iPEAR: Iterative Pyramid Estimation with Attention and Residuals for Deformable Medical Image Registration

Existing pyramid registration networks may accumulate anatomical misalignments and lack an effective mechanism to dynamically determine the number of optimization iterations under varying deformation requirements across images, leading to degraded performance. To solve these limitations, we propose iPEAR. Specifically, iPEAR adopts our proposed Fused Attention-Residual Module (FARM) for decoding, which comprises an attention pathway and a residual pathway to alleviate the accumulation of anatomical misalignment. We further propose a dual-stage Threshold-Controlled Iterative (TCI) strategy that adaptively determines the number of optimization iterations for varying images by evaluating registration stability and convergence. Extensive experiments on three public brain MRI datasets and one public abdomen CT dataset show that iPEAR outperforms state-of-the-art (SOTA) registration networks in terms of accuracy, while achieving on-par inference speed and model parameter size. Generalization and ablation studies further validate the effectiveness of the proposed FARM and TCI.

💡 Research Summary

iPEAR (Iterative Pyramid Estimation with Attention and Residuals) addresses two fundamental shortcomings of existing pyramid‑based deformable registration networks: the accumulation of anatomical mis‑alignment across scales and the lack of an adaptive mechanism to decide how many iterative refinements each scale requires.
The proposed solution consists of two novel components. First, the Fused Attention‑Residual Module (FARM) replaces the conventional vanilla residual decoder. FARM has a dual‑pathway design: an Attention Pathway (AP) that combines a 3‑D Squeeze‑Excitation Block (SEB) for channel‑wise importance weighting and a Spatial Attention Block (SAB) for spatial saliency, and a Residual Pathway (RP) that contains two 3‑D convolutions followed by two residual blocks to capture fine‑grained anatomical details. By suppressing irrelevant features while preserving local structure, FARM mitigates the propagation of errors from coarse to fine scales.
Second, the Threshold‑Controlled Iterative (TCI) strategy introduces a two‑stage stopping criterion for the iterative refinement at each pyramid level. In the first stage, a sliding window of the most recent t similarity differences (e.g., NCC) is examined; the standard deviation εₗ of these differences quantifies registration stability. When εₗ falls below a predefined threshold, the process proceeds to the second stage, where the absolute change Δs between the two latest similarity scores is computed. If Δs is also below a second threshold, the iteration stops. This dual‑stage check prevents premature termination on easy cases while allowing additional refinements for challenging deformations, all without adding noticeable computational overhead.
Architecturally, iPEAR employs a shared 4‑level encoder (each level: 3‑D convolution → Neighborhood Attention → average pooling) to extract multi‑scale feature maps {Fₗ} and {Mₗ} from the fixed and moving images. Decoding starts at the lowest resolution (level 4) and proceeds to higher resolutions. At each level l, the current moving feature map is warped by the deformation field estimated at the previous level (or previous iteration) using a Spatial Transformer Network, then concatenated with the fixed feature map and fed into FARM to produce a new deformation field φₖ,ₗ. The TCI controller monitors the similarity metrics and dynamically determines the number of iterations k for that level.
The authors evaluate iPEAR on three public brain MRI datasets (OASIS‑3, ADNI, HCP) covering a range of deformation magnitudes and on the BTCV abdominal CT dataset. Compared with six state‑of‑the‑art methods—including VoxelMorph, VoxelMorph‑Cascade, RDP, and the recent single‑stage adaptive method—iPEAR consistently achieves higher Dice scores (2–4 % absolute improvement), better NCC and lower Hausdorff distances, while maintaining a comparable model size (~12 M parameters) and inference time (~0.12 s per volume). Ablation studies show that removing either FARM or TCI degrades performance by roughly 1.5–2 % Dice, confirming that both components contribute synergistically. Moreover, cross‑dataset generalization tests demonstrate that a model trained on one modality or institution retains high accuracy on unseen data, indicating robust feature learning.
Limitations noted by the authors include the high GPU memory demand of processing full 3‑D volumes and the need to manually set the two TCI thresholds for each new dataset. Future work may explore memory‑efficient slice‑based processing, meta‑learning to automatically tune thresholds, and extension to multimodal registration (e.g., CT‑MRI).
In summary, iPEAR introduces a principled attention‑residual decoder and a dual‑stage adaptive iteration scheme that together overcome the primary sources of error in pyramid registration networks. The resulting system delivers state‑of‑the‑art accuracy without sacrificing speed or model compactness, making it a promising candidate for real‑time clinical applications such as image‑guided surgery, longitudinal study alignment, and multi‑center data harmonization.