Source-linked AI summary
Really Need Recursion? Revisiting One-pass Registration with Multi-level Multi-space Contextual Guidance
Yue Lin, Shenghai Wu, Mingyuan Meng
TL;DR
The paper asks when recursion is needed for deformable registration, given its accuracy benefits and computational overhead. M2CG uses one-pass pyramid registration guided by multi-level, multi-space context; it reports state-of-the-art results across two tasks and eight datasets. Its findings suggest that reliance on recursion may reflect insufficient contextual guidance.
Problem
Recursive registration can improve accuracy but adds computational overhead, and few studies have critically examined why or when it is needed.
Source: §1 · p. 2Method
M2CG performs one-pass progressive pyramid registration using contextual cues from previous, current, and subsequent steps across four spaces.
Source: §1 · p. 2, §2.2 · p. 4Results
M2CG achieved the highest DSCs across evaluation datasets in both supervision settings and state-of-the-art performance across two tasks and eight datasets.
Source: §4 · p. 6, §5 · pp. 8–9Takeaways & Limitations
The findings suggest that reliance on recursion may stem from insufficient contextual guidance rather than inherent limits of one-pass architectures.
Source: §5 · p. 9The formulation assumes single-channel, grayscale 3D volumes.
Source: §2 · p. 3
Abstract
from arXiv · showhide
Deformable image registration is a fundamental challenge in medical image analysis. Recent deep learning-based methods typically adopt either one-pass architectures that complete registration in a single network iteration, or recursive architectures that refine registration through multiple passes. While recursive methods often improve accuracy, they incur substantial computational overhead and slower inference speed. In this paper, we challenge the necessity of recursion for high-performance deformable registration and propose M2CG, a one-pass registration network with Multi-level Multi-space Contextual Guidance. It performs multi-step progressive pyramid registration within a single iteration, guided at each step by contextual cues from three levels (previous, current, and subsequent steps) and four spaces (image space, image feature space, segmentation space, and segmentation feature space). Our M2CG supports both unsupervised (without any labels) and semi-supervised (with segmentation labels for partial training data) registration, and has achieved state-of-the-art performance across two well-benchmarked inter-/intra-patient medical registration tasks with eight public datasets. Our findings suggest that the current reliance on recursion may stem from the absence of sufficient contextual cues that comprehensively characterize the remaining distance to the target alignment.
1 Introduction
Deep registration methods trade one-pass speed for recursive refinement, while the conditions under which recursion is needed remain insufficiently examined. M2CG tests whether richer contextual guidance can support accurate one-pass progressive registration.
1 Introduction
- One-pass methods align images in one network iteration, whereas recursive methods refine alignment through multiple passes and typically incur greater computational cost.The introduction describes recursion as often improving accuracy, but also notes its computational overhead and slower inference.Source: §1 · p. 2
- The paper hypothesizes that recursion may compensate for insufficient step-wise context, rather than being inherently necessary for accurate pyramid registration.It proposes that contextual cues describing each step's status could guide its optimization direction; recursion may also introduce accumulated errors.Source: §1 · p. 2
- M2CG tests the hypothesis with one-pass progressive pyramid registration guided by cues from three registration steps and four information spaces.The spaces are image, image feature, segmentation, and segmentation feature space.Source: §1 · p. 2
- M2CG supports unsupervised and semi-supervised registration and was evaluated on two inter-/intra-patient tasks with eight public datasets.The semi-supervised setting uses segmentation labels for partial training data.Source: §1 · p. 3
2 Method
M2CG uses shared image features and, when labels are available, an auxiliary segmentation decoder alongside a coarse-to-fine registration decoder. Its training loss is unsupervised or semi-supervised according to label availability.
