows compared included: (1) traditional intensity-based (No PD), (2) auto-segmented contours on sCT (CycleGAN PD), and (3) expert manual contours (Expert PD). A 3D U-Net model trained on 56 images and validated on 14 cases segmented the prostate, bladder, and rectum. DIR accuracy was assessed using Dice Similarity Coefficient (DSC), 95% Hausdorff Distance (HD), and fiducial separation. Results: The hybrid metric improved DIR accuracy. For the prostate, DSC increased from 0.61+/-0.18 (No PD) to 0.82+/-0.13 (CycleGAN PD) and 0.89+/-0.05 (Expert PD), with reductions in 95% HD from 11.75 mm to 4.86 mm and 3.27 mm, respectively. Fiducial separation decreased from 8.95 mm to 4.07 mm (CycleGAN PD) and 4.11 mm (Expert PD) (p < 0.05). Improvements were also observed for the bladder and rectum. Conclusion: This study demonstrates that a hybrid similarity metric using CycleGAN-based auto-segmentation improves DIR accuracy, particularly for low-contrast CBCT images. These findings highlight the potential for integrating AI-based image correction and segmentation into ART workflows to enhance precision and streamline clinical processes.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

1 extracted references

  1. [1]

    Chen L, Liang X, Shen C, Nguyen D, Jiang S and Wang J 2021 Synthetic CT generation from CBCT images via unsupervised deep learning Phys. Med. Biol. 66 115019 Çiçek Ö, Abdulkadir A, Lienkamp S S, Brox T and Ronneberger O 2016 3D U-Net: Learning Dense Volumetric Segmentation from Sparse Annotation Medical Image Computing and Computer-Assisted Intervention –...

    arXiv 2021

Pith tools

Reviewed August 12, 2026 · model on record in the stance chip above.