Enhancing Prostate Cancer Segmentation for Multi-Domain Generalization using a novel Parallel-Route Coherent Mixup Regularization Training

Daniel Gorovets; Harini Veeraraghavan; Himanshu Nagar; James Janopaul-Naylor; Josiah Simeth; Justin Haseltine; Neelam Tyagi; Sean McBride; Sudharsan Madhavan; Victoria Brennan

arxiv: 2607.00370 · v1 · pith:B7KIYMKOnew · submitted 2026-07-01 · 📡 eess.IV

Enhancing Prostate Cancer Segmentation for Multi-Domain Generalization using a novel Parallel-Route Coherent Mixup Regularization Training

Josiah Simeth , Sudharsan Madhavan , Victoria Brennan , Sean McBride , James Janopaul-Naylor , Justin Haseltine , Daniel Gorovets , Himanshu Nagar

show 2 more authors

Neelam Tyagi Harini Veeraraghavan

This is my paper

Pith reviewed 2026-07-02 04:51 UTC · model grok-4.3

classification 📡 eess.IV

keywords prostate cancer segmentationMRImulti-domain generalizationmixup regularizationdeep learningresidual networks

0 comments

The pith

PaRC-mix applies linear feature combinations at multiple network layers to improve prostate lesion segmentation across MRI scanners and lesion types.

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

The paper introduces PaRC-mix, a training method that augments features inside residual networks by linearly mixing intermediate activations from different samples in a batch. It tests this on MRRN and UNet++ models trained on one 3T GE dataset and evaluated on five external datasets from Siemens, Philips, and Elekta scanners. The approach yields higher Dice scores, lower surface distances, and better detection of both aggressive and non-aggressive lesions than standard training or input-level mixup. The accuracy difference between aggressive and non-aggressive lesions shrinks markedly when PaRC-mix is used. The method is presented as network-agnostic and easy to add to multi-stream architectures for single-source to multi-domain generalization.

Core claim

PaRC-mix training creates coherent feature augmentations at multiple layers by linear combination of features from different training samples; when applied to MRRN and UNet++ networks for prostate cancer lesion segmentation, it produces significantly higher composite accuracy on external multi-scanner data and narrows the performance gap between aggressive and non-aggressive lesions from roughly 20 points to under 8 points.

What carries the argument

parallel-route coherent mixup (PaRC-mix), which performs linear combinations of intermediate features drawn from different samples at several depths inside a residual network.

If this is right

Models trained with PaRC-mix achieve higher tumor detection recall at fixed precision on data from different manufacturers.
The performance difference between aggressive and non-aggressive lesions shrinks when the same architecture is trained with PaRC-mix instead of standard or input-mixup methods.
The method can be inserted into any multi-stream residual network without changing the architecture itself.
PaRC-mix outperforms both no-mixup training and mixup applied only at the input or backbone layers on the reported external datasets.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

The same layer-wise mixing strategy could be tried on other medical image segmentation tasks that suffer from scanner-domain shift.
Because the mixing occurs inside the network, it may interact differently with residual connections than with purely convolutional backbones.
If the linear combinations are applied only on selected layers, the benefit might be retained with lower computational cost during training.

Load-bearing premise

Linear combinations of intermediate features from different samples preserve the lesion-specific signal without creating new correlations that only look useful on the tested external datasets.

What would settle it

A new external test set from an unseen scanner vendor or field strength where PaRC-mix models show no improvement in composite DSC-HD95-MSD score over non-mixup baselines.

Figures

Figures reproduced from arXiv: 2607.00370 by Daniel Gorovets, Harini Veeraraghavan, Himanshu Nagar, James Janopaul-Naylor, Josiah Simeth, Justin Haseltine, Neelam Tyagi, Sean McBride, Sudharsan Madhavan, Victoria Brennan.