2 Method
- M2CG combines an image encoder, an optional segmentation decoder, and a registration decoder for one-pass pyramid registration.The segmentation decoder is used only for semi-supervised registration; the model parameters can be optimized in either setting.Source: §2 · p. 3, §2.1 · p. 4
2.1 Image Encoder and Segmentation Decoder
- The shared encoder extracts multi-scale features, while the segmentation decoder progressively predicts a four-level mask pyramid when partial labels are available.The segmentation decoder is omitted for unsupervised registration.Source: §2.1 · pp. 3–4
2.2 Registration Decoder with Multi-level Multi-space Contextual Guidance
- The registration decoder uses previous-, current-, and subsequent-step cues across four spaces to guide coarse-to-fine registration at four scale levels.Its look-ahead cue uses wavelet-decomposed images to expose finer-scale details during the current step; unsupervised registration omits segmentation-related operations.Source: §2.2 · pp. 4–5
2.3 Unsupervised and Semi-supervised Learning
- Unsupervised training combines image similarity and transformation regularization, while semi-supervised training adds segmentation and joint registration-segmentation constraints.Semi-supervised training can use both labeled and unlabeled data, excluding label-dependent terms when labels are unavailable.Source: §2.3 · p. 5
3 Experimental Setup
The evaluation covers inter-patient brain and intra-patient cardiac registration using eight public datasets. Experiments compare M2CG with existing methods and test contextual cues and added recursion.
3.1 Datasets and Preprocessing
- Eight public datasets span inter-patient brain and intra-patient cardiac registration, including independent brain tests and cardiac ED-ES registration.Brain training combines 314 labeled OASIS images with 2,656 unlabeled images; Mindboggle and Buckner are independent tests, and ACDC supplies cardiac ED-ES data.Source: §3.1 · p. 5
3.2 Implementation Details
- Brain models trained for 100,000 iterations and cardiac models for 50,000, with training schedules differing by task and supervision setting.The semi-supervised brain schedule begins with labeled pairs; cardiac schedules include ED-ES pairs in the final 10,000 iterations.Source: §3.2 · p. 6
3.3 Experimental Designs
- Comparisons cover traditional, one-pass, recursive, and segmentation-assisted methods; ablations test each contextual cue and whether recursion still helps.Registration accuracy is evaluated by Dice similarity coefficient, and transformation smoothness by percentage of Negative Jacobian Determinants.Source: §3.3 · p. 6
4 Results and Discussion
M2CG achieved the highest DSCs across evaluation datasets in both settings while retaining lower or similar NJDs. Adding recursion to M2CG yielded limited accuracy gains at greater computational cost.
4 Results and Discussion
- In the qualitative comparison, M2CG’s registration results were most consistent with the fixed image.Figure 2 shows intensity-difference error maps below each image and marks mean absolute error in red.Source: §4 · p. 7, Figure 2 · p. 7
- M2CG achieved the highest DSCs across all evaluation datasets in both settings and outperformed recursive methods at one-pass computational cost.Table 1 compares registration methods; M2CG also exhibited lower or similar NJDs across datasets and settings.Source: §4 · pp. 6–7, Table 1 · p. 7
- Recursion performed twice nearly doubled computational cost and GPU memory use, and increased NJDs.The recursion depth was limited to two repetitions by GPU memory constraints; the authors note that errors or transformation folds may accumulate with repeated registration.Source: §4 · p. 8
- Removing any contextual cue degraded performance in both settings, while adding recursion produced only marginal unsupervised gains and no semi-supervised improvement.The tested recursion mechanisms were network, decoder, and pyramid-scale recursion.Source: Table 2 · p. 8, §4 · p. 8
5 Conclusion
M2CG uses contextual guidance across three progressive levels and four complementary spaces to perform pyramid registration in one network iteration. Across eight public datasets and two inter-/intra-patient tasks, it achieved state-of-the-art performance in both training settings, while added recursion brought limited gains at increased cost and transformation folding.
5 Conclusion
- Across eight public datasets and two inter-/intra-patient tasks, M2CG demonstrated state-of-the-art performance in both unsupervised and semi-supervised settings.The conclusion scopes this finding to the evaluated uni-modal settings.Source: §5 · pp. 8–9
- Adding recursion to M2CG provided limited accuracy gains while increasing computational cost and transformation folding.The authors suggest reliance on recursion may stem from insufficient contextual guidance rather than inherent limitations of one-pass architectures.Source: §5 · p. 9
- Validation on more anatomical sites, cross-modal registration, and clinically heterogeneous scenarios remains future work.The stated conclusion is limited to the evaluated uni-modal settings [37].Source: §5 · p. 9