ProRSeg: Progressive Reg. & Segmentation

• The paper presents a novel joint registration-segmentation framework that iteratively refines deformation fields using recurrent networks for accurate organ alignment.
• It employs 5-fold cross-validation on MRI and CBCT scans, achieving high Dice scores and demonstrating reliable registration consistency essential for dosimetric analysis.
• ProRSeg integrates anatomical regularization via segmentation, enabling effective downstream dose accumulation even in variable and complex organ geometries.

Progressively Refined Registration and Segmentation (ProRSeg) is a deep joint registration-segmentation framework introduced for gastrointestinal organs at risk, with an initial application to MRI and cone-beam CT in a treatment-fraction setting (Jiang et al., 2022). In the published record, it is characterized by joint estimation of organ segmentations and deformable alignment, evaluation through segmentation accuracy and registration consistency, and downstream use for dose accumulation. A later prostate MR-guided radiotherapy study describes ProRSeg as a recurrent registration-segmentation framework with incremental deformation vector fields and segmentation-derived anatomical regularization, indicating that the “progressively refined” designation refers to iterative refinement of alignment rather than a single-shot deformation prediction (Madhavan et al., 9 Jul 2025).

1. Definition and original clinical scope

ProRSeg was introduced as “progressively refined deep joint registration segmentation (ProRSeg) of gastrointestinal organs at risk: Application to MRI and cone-beam CT” (Jiang et al., 2022). Its original target structures were liver, large bowel, small bowel, and stomach-duodenum, and its empirical setting was longitudinal imaging across treatment fractions. In that study, ProRSeg was trained using 5-fold cross-validation with 110 T2-weighted MRI acquired at 5 treatment fractions from 10 different patients, with the constraint that same patient scans were not placed in training and testing folds. Applicability to cone-beam CT was also evaluated on 80 scans using 5-fold cross-validation.

The original evaluation protocol separated the registration and segmentation aspects explicitly. Segmentation accuracy was measured using Dice similarity coefficient (DSC) and Hausdorff distance at 95th percentile (HD95). Registration consistency was measured using coefficient of variation (CV) in displacement of organs at risk. Ablation tests and accuracy comparisons against multiple methods were also reported. This framing places ProRSeg in the class of methods that treat contour propagation, anatomical overlap, and geometric stability as coupled but distinct criteria.

A plausible implication is that ProRSeg was designed for settings in which anatomically meaningful segmentation and deformable alignment must both be reliable enough to support downstream dose analysis. That interpretation is consistent with the study’s inclusion of dose accumulation accounting for intra-fraction and inter-fraction motion.

2. Reported performance on MRI and cone-beam CT

The original study reports that ProRSeg processed 3D volumes of size $128 \times 192 \times 128$ in 3 secs on a NVIDIA Tesla V100 GPU and that its segmentations were significantly more accurate ($p<0.001$) than compared methods (Jiang et al., 2022). On MRI, the reported DSC values were $0.94 \pm 0.02$ for liver, $0.88 \pm 0.04$ for large bowel, $0.78 \pm 0.03$ for small bowel, and $0.82 \pm 0.04$ for stomach-duodenum. On CBCT, the reported DSC values were $0.72 \pm 0.01$ for small bowel and $0.76 \pm 0.03$ for stomach-duodenum.

Structure	MRI DSC	CBCT DSC
Liver	$0.94 \pm 0.02$	—
Large bowel	$0.88 \pm 0.04$	—
Small bowel	$p<0.001$0	$p<0.001$1
Stomach-duodenum	$p<0.001$2	$p<0.001$3

These results establish two features of the framework’s early profile. First, the method was not confined to a single modality, since MRI and CBCT were both evaluated. Second, the organ-wise spread in DSC indicates that ProRSeg addressed structures of substantially different geometric and contrast complexity, with liver segmentation markedly higher than bowel segmentation. This suggests that the framework’s difficulty profile was organ dependent rather than uniform across the abdominal field.

3. Registration consistency and dose accumulation

Beyond overlap-based segmentation metrics, the original report emphasizes registration consistency through coefficient of variation in organ displacement (Jiang et al., 2022). ProRSeg registrations resulted in the lowest CV in displacement for stomach-duodenum ($p<0.001$4, $p<0.001$5, $p<0.001$6), small bowel ($p<0.001$7, $p<0.001$8, $p<0.001$9), and large bowel ($0.94 \pm 0.02$0, $0.94 \pm 0.02$1, $0.94 \pm 0.02$2).

The same study linked registration to dosimetric analysis. ProRSeg-based dose accumulation accounting for intra-fraction motion from pre-treatment to post-treatment MRI scan and inter-fraction motion showed that organ dose constraints were violated in 4 patients for stomach-duodenum and in 3 patients for small bowel. In this formulation, ProRSeg is not only a contouring or correspondence engine; it is also a mechanism for transporting dose information through anatomically estimated motion fields.

A common misunderstanding is to reduce ProRSeg to a segmentation benchmark. The dose-accumulation result indicates that the intended use case extends to treatment-compliance analysis. This suggests that geometric plausibility of the registration field is central, not merely secondary to mask overlap.

