Machine-agnostic Automated Lumbar MRI Segmentation using a Cascaded Model Based on Generative Neurons

Anwarul Hasan; Enamul Hoque Bhuiyan; Israa Al-Hashimi; Muhammad E. H. Chowdhury; Muhammad Salman Khan; Promit Basak; Rusab Sarmun; Saidul Kabir

REVIEW 5 major objections 6 minor 12 references

Machine-agnostic Automated Lumbar MRI Segmentation using a Cascaded Model Based on Generative Neurons

T0 review · 5 major / 6 minor · reviewed 2026-08-12 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read A cascaded YOLOv8 plus Self-ONN model segments each lumbar vertebra and disc across 12 MRI scanners.

desk verdict Competent lumbar MRI segmentation pipeline whose 'machine-agnostic' and 'first of its kind' claims outrun the evidence; scanner-disjoint evaluation needed. read the letter →

arxiv 2411.15656 v1 pith:C7WXDFM2 submitted 2024-11-23 eess.IV cs.CVcs.LG

Promit Basak , Rusab Sarmun , Saidul Kabir , Israa Al-Hashimi , Enamul Hoque Bhuiyan , Anwarul Hasan , Muhammad Salman Khan , Muhammad E. H. Chowdhury This is my paper · ORCID

classification eess.IVcs.CVcs.LG

keywords lumbarspineMRIsegmentationintervertebraldiscvertebralbodySelf-ONNgenerativeneuronsYOLOv8machine-agnosticmulti-scanner

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

This paper tries to establish that lumbar spine MRI segmentation can be made machine-agnostic: one automated pipeline can locate and outline each lumbar vertebral body (L1–L5) and intervertebral disc (L1-2 through L4-5) in images from many different scanners. The proposed method is a cascade: a YOLOv8 medium detector first crops the lumbar region from each slice, and a Self-ONN-based encoder-decoder network with a DenseNet121 encoder then produces nine-class segmentation masks. On a 34-subject, 12-scanner dataset, evaluated with subject-wise 10-fold cross-validation, the authors report a mean Intersection over Union of 83.66%, a Dice similarity coefficient of 91.03%, and a sensitivity of 91.44%. They also report that the cascaded design matters: using the detected ROI improves Dice by roughly 7 points and IoU by roughly 10 points over running the segmenter on full images, and a ceiling analysis with perfect ROIs reaches 93.73% Dice. If the claims hold, automated per-level spine measurement and diagnosis support become feasible across scanner vendors without retraining.

What carries the argument

The load-bearing mechanism is the cascade itself plus the generative neurons of Self-ONN. The YOLOv8 medium model localizes the lumbar spine (mAP 0.5:0.95 of 0.916), and the segmenter is a DenseNet121 encoder feeding a five-block Self-ONN decoder with skip connections and no bottleneck. A Self-ONN generative neuron replaces the fixed linear convolution of a CNN with a learnable composite nodal operator expressed as a Taylor polynomial of order Q; the neuron output is the sum of Q convolutions of the input raised to powers 1 through Q, so each neuron can synthesize nonlinear operators during backpropagation. This is what the paper credits for capturing tissue boundaries that low-contrast MRI makes indistinct. The ground-truth masks themselves are formed by fusing manual annotations with the STAPLE algorithm, and the evaluation uses subject-preserved 10-fold cross-validation to prevent data leakage.

What would settle it

Run the trained cascade on held-out T2, STIR, and Dixon volumes, or on a scanner absent from the 12 used, without any slice trimming, and measure per-class Dice and IoU; a substantial drop would falsify the machine-agnostic claim.

