Refined-Segmentation R-CNN: A Two-stage Convolutional Neural Network for Punctate White Matter Lesion Segmentation in Preterm Infants

Jian Yang; Jie Li; Miaomiao Wang; Xianjun Li; Xingbo Gao; Yalong Liu; Ying Wang; Zhicheng Jiao

arxiv: 1906.09684 · v2 · pith:GFIBG4CRnew · submitted 2019-06-24 · 📡 eess.IV · cs.CV

Refined-Segmentation R-CNN: A Two-stage Convolutional Neural Network for Punctate White Matter Lesion Segmentation in Preterm Infants

Yalong Liu , Jie Li , Ying Wang , Miaomiao Wang , Xianjun Li , Zhicheng Jiao , Jian Yang , Xingbo Gao This is my paper

Pith reviewed 2026-05-25 17:31 UTC · model grok-4.3

classification 📡 eess.IV cs.CV

keywords PWML segmentationpreterm infantsT1w MRIR-CNNtwo-stage CNNheuristic RPNmedical image segmentationDCRF

0 comments

The pith

A two-stage network segments punctate white matter lesions in preterm infant T1w MRIs with Dice score 0.6616.

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

The paper develops an automatic segmentation method for punctate white matter lesions in preterm infant brains to lower risks of postnatal developmental problems. It builds RS RCNN, a two-stage CNN that adds a heuristic region proposal network to draw on surrounding context around lesions, pairs it with a lightweight segmentation head, and applies dense conditional random field refinement. The approach works from T1w MRI alone and reports better overlap, sensitivity, and boundary accuracy than earlier algorithms even for single-pixel lesions.

Core claim

The RS RCNN constructs an efficient two-stage PWML semantic segmentation network based on the characteristics of the lesion. A heuristic RPN utilizes surrounding information around the PWMLs for heuristic segmentation, a lightweight segmentation network segments the lesion quickly, and DCRF optimizes the results. Using only T1w MRIs, the model segments lesions of ordinary size or pixel size, achieving Dice similarity coefficient of 0.6616, sensitivity 0.7069, specificity 0.9997, and Hausdorff distance 52.9130, outperforming the state-of-the-art algorithm.

What carries the argument

The heuristic RPN (H-RPN) that utilizes surrounding information around the PWMLs for heuristic segmentation inside the two-stage pipeline.

If this is right

Lesions as small as single pixels become detectable without additional MRI contrasts.
Only T1-weighted images suffice, removing the need to acquire and align multiple sequences.
Higher specificity of 0.9997 reduces false positives that could trigger unnecessary follow-up.
The DCRF post-processing step improves boundary precision measured by Hausdorff distance.
A lightweight segmentation head keeps inference fast enough for clinical workflows.

Where Pith is reading between the lines

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

The surrounding-context heuristic may transfer to segmentation of other small or low-contrast lesions such as microbleeds or early demyelination plaques.
Adding multi-center or longitudinal test sets would reveal whether the current metrics reflect overfitting to a single scanner or population.
Replacing the heuristic RPN with learned attention modules could test whether explicit surrounding priors remain necessary once more data are available.

Load-bearing premise

The heuristic RPN actually extracts useful surrounding context and the reported performance holds on new cases, given that dataset size, diversity, and validation procedure receive no details in the abstract.

What would settle it

Running the released code on an independent collection of T1w MRIs from preterm infants and obtaining a Dice score below the previous state-of-the-art method would refute the superiority claim.

read the original abstract

Accurate segmentation of punctate white matter lesion (PWML) in infantile brains by an automatic algorithm can reduce the potential risk of postnatal development. How to segment PWML effectively has become one of the active topics in medical image segmentation in recent years. In this paper, we construct an efficient two-stage PWML semantic segmentation network based on the characteristics of the lesion, called refined segmentation R-CNN (RS RCNN). We propose a heuristic RPN (H-RPN) which can utilize surrounding information around the PWMLs for heuristic segmentation. Also, we design a lightweight segmentation network to segment the lesion in a fast way. Densely connected conditional random field (DCRF) is used to optimize the segmentation results. We only use T1w MRIs to segment PWMLs. The result shows that our model can well segment the lesion of ordinary size or even pixel size. The Dice similarity coefficient reaches 0.6616, the sensitivity is 0.7069, the specificity is 0.9997, and the Hausdorff distance is 52.9130. The proposed method outperforms the state-of-the-art algorithm. (The code of this paper is available on https://github.com/YalongLiu/Refined-Segmentation-R-CNN)