**Figure 1.** Figure 1: (A) Model training and validation to select the best model was performed on an internal dataset containing 2029 examples. (B) All models were tested on four distinct out of domain datasets, totaling 1428 examples, provided to model only after training, to assess segmentation quality and lesion detection accuracy across datasets and lesion characteristics. (C,D) Multiple training strategies were implemented… view at source ↗

read the original abstract

MRI guided adaptive radiotherapy (MRgART) for prostate cancer (PCa) targets tumors while sparing organs from unnecessary radiation. Daily treatment adaptation requires accurate segmentation of tumors and organs. Manual delineation can be time and cost prohibitive. Deep learning segmentation methods have limited success applied to datasets distinct from training, hampering generalizability and adoption of MRgART. We develop a novel parallel route coherent mixup (PaRC-mix) training approach for single source to multi-domain generalization. PaRC-mix creates feature augmentations at multiple network layers through linear combination of features from different training samples in a batch. PaRC-mix training was implemented on two deep and residually connected networks, a multiple resolution residual network (MRRN) and UNet++ to segment PCa dominant intraprostatic lesions from apparent diffusion coefficient images. Models were trained on 2,029 samples from 3.0T GE MRI and tested on 1,547 PCa samples from 5 datasets acquired using 3T Siemens, 3T Philips, and 1.5T Elekta Unity MR-Linac scanners. PaRC-mix training led to significantly more accurate tumor detection and segmentation for both networks compared to training without mixup as well as input-mix training. PaRC-mix also achieved better recall to precision tradeoff than mixup applied only on the network backbone or input-mixup. Using a normalized composite DSC, HD95, and MSD score the accuracy gap between aggressive and non-aggressive lesions decreased from 21.1 and 19.5 for MRRN and UNet++ models trained without mixup to 5.2 and 7.9 with same models trained with PaRC-mix. This paper presents an easy to implement network agnostic approach to feature augmentation in multi-stream networks that enhances generalizability for the difficult problem of prostate cancer lesion segmentation.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

PaRC-mix is a straightforward multi-layer feature mixup variant that narrows the scanner gap in prostate lesion segmentation on external data, but the abstract supplies too few controls and statistics to confirm the gains reflect genuine domain invariance rather than regularization artifacts.

read the letter

The paper's core contribution is applying linear interpolation of intermediate features from different samples across parallel routes at multiple layers inside residual networks, then showing this helps MRRN and UNet++ models trained only on 3T GE data perform better on five external datasets from Siemens, Philips, and Elekta scanners. The reported drop in the aggressive versus non-aggressive lesion gap on a normalized composite of DSC, HD95, and MSD is the practical result worth noting.

It does a clean job of comparing the new scheme against no-mixup and input-only mixup baselines on the same two architectures, and the method itself is network-agnostic and simple to add. That matches the kind of incremental regularization work that can actually get used in clinics.

The soft spots sit in the missing details. The abstract states statistically significant gains but gives no error bars, p-values, exact interpolation coefficients, or ablation tables that isolate the parallel coherent placement from other mixup choices. The stress-test concern lands: without results on a held-out partition from the original GE domain or on synthetic shifts that do not match the test scanners, it remains unclear whether the interpolated features truly suppress scanner cues or simply produce activations that score well on these particular external sets. The assumption that linear combinations preserve lesion semantics while removing domain signals is stated but not directly tested.

This work is aimed at researchers doing multi-vendor MRI segmentation who need a lightweight way to improve generalization. A reader already familiar with mixup variants will see the specific implementation and the external-scanner numbers, but will still want the full methods and statistics before adopting it.

The paper deserves peer review because the task is clinically relevant and the approach is reproducible enough to check, even though the current evidence is thin on controls.

Referee Report

4 major / 0 minor

Summary. The manuscript proposes a novel Parallel-Route Coherent Mixup (PaRC-mix) regularization for single-source to multi-domain generalization in prostate cancer lesion segmentation on ADC MRI. PaRC-mix performs linear interpolation of intermediate activations from different training samples (all from one 3T GE source domain) at multiple layers of MRRN and UNet++ networks. Models trained with PaRC-mix are reported to outperform no-mixup and input-mixup baselines on 1547 samples from five external scanners (3T Siemens, 3T Philips, 1.5T Elekta), with improved DSC/HD95/MSD and a narrowed accuracy gap between aggressive and non-aggressive lesions (composite score gap reduced from 21.1/19.5 to 5.2/7.9).

