pith. sign in

arxiv: 2605.20760 · v1 · pith:64SIZURMnew · submitted 2026-05-20 · 💻 cs.CV

SpineContextResUNet: A Computationally Efficient Residual UNet for Spine CT Segmentation

K S Nithurshen , Saurabh J. Shigwan This is my paper

Pith reviewed 2026-05-21 05:58 UTC · model grok-4.3

classification 💻 cs.CV
keywords spine segmentationCT imagingresidual U-Netcontext blockdilated convolutionslightweight modelsmedical image segmentationvertebral column
0
0 comments X

The pith

A residual U-Net with parallel multi-dilated convolutions segments the spine in CT scans while running on commodity hardware where transformers and large ensembles fail.

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

The paper introduces SpineContextResUNet to automate vertebral column segmentation in CT scans without demanding high-end GPUs. It adds a lightweight context block of parallel multi-dilated convolutions to a 3D residual U-Net so that long-range anatomical context can be gathered at low latency and memory cost. On the VerSe2020 and CTSpine1K benchmarks the model reaches Dice scores of 88.17 percent and 88.13 percent while staying inside a footprint of roughly 1.7 million parameters. The same hardware budget causes a size-matched transformer to lose accuracy and causes heavier baselines to exhaust memory on an Intel Core i5 with 8 GB RAM. This matters because it could make reliable spine segmentation available for point-of-care diagnostics and edge devices such as the Nvidia Jetson Orin Nano.

Core claim

SpineContextResUNet integrates a lightweight Context Block that employs parallel multi-dilated convolutions into a 3D Residual U-Net to capture long-range anatomical dependencies without RNN latency or self-attention memory overhead. The resulting model achieves Dice scores of 88.17 percent on VerSe2020 and 88.13 percent on CTSpine1K while supporting inference on commodity hardware where a constrained SwinUNETR degrades and TotalSegmentator fails due to memory exhaustion.

What carries the argument

The lightweight Context Block that runs parallel multi-dilated convolutions to gather long-range spatial context inside the residual U-Net.

If this is right

  • The model reaches Dice scores above 88 percent on two public spine CT benchmarks while using only about 1.7 million parameters.
  • Inference succeeds on an Intel Core i5 processor with 8 GB RAM and on the Nvidia Jetson Orin Nano.
  • A size-matched transformer loses accuracy in the same limited-data setting because it lacks spatial inductive biases.
  • Heavy ensemble baselines cannot load into memory on the same commodity hardware.

Where Pith is reading between the lines

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

  • The same block design could be inserted into other 3D segmentation networks that need distant context but must stay small.
  • Further parameter reduction might allow real-time spine segmentation directly on portable ultrasound or X-ray devices.
  • Testing generalization across scanner vendors would clarify how much the dilated-convolution context compensates for limited training data.

Load-bearing premise

The parallel multi-dilated convolutions supply enough long-range anatomical context to keep segmentation accuracy high without attention mechanisms or greater network capacity.

What would settle it

A large drop in Dice score on CT scans from unseen scanners or patient cohorts would show that the context block does not supply adequate long-range dependencies.

Figures

Figures reproduced from arXiv: 2605.20760 by K S Nithurshen, Saurabh J. Shigwan.

Figure 1
Figure 1. Figure 1: Overview of the proposed SpineContextResUNet architecture. Blocks [8]. Each residual block consists of two 3 × 3 × 3 convolutions with Batch Normalization (BN) and ReLU activation, linked by a residual shortcut connection. This design facilitates gradient flow and allows for the training of deeper networks without degradation. The network processes a 3D input volume X ∈ R 1×D×H×W . It consists of three enc… view at source ↗
Figure 2
Figure 2. Figure 2: The visualization confirms the efficacy of the proposed Parallel Context Block. The Left Panel displays the activation map of SpineContextResUNet, which demonstrates a cohesive and high-intensity focus strictly aligned with the vertebral bodies. The constrained SwinUNETR [1] exhibits erroneous high-intensity activations on the ribs and soft tissue. Because the Transformer’s capacity was artificially limite… view at source ↗
Figure 2
Figure 2. Figure 2: Grad-CAM Visualization of Vertebral Attention. [PITH_FULL_IMAGE:figures/full_fig_p008_2.png] view at source ↗
read the original abstract