Watch

Extended reading notes

Core claim

The central discovery is that a two-stage architecture—ROI extraction followed by segmentation—lets a Self-ONN decoder with a DenseNet121 encoder segment individual lumbar vertebrae and intervertebral discs at 91.03% Dice and 83.66% mean IoU, averaged over ten subject-disjoint folds, using T1-weighted images from 12 MRI scanners. This is presented as the first system to segment each lumbar vertebral body and each disc as separate classes rather than treating the spine as one or two structures. The authors also show that the ROI stage is the current bottleneck: when ground-truth ROIs replace detector outputs, Dice rises to 93.73%, whereas removing the ROI stage entirely drops Dice to 74.1%.

Load-bearing premise

The load-bearing premise is that T1-weighted slices, after trimming 20% from each volume and deleting 53 corrupted scans, fairly represent the multi-scanner clinical reality the model claims to handle; if the excluded modalities or edge slices behave differently, the machine-agnostic performance numbers could overstate real-world accuracy.

Editorial extensions

If this is right

A single trained pipeline can label each lumbar vertebra and disc separately across scanner vendors, enabling automated level-specific measurements such as disc height, vertebral area, or stenosis indices.
The cascade improves segmentation by roughly 7 Dice points over whole-image segmentation, so localizing the anatomy before pixel classification is a practical strategy for low-contrast medical images.
Because the ceiling analysis shows 93.73% Dice with perfect ROIs, improving the detector or adding slice-level rejection is the clearest path to higher overall accuracy.
The Self-ONN decoder outperforms standard U-Net and U-Net++ decoders with the same DenseNet121 encoder on this dataset, suggesting learnable nonlinear operators help in this modality.
Subject-wise 10-fold cross-validation gives a more realistic estimate of generalization than random splits, and the reported numbers are the averaged per-fold results, not a single lucky test set.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

If the T1-only training set is representative, the machine-agnostic claim should be read as scanner-agnostic for T1; the paper's own limitation note concedes that other MRI modalities remain untested.
The heuristic removal of the first and last 20% of slices and the exclusion of 53 corrupted scans could bias the reported metrics upward, so a prospective test on full, unedited volumes would give a truer clinical picture.
The same ROI-then-Self-ONN recipe could transfer to other anatomical regions where background dominates the image and target structures vary across patients, such as cervical spine or knee cartilage segmentation.
With per-class labels, this type of model could feed downstream decision tools that report exactly which disc level is abnormal, rather than only flagging the presence of a lesion.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

The headline is that this is a competent engineering paper whose central claims are two sizes too big. The cascade—YOLOv8m ROI detector plus DenseNet121 encoder and Self-ONN decoder—is new on this public multi-scanner dataset, and the reported 91.03% DSC/83.66% IoU are honestly measured with subject-preserved 10-fold CV. The ceiling analysis and ablations are a plus; they show the ROI stage costs about 2.7 DSC and that preprocessing matters. That part is useful and reproducible.

The soft spots are in the framing, not the code. 'Machine-agnostic' is only supported for T1-weighted images: T2, STIR, and Dixon are used in ROI detection but the segmentation model never sees them. So at best you have scanner-agnostic T1 segmentation. Worse, the CV is subject-preserved but not scanner-disjoint. With 34 subjects and 12 scanners, a scanner can contribute subjects to both training and test folds, which lets the model memorize scanner-specific intensity/artifact patterns. The reported numbers can therefore overstate cross-scanner generalization. The paper needs a scanner-disjoint split or per-scanner held-out analysis before that claim flies.

The 'first of its kind' novelty claim is contradicted by their own Related Works: SAFNet (Yilizati-Yilihamu 2023) already segments individual vertebrae and discs, and Spine-GAN (Han 2018) segments multiple spinal structures. That sentence should be deleted. The preprocessing—dropping 20% of slices and 53 corrupted scans—could bias accuracy upward, but the authors disclose it; I'd call that a minor issue, not a fatal one. The comparison table mixes datasets and class definitions, and the only same-dataset comparison (Kim et al. 2024) is within 1 point on DSC while segmenting only discs, so 'excellent performance' relative to SOTA is not really established.