Significance. If the reported gains reflect genuine domain-invariant lesion features rather than test-set-specific artifacts, the method offers a practical, network-agnostic way to improve generalization for MR-guided adaptive radiotherapy without multi-source training data, which would address a clinically relevant barrier to automated segmentation adoption.

major comments (4)

Abstract and results: claims of 'significantly more accurate' tumor detection and segmentation, plus the composite-score gap reductions, are presented without error bars, p-values, confidence intervals, or any statistical test details, undermining the ability to assess whether the improvements exceed experimental variability.
Methods: the mixup interpolation strength (lambda) and the precise definition of 'parallel routes' and affected layers are not specified, nor are ablation results isolating PaRC-mix from other regularization effects or from input-mixup.
Evaluation design: the central multi-domain generalization claim rests on external-scanner performance, yet no control experiments are described that evaluate the same models on a held-out partition of the source 3T GE domain or on synthetic domain shifts; without these, it is impossible to distinguish true invariance from scanner-specific regularization artifacts.
Results: the normalized composite DSC/HD95/MSD score used to quantify the aggressive/non-aggressive gap is introduced without definition of the normalization, weighting scheme, or exact formula, making the reported reductions (21.1/19.5 to 5.2/7.9) difficult to interpret or reproduce.

Simulated Author's Rebuttal

4 responses · 0 unresolved

We thank the referee for the detailed and constructive comments on our manuscript. We address each of the major comments below and will make the necessary revisions to improve the clarity and rigor of the paper.

read point-by-point responses

Referee: Abstract and results: claims of 'significantly more accurate' tumor detection and segmentation, plus the composite-score gap reductions, are presented without error bars, p-values, confidence intervals, or any statistical test details, undermining the ability to assess whether the improvements exceed experimental variability.

Authors: We agree that statistical details are essential. In the revised manuscript, we will report standard deviations across multiple training runs, include p-values from paired statistical tests comparing PaRC-mix to baselines, and provide confidence intervals for the key metrics including the composite scores. revision: yes
Referee: Methods: the mixup interpolation strength (lambda) and the precise definition of 'parallel routes' and affected layers are not specified, nor are ablation results isolating PaRC-mix from other regularization effects or from input-mixup.

Authors: The full methods section in the manuscript provides some details, but we acknowledge the need for more precision. We will explicitly define the lambda distribution used for interpolation, clarify that parallel routes refer to the multiple resolution paths in MRRN and the nested skip connections in UNet++, specify the layers where mixup is applied, and include additional ablation experiments to isolate the effect of PaRC-mix. revision: yes
Referee: Evaluation design: the central multi-domain generalization claim rests on external-scanner performance, yet no control experiments are described that evaluate the same models on a held-out partition of the source 3T GE domain or on synthetic domain shifts; without these, it is impossible to distinguish true invariance from scanner-specific regularization artifacts.

Authors: This point is well-taken. While the primary goal is multi-domain generalization, we will add results evaluating the models on a held-out subset of the source domain to confirm that performance is maintained or improved. We will also discuss the rationale for focusing on real external scanner data rather than synthetic shifts. revision: yes
Referee: Results: the normalized composite DSC/HD95/MSD score used to quantify the aggressive/non-aggressive gap is introduced without definition of the normalization, weighting scheme, or exact formula, making the reported reductions (21.1/19.5 to 5.2/7.9) difficult to interpret or reproduce.

Authors: We will revise the manuscript to include a clear definition of the composite score, including the normalization method (e.g., min-max scaling per metric), the weighting scheme (equal weights or specified), and the exact formula used to compute the gap reductions. revision: yes

Circularity Check

0 steps flagged

No circularity: empirical method evaluated on held-out external domains

full rationale

The paper proposes an empirical training regularization (PaRC-mix feature interpolation) and reports segmentation metric gains on 5 held-out scanner domains never seen in training. No equations, predictions, or derivations are presented that reduce the reported improvements to a fitted parameter or self-referential definition drawn from the test data. No load-bearing self-citations, uniqueness theorems, or ansatzes are invoked. The evaluation protocol is standard single-source to multi-domain testing and remains self-contained against external benchmarks.