Automated segmentation of the vertebral column in Computed Tomography (CT) scans is a prerequisite for pathological assessment and surgical planning. However, state-of-the-art methods, particularly those based on Transformers or large-scale ensembles, demand substantial GPU resources, creating a barrier for clinical adoption in resource-constrained environments or on edge devices. To address this, we introduce SpineContextResUNet, a computationally efficient 3D Residual U-Net designed for rapid spinal localization. Our architecture integrates a lightweight Context Block that employs parallel multi-dilated convolutions to capture long-range anatomical dependencies without the high latency of Recurrent Neural Networks (RNNs) or the memory overhead of Self-Attention mechanisms. Extensive validation on two public benchmarks, VerSe2020 and CTSpine1K, demonstrates that our model achieves a Dice score of 88.17% and 88.13% respectively. To evaluate performance under strict hardware constraints, we compared our model against a bottlenecked SwinUNETR scaled to match our ~1.7M hardware footprint. While the constrained Transformer suffers severe performance degradation due to a lack of spatial inductive biases in a limited-data regime, our CNN-based approach successfully maintains high accuracy. Crucially, heavy baselines like TotalSegmentator fail due to memory exhaustion on commodity hardware (Intel Core i5, 8GB RAM), our model performs robust inference, making it a viable solution for point-of-care diagnostics and deployment on edge platforms like the Nvidia Jetson Orin Nano.

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.

Referee Report

2 major / 2 minor

Summary. The manuscript introduces SpineContextResUNet, a computationally efficient 3D Residual U-Net for vertebral column segmentation in CT scans. It integrates a lightweight Context Block with parallel multi-dilated convolutions to capture long-range dependencies without RNN latency or self-attention memory overhead. The central claims are Dice scores of 88.17% on VerSe2020 and 88.13% on CTSpine1K, plus robust inference on commodity hardware (Intel Core i5, 8GB RAM) where TotalSegmentator fails due to memory exhaustion and a parameter-matched SwinUNETR degrades severely due to missing spatial inductive biases in limited-data regimes.

Significance. If the empirical results and baseline comparison hold after clarification, the work offers a practical, low-footprint (~1.7M parameters) CNN alternative for spine segmentation that could enable deployment on edge devices and resource-constrained clinical environments. The focus on inductive biases versus attention mechanisms in limited-data medical imaging is timely and could inform efficient architecture design.

major comments (2)
  1. [Experiments] Experiments section (and abstract): The scaling procedure for the 'bottlenecked SwinUNETR scaled to match our ~1.7M hardware footprint' is unspecified. No details are given on whether this involved uniform channel reduction, layer pruning, or other modifications, nor is there an ablation confirming that performance degradation arises from absent spatial inductive biases rather than reduced capacity alone. This directly undercuts the load-bearing contrast drawn between the Context Block and the constrained Transformer.
  2. [Methods/Experiments] Methods and Experiments sections: Training protocol details (data splits, augmentations, optimizer, loss, epochs, and any statistical tests or confidence intervals on the reported Dice scores) are absent. Without these, the concrete performance numbers cannot be reproduced or verified, weakening the central empirical claims.
minor comments (2)
  1. [Abstract/Experiments] The abstract and text refer to 'heavy baselines like TotalSegmentator' failing on 8GB RAM; clarify the exact memory footprint measured and whether any optimization (e.g., mixed precision) was attempted for fairness.
  2. [Architecture] Notation for the Context Block (parallel multi-dilated convolutions) should include a diagram or explicit equation showing dilation rates and fusion to improve clarity.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive feedback on our manuscript. We address each major comment below and will revise the paper accordingly to improve clarity and reproducibility.