Who gets value: groups working on lumbar MRI segmentation who want a baseline on the multi-scanner dataset. It deserves a serious referee—the empirical work is real and the dataset choice is useful—but the decision should be major revision, with scanner-disjoint evaluation and scaled-back claims.

Referee Report

5 major / 6 minor

Summary. The paper presents a two-stage deep learning pipeline for lumbar spine MRI segmentation: a YOLOv8-medium detector localizes the lumbar region, and a DenseNet121 encoder with a Self-ONN decoder segments five lumbar vertebrae and four intervertebral discs into nine classes. The authors evaluate on a public multi-scanner, multi-modal dataset (12 scanners, 34 subjects) using subject-preserved 10-fold cross-validation, reporting a mean IoU of 83.66%, sensitivity of 91.44%, and DSC of 91.03% for the segmentation stage, along with a YOLOv8 ROI detection mAP of 0.916. The paper claims the system is 'machine-agnostic', generalizing across scanners and addressing the challenge of diverse MRI modalities.

Significance. If the reported performance holds under a stricter evaluation, the cascaded design and the use of Self-ONN for spine segmentation could be a useful engineering contribution. The use of a public multi-scanner dataset, subject-preserved cross-validation, and comparisons with several decoder architectures are commendable. However, the central claims are not yet substantiated: the segmentation model is trained only on T1-weighted scans, and the cross-validation does not prevent scanner overlap between training and test folds. Therefore, the 'machine-agnostic' and multi-modal generalization claims are not supported by the evidence presented. The paper also contains a contradictory novelty claim and several technical errors that need correction.

major comments (5)

[Section 3.2, Table 2] The segmentation model is trained and evaluated exclusively on T1-weighted images (T1 non-contrast, T1 contrast-enhanced, T1 fat-saturated), while the abstract and introduction claim that the method addresses 'diverse MRI modalities' and is 'machine-agnostic'. The paper explicitly states that T2, STIR, and Dixon scans were excluded because of limited sample size, and the model is never tested on those modalities. This is a load-bearing overstatement: the claim should be restricted to T1-weighted multi-scanner segmentation, or the authors must provide evidence of performance on the other modalities.
[Section 4.2, Table 3] The 10-fold cross-validation is subject-preserved, as stated in Section 4.2, but it is not scanner-disjoint. Because multiple subjects in the dataset are scanned on the same scanner (Table 1 reports 34 subjects and 12 scanners), the same scanner's intensity characteristics can appear in both training and test folds. This allows the model to exploit scanner-specific appearance rather than learn general anatomy, potentially inflating the reported DSC and IoU. To support the 'machine-agnostic' claim, the authors should run a scanner-disjoint cross-validation or, at minimum, report per-scanner performance and the degree of scanner overlap across folds.
[Section 4.2, Table 3] The ROI detector is trained on only 2 subjects (9 scans, 160 slices) and validated/tested on one subject each. With such a small training set, the reported mAP of 0.916 and the downstream segmentation performance are not convincing evidence of robust localization across 12 scanners. The authors should either increase the ROI training set (if possible) or explicitly discuss the statistical limitations of this split.
[Section 1 and Section 4.6, Table 7] The claim that this is 'the first of its kind' in segmenting individual lumbar vertebral bodies and intervertebral discs is contradicted by the same paper's Table 7, which lists Yilizati-Yilihamu et al. (2023) as segmenting 10 vertebrae and 9 IVDs individually. The novelty statement should be revised to a precise and defensible claim, e.g., regarding the cascaded Self-ONN approach on this dataset, rather than being the first to segment individual lumbar structures.
[Section 3.2 and Section 4.1] The preprocessing removes 20% of slices from the beginning and end of each scan (heuristically chosen) and excludes 53 corrupted scans (about 25% of the dataset). Both decisions may preferentially remove hard or ambiguous cases and thus bias the reported metrics upward. No analysis is provided to assess this bias. The authors should either quantify the impact of these exclusions (e.g., by reporting metrics with and without them, if feasible) or explicitly acknowledge this limitation in the discussion.