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

The paper adapts R-CNN with a heuristic RPN and DCRF for small PWML segmentation on T1w neonatal MRI but supplies no cohort size, split protocol, or named baseline, so the reported outperformance cannot be evaluated.

read the letter

The core of this work is a two-stage network that starts from Faster R-CNN, adds a heuristic region proposal step meant to use local context around tiny lesions, swaps in a lightweight segmentation head, and finishes with dense CRF refinement. All of it runs on T1-weighted scans only. The authors report Dice 0.6616, sensitivity 0.7069, specificity 0.9997, and Hausdorff distance 52.91, and they state that the method beats prior art. They also released the code on GitHub, which is a concrete plus for anyone who wants to test the pipeline themselves. That combination of an established detector backbone with task-specific tweaks for very small lesions is the actual new piece; it is not a new theoretical framework but a targeted engineering extension for this narrow clinical target. The limitation that stands out is the complete absence of dataset size, patient demographics, train/validation/test protocol, or any description of the baseline method and how it was re-implemented. Without those elements the numerical claims are impossible to interpret or reproduce from the text alone. The stress-test note is accurate on this point. If the full manuscript later supplies a documented cohort of reasonable size, a clear cross-validation scheme, and a reproducible comparison, the method could be of interest to groups already working on neonatal white-matter lesion segmentation. As it stands, the paper is an incremental methods note whose central result cannot be assessed. I would bring it to a reading group only as a maybe, to discuss the heuristic RPN idea if the experiments turn out to be solid. It is worth sending to referees provided the missing experimental details are present in the full version, because the clinical task is real and the code release lowers the barrier to checking the claims.

Referee Report

3 major / 2 minor

Summary. The paper proposes Refined-Segmentation R-CNN (RS-RCNN), a two-stage CNN for semantic segmentation of punctate white matter lesions (PWML) on T1-weighted MRIs of preterm infants. It introduces a heuristic RPN (H-RPN) to exploit surrounding context, a lightweight segmentation head, and post-processing with densely connected CRF (DCRF). Using only T1w images, the method reports DSC = 0.6616, sensitivity = 0.7069, specificity = 0.9997 and Hausdorff distance = 52.9130, claiming to outperform prior state-of-the-art algorithms. Code is released on GitHub.

Significance. If the reported metrics are obtained on a properly sized, multi-center cohort with documented splits and reproducible baselines, the work would offer a clinically relevant tool for early PWML detection that relies solely on T1w contrast. The heuristic use of surrounding context for small/pixel-scale lesions and the lightweight design are potentially useful contributions to neonatal neuroimaging segmentation.

major comments (3)

[Experimental results] Experimental results / methods: the abstract and manuscript report concrete performance numbers and an outperformance claim, yet supply no cohort size, acquisition parameters, train/validation/test split ratios, or cross-validation protocol. These omissions are load-bearing for the central empirical claim.
[Experimental results] Comparison to prior work: the state-of-the-art baseline is never named and no description is given of whether the comparison used identical data, identical splits, or the same evaluation protocol. Without this, the headline numerical superiority cannot be verified.
[Experimental results] Statistical grounding: no p-values, confidence intervals, or statistical tests are reported for the claimed improvements over the baseline. This is required to support the assertion that the method “outperforms the state-of-the-art.”

minor comments (2)

[Abstract] Abstract: the phrase “segment the lesion of ordinary size or even pixel size” is imprecise; clarify what lesion-size range is actually evaluated.
[Methods] The manuscript should explicitly state whether the reported metrics are obtained on a held-out test set or on cross-validation folds.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for the detailed and constructive feedback. The comments correctly identify several omissions in the experimental reporting that undermine the verifiability of our claims. We will perform a major revision to address all three points by adding the required details, naming baselines, and including statistical tests. These changes will be incorporated into the next version of the manuscript.

read point-by-point responses

Referee: [Experimental results] Experimental results / methods: the abstract and manuscript report concrete performance numbers and an outperformance claim, yet supply no cohort size, acquisition parameters, train/validation/test split ratios, or cross-validation protocol. These omissions are load-bearing for the central empirical claim.