Axiom & Free-Parameter Ledger

1 free parameters · 1 axioms · 0 invented entities

The approach rests on the untested premise that feature-space linear interpolation improves domain invariance for lesion segmentation; no free parameters beyond standard mixup strength are enumerated in the abstract.

free parameters (1)

mixup interpolation strength
Hyperparameter controlling how strongly features from different samples are blended; value chosen to optimize validation performance.

axioms (1)

domain assumption Feature mixing at multiple layers produces augmentations that generalize across scanner domains
Core premise of PaRC-mix; invoked when claiming improved multi-domain performance.

pith-pipeline@v0.9.1-grok · 5921 in / 1245 out tokens · 22188 ms · 2026-07-02T04:51:27.593191+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

3 extracted references · 3 canonical work pages

[1]

SAM-UNETR: Clinically Significant Prostate Cancer Segmentation Using Transfer Learning From Large Model,

Available at: https://doi.org/10.1186/1748-717X-8-84. Alzate-Grisales, J.A. et al. (2023) “SAM-UNETR: Clinically Significant Prostate Cancer Segmentation Using Transfer Learning From Large Model,” IEEE Access, 11, pp. 118217–118228. Available at: https://doi.org/10.1109/ACCESS.2023.3326882. Aminbeidokhti, M. et al. (2024) “Domain Generalization by Rejecti...

work page doi:10.1186/1748-717x-8-84 2023
[2]

Deep learning-based dominant index lesion segmentation for MR-guided radiation therapy of prostate cancer,

Available at: https://doi.org/10.4172/2167-7964.1000170. Simeth, J. et al. (2023) “Deep learning-based dominant index lesion segmentation for MR-guided radiation therapy of prostate cancer,” Medical Physics, 50(8), pp. 4854–4870. Available at: https://doi.org/10.1002/mp.16320. Simeth, J. et al. (2026) “Parallel Route Coherent Mixup Deep Learning for Singl...

work page doi:10.4172/2167-7964.1000170 2023
[3]

Generalist medical foundation model improves prostate cancer segmentation from multimodal MRI images,

Available at: https://doi.org/10.1109/TMI.2020.2973595. Zhang, Y. et al. (2025) “Generalist medical foundation model improves prostate cancer segmentation from multimodal MRI images,” npj Digital Medicine, 8(1), p. 372. Available at: https://doi.org/10.1038/s41746-025-01756-2. Zhou, K. et al. (2021) “Domain Generalization with MixStyle.” arXiv. Available ...

work page doi:10.1109/tmi.2020.2973595 2020

[1] [1]

SAM-UNETR: Clinically Significant Prostate Cancer Segmentation Using Transfer Learning From Large Model,

Available at: https://doi.org/10.1186/1748-717X-8-84. Alzate-Grisales, J.A. et al. (2023) “SAM-UNETR: Clinically Significant Prostate Cancer Segmentation Using Transfer Learning From Large Model,” IEEE Access, 11, pp. 118217–118228. Available at: https://doi.org/10.1109/ACCESS.2023.3326882. Aminbeidokhti, M. et al. (2024) “Domain Generalization by Rejecti...

work page doi:10.1186/1748-717x-8-84 2023

[2] [2]

Deep learning-based dominant index lesion segmentation for MR-guided radiation therapy of prostate cancer,

Available at: https://doi.org/10.4172/2167-7964.1000170. Simeth, J. et al. (2023) “Deep learning-based dominant index lesion segmentation for MR-guided radiation therapy of prostate cancer,” Medical Physics, 50(8), pp. 4854–4870. Available at: https://doi.org/10.1002/mp.16320. Simeth, J. et al. (2026) “Parallel Route Coherent Mixup Deep Learning for Singl...

work page doi:10.4172/2167-7964.1000170 2023

[3] [3]

Generalist medical foundation model improves prostate cancer segmentation from multimodal MRI images,

Available at: https://doi.org/10.1109/TMI.2020.2973595. Zhang, Y. et al. (2025) “Generalist medical foundation model improves prostate cancer segmentation from multimodal MRI images,” npj Digital Medicine, 8(1), p. 372. Available at: https://doi.org/10.1038/s41746-025-01756-2. Zhou, K. et al. (2021) “Domain Generalization with MixStyle.” arXiv. Available ...

work page doi:10.1109/tmi.2020.2973595 2020