read point-by-point responses

  1. Referee: [Experiments] Experiments section (and abstract): The scaling procedure for the 'bottlenecked SwinUNETR scaled to match our ~1.7M hardware footprint' is unspecified. No details are given on whether this involved uniform channel reduction, layer pruning, or other modifications, nor is there an ablation confirming that performance degradation arises from absent spatial inductive biases rather than reduced capacity alone. This directly undercuts the load-bearing contrast drawn between the Context Block and the constrained Transformer.

    Authors: We acknowledge the need for greater detail on the SwinUNETR scaling. The procedure involved uniform reduction of channel dimensions across all layers to reach approximately 1.7M parameters while preserving the overall architecture. This ensures a direct comparison under identical hardware constraints. We attribute the observed degradation to the lack of spatial inductive biases in the limited-data regime, as our CNN model with the Context Block maintains performance. In the revision we will explicitly describe the scaling method and add a brief discussion or supporting analysis to separate capacity effects from inductive bias effects. revision: yes

  2. Referee: [Methods/Experiments] Methods and Experiments sections: Training protocol details (data splits, augmentations, optimizer, loss, epochs, and any statistical tests or confidence intervals on the reported Dice scores) are absent. Without these, the concrete performance numbers cannot be reproduced or verified, weakening the central empirical claims.

    Authors: We agree that these implementation details were omitted from the original submission. The revised manuscript will include the exact data splits, augmentation pipeline, optimizer settings, loss function, training epochs, and any statistical tests or confidence intervals associated with the Dice scores to support full reproducibility. revision: yes

Circularity Check

0 steps flagged

No circularity: empirical validation on independent benchmarks

full rationale

The paper proposes a new CNN architecture (SpineContextResUNet with Context Block) motivated by efficiency needs and evaluates it via standard Dice scores on external public datasets (VerSe2020, CTSpine1K). No derivation chain exists that reduces predictions or uniqueness claims to self-referential fits, self-citations, or definitional loops; architectural choices and performance claims are independent of the reported metrics and rest on reproducible benchmark comparisons rather than tautological constructions.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

The design rests on the domain assumption that multi-dilated parallel convolutions provide sufficient context for vertebral anatomy; no free parameters or invented entities are explicitly introduced beyond standard network hyperparameters.

axioms (1)
  • domain assumption Parallel multi-dilated convolutions capture long-range anatomical dependencies without RNN latency or attention memory cost
    Invoked to justify the Context Block as a lightweight replacement for recurrent or attention mechanisms.

pith-pipeline@v0.9.0 · 5805 in / 1172 out tokens · 32216 ms · 2026-05-21T05:58:00.318712+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

What do these tags mean?
matches
The paper's claim is directly supported by a theorem in the formal canon.
supports
The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends
The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses
The paper appears to rely on the theorem as machinery.
contradicts
The paper's claim conflicts with a theorem or certificate in the canon.
unclear
Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Reference graph

Works this paper leans on

21 extracted references · 21 canonical work pages · 2 internal anchors

  1. [1]

    Swinunetr-v2: Stronger swin transformers with stagewise convolutions for 3d med- 8 ical image segmentation

    Yufan He, Vishwesh Nath, Dong Yang, Yucheng Tang, Andriy Myronenko, and Daguang Xu. Swinunetr-v2: Stronger swin transformers with stagewise convolutions for 3d med- 8 ical image segmentation. InInternational Conference on Medical Image Computing and Computer-Assisted Intervention, pages 416–426. Springer, 2023

    work page 2023

  2. [2]

    An image is worth 16x16 words: Transformers for image recognition at scale

    Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. In International Conference on Learning Representations, 2021

    work page 2021

  3. [3]

    nnu-net: a self-configuring method for deep learning-based biomedical image segmentation