Authors: We acknowledge that these details were omitted from the main text. In the revised manuscript we will insert a new 'Materials and Methods' subsection that explicitly states the cohort size (number of subjects and images), MRI acquisition parameters (field strength, sequence details, resolution), the train/validation/test split ratios, and the cross-validation protocol (including number of folds). This information will be moved from any supplementary material into the primary narrative to make the empirical results fully reproducible. revision: yes
Referee: [Experimental results] Comparison to prior work: the state-of-the-art baseline is never named and no description is given of whether the comparison used identical data, identical splits, or the same evaluation protocol. Without this, the headline numerical superiority cannot be verified.

Authors: We agree that the baseline must be named and the comparison protocol clarified. The revised paper will explicitly name the prior state-of-the-art method, state that the comparison was performed on the identical dataset and splits, and confirm that the same evaluation metrics and protocol were applied. Any differences in implementation will be noted to allow readers to assess fairness. revision: yes
Referee: [Experimental results] Statistical grounding: no p-values, confidence intervals, or statistical tests are reported for the claimed improvements over the baseline. This is required to support the assertion that the method “outperforms the state-of-the-art.”

Authors: We recognize the need for statistical support. In the revision we will add p-values (from paired statistical tests across cross-validation folds) and 95% confidence intervals for the key metrics (Dice, sensitivity, etc.) when comparing against the baseline. These will be reported in the results tables and text. revision: yes

Circularity Check

0 steps flagged

No circularity; empirical CNN training and evaluation

full rationale

The paper proposes an RS-RCNN architecture (heuristic RPN + lightweight segmentation head + DCRF) and reports empirical metrics (DSC 0.6616, sensitivity 0.7069, etc.) obtained by training and testing the network on T1w MRI data. No equations, parameter-fitting steps, or derivations appear that would reduce any reported quantity to a fitted input or self-defined quantity by construction. The result is a standard empirical ML claim whose validity depends on external data and validation protocol rather than internal self-reference.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

The central claim depends on standard supervised training of a convolutional network on an unspecified collection of infant MRI scans; no explicit free parameters, axioms, or invented physical entities are stated in the abstract.

pith-pipeline@v0.9.0 · 5794 in / 1099 out tokens · 42383 ms · 2026-05-25T17:31:23.412676+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

16 extracted references · 16 canonical work pages

[1]

A., et al .: Comparing 3T T1- weighted sequences in identifying hyperintense punctate lesions in preterm neonates

Tortora, D., Panara, V., Mattei, P. A., et al .: Comparing 3T T1- weighted sequences in identifying hyperintense punctate lesions in preterm neonates. American Journal of Neuro- radiology 36(3), 581-586 (2015)

work page 2015
[2]

J., Benders, M

Kersbergen, K. J., Benders, M. J., Groenendaal, F., et al .: Different patterns of punctate white matter lesions in serially scanned preterm infants. PloS one 9(10), e108904 (2014)

work page 2014
[3]

S., Wan, M

Li, X., Gao, J., Wang, M., Zheng, J., Li, Y., Hui, E. S., Wan, M. & Yang, J.: Characteriza- tion of extensive microstructural variations associated with punctate white matter lesions in preterm neonates. American Journal of Neuroradiology 38(6), 1228-1234 (2017)

work page 2017
[4]

: White matter injury detection in neonatal MRI

Cheng, I., et al. : White matter injury detection in neonatal MRI. Proceedings of the Inter- national Society for Optical Engineering. vol.8670, pp.86702L. SPIE, Florida (2013)

work page 2013
[5]

P., Duerden, E

Cheng, I., Miller, S. P., Duerden, E. G., et al.: Stochastic process for white matter injury detection in preterm neonates. NeuroImage: Clinical 7, 622-630 (2015)

work page 2015
[6]

Medical & Biological Engineer- ing & Computing 57(1), 71-87 (2019)