4. Recurrent progressive refinement in later prostate adaptation

A later study in MR-guided prostate cancer radiotherapy explicitly describes ProRSeg as a joint registration-segmentation framework originally introduced in prior work for gastrointestinal organs and adapted to pelvic MRI (Madhavan et al., 9 Jul 2025). In that application, ProRSeg was described as consisting of two recurrent subnetworks trained jointly: a Recurrent Registration Network (RRN) and a Recurrent Segmentation Network (RSN). The RRN computes deformation vector fields through 8 CLSTM iterations, with each CLSTM step connected to a spatial transformer network to generate an incremental deformation vector field, yielding progressively refined registration. The RSN generates multi-class segmentations through 9 CLSTM iterations and uses a 3D U-Net style segmentation backbone recurrently refined via CLSTM blocks.

The coupling in that prostate application is through multi-task learning. The registration loss comprises image similarity, deformation smoothness, and weighted segmentation consistency, while the segmentation branch is trained with multi-category cross-entropy at each recurrent step. The weighted segmentation consistency is defined over $0.94 \pm 0.02$3 with weights $0.94 \pm 0.02$4, $0.94 \pm 0.02$5, and $0.94 \pm 0.02$6. The total registration loss is reported as

$0.94 \pm 0.02$7

with best values $0.94 \pm 0.02$8 and $0.94 \pm 0.02$9.

This later description clarifies the meaning of “progressively refined” within the ProRSeg lineage. The progressive component is implemented through multiple ConvLSTM iterations and incremental deformation updates, with segmentation consistency computed from the individual segmentations so as to provide a deep supervision segmentation consistency loss. At inference in that study, only the registration network was kept, while the segmentation branch was used only during training as a regularizer. A plausible implication is that ProRSeg, at least in this later form, should be understood as weakly supervised deformable registration regularized by segmentation rather than as a symmetric requirement for dual-branch inference.

5. Position within joint registration-segmentation research

ProRSeg belongs to a broader family of methods that couple anatomical labeling and spatial alignment, but its “progressively refined” aspect distinguishes it from several adjacent formulations. DeepAtlas jointly learns networks for image registration and image segmentation and is highly relevant as background, but it is not itself a progressively refined framework; it uses alternating optimization of two fixed networks rather than explicit multi-stage progressive refinement, recurrent correction, or cascaded coarse-to-fine registration-segmentation updates (Xu et al., 2019).

One-shot PACS is more closely aligned with ProRSeg-style design because it uses a recurrent registration network that produces a progressively deformed sequence of images and a recurrent segmentation network that consumes progressively warped patient-specific anatomic context and shape context, but its clinical target is longitudinal thoracic CBCT segmentation rather than gastrointestinal organs at risk (Jiang et al., 2022). By contrast, DAFF-Net simultaneously achieves the segmentation masks and dense deformation fields in a single-step estimation, with a global encoder, a segmentation decoder, and a coarse-to-fine pyramid registration decoder; it is multi-task and coarse-to-fine, but not described as a progressively refined recurrent framework in the ProRSeg sense (Zhou et al., 2024).

A different conceptual contrast is provided by SegReg, which is explicitly segmentation-driven and region-wise: it first decomposes input moving and fixed images into anatomically coherent subregions through segmentation, then processes localized domains by the same registration backbone, and finally integrates partial deformation fields into a global deformation field. That work is not itself a progressively refined joint registration-segmentation framework and instead argues that registration accuracy has a near-linear dependence on segmentation quality (Chen et al., 19 Sep 2025). Relative to such approaches, ProRSeg remains more tightly associated with iterative refinement of correspondence estimation under segmentation-derived anatomical constraints.

6. Limitations, misconceptions, and significance

The original gastrointestinal study identified two explicit limitations: lack of independent testing and lack of ground truth phantom datasets to measure dose accumulation accuracy (Jiang et al., 2022). These limitations are consequential because the method was used not only for segmentation and registration benchmarks but also for dose accumulation. The absence of independent testing constrains claims about generalization, while the absence of ground truth phantom datasets constrains claims about dosimetric fidelity.

A later prostate study shows that ProRSeg can generalize across same-domain, cross-domain, and mixed-domain MR settings, but also reports that mixed-domain MRSim-to-MRL remained difficult and that urethra performance was weak and variable (Madhavan et al., 9 Jul 2025). This suggests that the progressive and segmentation-regularized formulation does not eliminate domain shift or small-structure difficulty. The most difficult scenarios remain those with stronger appearance mismatch or limited structural support.

A recurring misconception is that any method combining registration and segmentation is therefore a ProRSeg-type method. The literature summarized here indicates otherwise. Some methods are joint but not progressively refined; some are progressively refined but only on the registration side; some are segmentation-driven without mutual refinement. ProRSeg is more precisely identified by the conjunction of joint registration-segmentation design, iterative or recurrent refinement of deformation estimation, and evaluation in clinically consequential settings where contour propagation and dose accumulation are both relevant. In that narrower sense, ProRSeg occupies a specific position in anatomy-aware deformable registration: it is a framework in which segmentation is not merely an endpoint, but a structural constraint on how registration is learned and, in later forms, progressively updated.