    Fabian Isensee, Paul F Jaeger, Simon AA Kohl, Jens Petersen, and Klaus H Maier-Hein. nnu-net: a self-configuring method for deep learning-based biomedical image segmentation. Nature methods, 18(2):203–211, 2021

    work page 2021

  4. [4]

    Segment anything

    Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. InProceedings of the IEEE/CVF international conference on computer vision, pages 4015–4026, 2023

    work page 2023

  5. [5]

    Segment anything in medical images.Nature communications, 15(1):654, 2024

    Jun Ma, Yuting He, Feifei Li, Lin Han, Chenyu You, and Bo Wang. Segment anything in medical images.Nature communications, 15(1):654, 2024

    work page 2024

  6. [6]

    A 3d coarse-to-fine framework for volumetric medical image segmentation

    Zhuotun Zhu, Yingda Xia, Wei Shen, Elliot Fishman, and Alan Yuille. A 3d coarse-to-fine framework for volumetric medical image segmentation. In2018 International conference on 3D vision (3DV), pages 682–690. IEEE, 2018

    work page 2018

  7. [7]

    3d u-net: learning dense volumetric segmentation from sparse annotation

    ¨Ozg¨ un C ¸ i¸ cek, Ahmed Abdulkadir, Soeren S Lienkamp, Thomas Brox, and Olaf Ron- neberger. 3d u-net: learning dense volumetric segmentation from sparse annotation. In International conference on medical image computing and computer-assisted intervention, pages 424–432. Springer, 2016

    work page 2016

  8. [8]

    Resunet-a: A deep learning framework for semantic segmentation of remotely sensed data.ISPRS Journal of Photogrammetry and Remote Sensing, 162:94–114, 2020

    Foivos I Diakogiannis, Fran¸ cois Waldner, Peter Caccetta, and Chen Wu. Resunet-a: A deep learning framework for semantic segmentation of remotely sensed data.ISPRS Journal of Photogrammetry and Remote Sensing, 162:94–114, 2020

    work page 2020

  9. [9]

    Haofu Liao, Addisu Mesfin, and Jiebo Luo. Joint vertebrae identification and localiza- tion in spinal ct images by combining short-and long-range contextual information.IEEE transactions on medical imaging, 37(5):1266–1275, 2018

    work page 2018

  10. [10]

    Spineparsenet: spine parsing for volumetric mr image by a two-stage segmentation framework with semantic image representation

    Shumao Pang, Chunlan Pang, Lei Zhao, Yangfan Chen, Zhihai Su, Yujia Zhou, Meiyan Huang, Wei Yang, Hai Lu, and Qianjin Feng. Spineparsenet: spine parsing for volumetric mr image by a two-stage segmentation framework with semantic image representation. IEEE Transactions on Medical Imaging, 40(1):262–273, 2020

    work page 2020

  11. [11]

    Liang-Chieh Chen, George Papandreou, Iasonas Kokkinos, Kevin Murphy, and Alan L Yuille. Deeplab: Semantic image segmentation with deep convolutional nets, atrous con- volution, and fully connected crfs.IEEE transactions on pattern analysis and machine intelligence, 40(4):834–848, 2017

    work page 2017

  12. [12]

    Rethinking Atrous Convolution for Semantic Image Segmentation

    Liang-Chieh Chen, George Papandreou, Florian Schroff, and Hartwig Adam. Rethinking atrous convolution for semantic image segmentation.arXiv preprint arXiv:1706.05587, 2017

    work page internal anchor Pith review Pith/arXiv arXiv 2017

  13. [13]

    V-net: Fully convolutional neural networks for volumetric medical image segmentation

    Fausto Milletari, Nassir Navab, and Seyed-Ahmad Ahmadi. V-net: Fully convolutional neural networks for volumetric medical image segmentation. In2016 fourth international conference on 3D vision (3DV), pages 565–571. Ieee, 2016. 9

    work page 2016

  14. [14]