Mukherjee, S., Cheng, I., Miller, S., et al .: A fast segmentation -free fully automated ap- proach to white matter injury detection in preterm infants. Medical & Biological Engineer- ing & Computing 57(1), 71-87 (2019)

work page 2019
[7]

A., Kroll, C., et al.: Hough-CNN: deep learning for segmentation of deep brain regions in MRI and ultrasound

Milletari, F., Ahmadi, S. A., Kroll, C., et al.: Hough-CNN: deep learning for segmentation of deep brain regions in MRI and ultrasound. Computer Vision and Image Understanding 164, 92-102 (2017)

work page 2017
[8]

In: International Conference on Medical Image Computing and Computer -Assisted Intervention, pp

Ghafoorian, M., Mehrtash, A., Kapur, T., et al .: Transfer learning for domain adaptation in MRI: Application in brain lesion segmentation. In: International Conference on Medical Image Computing and Computer -Assisted Intervention, pp. 516-524. Springer, Cham (2017)

work page 2017
[9]

In: International Conference on Computer Vision, pp

He, K., Gkioxari, G., Doll´ ar, P., Girshick, R.: Mask R-CNN. In: International Conference on Computer Vision, pp. 2980–2988 (2017)

work page 2017
[10]

In: NIPS, pp

Ren, S., He, K., Girshick, R., Sun, J.: Faster R -CNN: towards real -time object detection with region proposal networks. In: NIPS, pp. 91–99 (2015)

work page 2015
[11]

In: IEEE Computer Vision Pattern Recognition, pp

He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: IEEE Computer Vision Pattern Recognition, pp. 770–778 (2016)

work page 2016
[12]

Y., Dollár, P., Girshick, R., He, K., Hariharan, B., Belongie, S.: Feature pyramid networks for object detection

Lin, T. Y., Dollár, P., Girshick, R., He, K., Hariharan, B., Belongie, S.: Feature pyramid networks for object detection. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pp. 2117-2125 (2017)

work page 2017
[13]

In: NIPS, pp

Kr¨ ahenb¨ uhl, P., Koltun, V.: Efficient inference in fully connected CRFs with gaussian edge potentials. In: NIPS, pp. 1–9 (2012). 9

work page 2012
[14]

C., Papandreou, G., Kokkinos, I., et al.: Deeplab: Semantic image segmentation with deep convolutional nets, atrous convolution, and fully connected crfs

Chen, L. C., Papandreou, G., Kokkinos, I., et al.: Deeplab: Semantic image segmentation with deep convolutional nets, atrous convolution, and fully connected crfs. IEEE Transac- tions on Pattern Analysis and Machine Intelligence 40(4), 834-848 (2018)

work page 2018
[15]

F., et al.: Efficient multi -scale 3D CNN with ful- ly connected CRF for accurate brain lesion segmentation

Kamnitsas, K., Ledig, C., Newcombe, V. F., et al.: Efficient multi -scale 3D CNN with ful- ly connected CRF for accurate brain lesion segmentation. Medical Image Analysis 36, 61 - 78 (2017)

work page 2017
[16]

P.: Practical bayesian optimization of machine learning algorithms

Snoek, J., Larochelle, H., Adams, R. P.: Practical bayesian optimization of machine learning algorithms. In Advances in neural information processing systems, pp. 2951 -2959 (2012)

work page 2012

[1] [1]

A., et al .: Comparing 3T T1- weighted sequences in identifying hyperintense punctate lesions in preterm neonates

Tortora, D., Panara, V., Mattei, P. A., et al .: Comparing 3T T1- weighted sequences in identifying hyperintense punctate lesions in preterm neonates. American Journal of Neuro- radiology 36(3), 581-586 (2015)

work page 2015

[2] [2]

J., Benders, M

Kersbergen, K. J., Benders, M. J., Groenendaal, F., et al .: Different patterns of punctate white matter lesions in serially scanned preterm infants. PloS one 9(10), e108904 (2014)

work page 2014

[3] [3]

S., Wan, M

Li, X., Gao, J., Wang, M., Zheng, J., Li, Y., Hui, E. S., Wan, M. & Yang, J.: Characteriza- tion of extensive microstructural variations associated with punctate white matter lesions in preterm neonates. American Journal of Neuroradiology 38(6), 1228-1234 (2017)

work page 2017

[4] [4]