    Combo loss: Handling input and output imbalance in multi-organ segmentation.Computerized Medical Imaging and Graphics, 75:24–33, 2019

    Saeid Asgari Taghanaki, Yefeng Zheng, S Kevin Zhou, Bogdan Georgescu, Puneet Sharma, Daguang Xu, Dorin Comaniciu, and Ghassan Hamarneh. Combo loss: Handling input and output imbalance in multi-organ segmentation.Computerized Medical Imaging and Graphics, 75:24–33, 2019

    work page 2019

  15. [15]

    A vertebral segmentation dataset with fracture grading.Radiology: Artificial Intelligence, 2(4):e190138, 2020

    Maximilian T L¨ offler, Anjany Sekuboyina, Alina Jacob, Anna-Lena Grau, Andreas Scharr, Malek El Husseini, Mareike Kallweit, Claus Zimmer, Thomas Baum, and Jan S Kirschke. A vertebral segmentation dataset with fracture grading.Radiology: Artificial Intelligence, 2(4):e190138, 2020

    work page 2020

  16. [16]

    Verse: a vertebrae labelling and segmentation benchmark for multi-detector ct images.Medical image analysis, 73:102166, 2021

    Anjany Sekuboyina, Malek E Husseini, Amirhossein Bayat, Maximilian L¨ offler, Hans Liebl, Hongwei Li, Giles Tetteh, Jan Kukaˇ cka, Christian Payer, Darko ˇStern, et al. Verse: a vertebrae labelling and segmentation benchmark for multi-detector ct images.Medical image analysis, 73:102166, 2021

    work page 2021

  17. [17]

    A computed tomography vertebral segmentation dataset with anatomical variations and multi-vendor scanner data.Scientific data, 8(1):284, 2021

    Hans Liebl, David Schinz, Anjany Sekuboyina, Luca Malagutti, Maximilian T L¨ offler, Amirhossein Bayat, Malek El Husseini, Giles Tetteh, Katharina Grau, Eva Niederreiter, et al. A computed tomography vertebral segmentation dataset with anatomical variations and multi-vendor scanner data.Scientific data, 8(1):284, 2021

    work page 2021

  18. [18]

    Yang Deng, Ce Wang, Yuan Hui, Qian Li, Jun Li, Shiwei Luo, Mengke Sun, Quan Quan, Shuxin Yang, You Hao, et al. Ctspine1k: A large-scale dataset for spinal vertebrae seg- mentation in computed tomography.Machine Learning for Biomedical Imaging, 3(Special Issue on MICCAI Open Data 2024-2025):824–832, 2025

    work page 2024

  19. [19]

    Totalsegmentator: robust segmentation of 104 anatomic structures in ct images.Radiology: Artificial Intelligence, 5(5):e230024, 2023

    Jakob Wasserthal, Hanns-Christian Breit, Manfred T Meyer, Maurice Pradella, Daniel Hinck, Alexander W Sauter, Tobias Heye, Daniel T Boll, Joshy Cyriac, Shan Yang, et al. Totalsegmentator: robust segmentation of 104 anatomic structures in ct images.Radiology: Artificial Intelligence, 5(5):e230024, 2023

    work page 2023

  20. [20]

    Grad-cam: Visual explanations from deep networks via gradient- based localization

    Ramprasaath R Selvaraju, Michael Cogswell, Abhishek Das, Ramakrishna Vedantam, Devi Parikh, and Dhruv Batra. Grad-cam: Visual explanations from deep networks via gradient- based localization. InProceedings of the IEEE international conference on computer vision, pages 618–626, 2017

    work page 2017

  21. [21]

    Multi-Scale Context Aggregation by Dilated Convolutions

    Fisher Yu and Vladlen Koltun. Multi-scale context aggregation by dilated convolutions. arXiv preprint arXiv:1511.07122, 2015. 10

    work page internal anchor Pith review Pith/arXiv arXiv 2015