: White matter injury detection in neonatal MRI

Cheng, I., et al. : White matter injury detection in neonatal MRI. Proceedings of the Inter- national Society for Optical Engineering. vol.8670, pp.86702L. SPIE, Florida (2013)

work page 2013

[5] [5]

P., Duerden, E

Cheng, I., Miller, S. P., Duerden, E. G., et al.: Stochastic process for white matter injury detection in preterm neonates. NeuroImage: Clinical 7, 622-630 (2015)

work page 2015

[6] [6]

Medical & Biological Engineer- ing & Computing 57(1), 71-87 (2019)

Mukherjee, S., Cheng, I., Miller, S., et al .: A fast segmentation -free fully automated ap- proach to white matter injury detection in preterm infants. Medical & Biological Engineer- ing & Computing 57(1), 71-87 (2019)

work page 2019

[7] [7]

A., Kroll, C., et al.: Hough-CNN: deep learning for segmentation of deep brain regions in MRI and ultrasound

Milletari, F., Ahmadi, S. A., Kroll, C., et al.: Hough-CNN: deep learning for segmentation of deep brain regions in MRI and ultrasound. Computer Vision and Image Understanding 164, 92-102 (2017)

work page 2017

[8] [8]

In: International Conference on Medical Image Computing and Computer -Assisted Intervention, pp

Ghafoorian, M., Mehrtash, A., Kapur, T., et al .: Transfer learning for domain adaptation in MRI: Application in brain lesion segmentation. In: International Conference on Medical Image Computing and Computer -Assisted Intervention, pp. 516-524. Springer, Cham (2017)

work page 2017

[9] [9]

In: International Conference on Computer Vision, pp

He, K., Gkioxari, G., Doll´ ar, P., Girshick, R.: Mask R-CNN. In: International Conference on Computer Vision, pp. 2980–2988 (2017)

work page 2017

[10] [10]

In: NIPS, pp

Ren, S., He, K., Girshick, R., Sun, J.: Faster R -CNN: towards real -time object detection with region proposal networks. In: NIPS, pp. 91–99 (2015)

work page 2015

[11] [11]

In: IEEE Computer Vision Pattern Recognition, pp

He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: IEEE Computer Vision Pattern Recognition, pp. 770–778 (2016)

work page 2016

[12] [12]

Y., Dollár, P., Girshick, R., He, K., Hariharan, B., Belongie, S.: Feature pyramid networks for object detection

Lin, T. Y., Dollár, P., Girshick, R., He, K., Hariharan, B., Belongie, S.: Feature pyramid networks for object detection. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pp. 2117-2125 (2017)

work page 2017

[13] [13]

In: NIPS, pp

Kr¨ ahenb¨ uhl, P., Koltun, V.: Efficient inference in fully connected CRFs with gaussian edge potentials. In: NIPS, pp. 1–9 (2012). 9

work page 2012

[14] [14]

C., Papandreou, G., Kokkinos, I., et al.: Deeplab: Semantic image segmentation with deep convolutional nets, atrous convolution, and fully connected crfs

Chen, L. C., Papandreou, G., Kokkinos, I., et al.: Deeplab: Semantic image segmentation with deep convolutional nets, atrous convolution, and fully connected crfs. IEEE Transac- tions on Pattern Analysis and Machine Intelligence 40(4), 834-848 (2018)

work page 2018

[15] [15]

F., et al.: Efficient multi -scale 3D CNN with ful- ly connected CRF for accurate brain lesion segmentation

Kamnitsas, K., Ledig, C., Newcombe, V. F., et al.: Efficient multi -scale 3D CNN with ful- ly connected CRF for accurate brain lesion segmentation. Medical Image Analysis 36, 61 - 78 (2017)

work page 2017

[16] [16]

P.: Practical bayesian optimization of machine learning algorithms

Snoek, J., Larochelle, H., Adams, R. P.: Practical bayesian optimization of machine learning algorithms. In Advances in neural information processing systems, pp. 2951 -2959 (2012)

work page 2012