minor comments (6)

[Equation (9)] The accuracy formula is incorrect: it should be (TP+TN)/(TP+FP+TN+FN), not (TP+FP)/total. Please correct the definition.
[Equation (14)] The Dice similarity coefficient formula has an error in the denominator: it should be |X|+|Y| (the sum of the two set sizes), not their union. Please fix the equation.
[Section 4.4] The experimental setup paragraph contains an out-of-place sentence about 'a sizable dataset of pediatric respiratory sounds', which appears to be a copy-paste error from another manuscript. This should be removed or corrected.
[Keywords and Title] The term 'Lumber vertebrae' is a typo; it should be 'Lumbar vertebrae'.
[Section 3.5] To enable reproduction, please provide full training details: loss function (e.g., Dice loss, cross-entropy), optimizer, learning rate schedule, batch size, number of epochs, input resolution, and augmentation parameters.
[Section 4.6] The comparison with prior work would be strengthened by reporting confidence intervals or per-class performance, especially because the dataset is small and folds vary in size.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the segmentation metrics are measured on held-out subject-disjoint folds of an external dataset, with no self-citation chain reducing the central claim to its own inputs.

full rationale

This is an empirical performance paper; the central numbers (mIoU 83.66%, DSC 91.03%, sensitivity 91.44%) are measured on held-out subject-disjoint folds of the external multi-scanner dataset, not derived from fitted parameters or from equations that assume the result. The Self-ONN component is cited from Kiranyaz et al. (2021), an independent prior work with no author overlap. The 'machine-agnostic' claim is indeed weakened by the absence of a scanner-disjoint split and by restriction of the segmentation training to T1-weighted scans, and the copy-paste error in Section 4.4 about pediatric respiratory sounds indicates quality-control issues; however, these are external-validity and correctness limitations, not circular reasoning. No load-bearing premise reduces to its own conclusion, no fitted input is relabeled as a prediction, and no uniqueness theorem is imported from the authors' own prior work. Therefore the paper has no significant circularity.

Assumptions & free parameters 4 free parameters · 4 assumptions · 0 invented entities

The central result is an empirical model performance measurement, not a derivation, so the ledger lists hand-tuned hyperparameters and dataset construction choices that the numbers depend on. No new physical entities or formal axioms are introduced.

free parameters (4)

Self-ONN Taylor order Q = Not reported
The generative neuron uses a polynomial of order Q (Eq. 3); Q controls the decoder's representational capacity and is never specified, yet it directly affects the segmentation result.
Slice removal fraction = 20%
Section 3.2: 'we removed 20% of slices from the beginning and end' and 'the value of 20% was chosen by heuristically investigating the dataset.' This data selection shapes every reported segmentation metric.
Learning rate = 0.001
Section 4.7 ablation selects 0.001 as optimal; learning rates of 0.01 and 0.1 substantially reduce Dice, so the headline number is conditional on this tuning.
Number of decoder blocks = 5
Section 4.7 ablation: 5 decoder blocks yields 84.2% Dice versus 80% for 3 and 82.3% for 7; the reported 91.03% assumes this architecture choice.

assumptions (4)

domain assumption Manual annotations in the source dataset are treated as ground truth after STAPLE fusion.
Section 3.1 says masks were unified with STAPLE; any rater error or STAPLE bias propagates to all reported metrics.
standard math Subject-preserved 10-fold cross-validation gives an unbiased estimate of generalization.
Section 4.2 states no subject appears in different folds; this is standard practice, but with only 34 subjects the variance of the estimate is high.
domain assumption The YOLOv8 ROI detector trained on 2 subjects generalizes to unseen patients and all modalities.
Section 4.2 trains ROI extraction on just 2 subjects (9 scans) and the segmentation pipeline depends on this detector; the ceiling analysis shows ROI errors cost about 2.7 Dice points.
ad hoc to paper Excluding 20% of slices and 53 corrupted scans does not selectively remove hard cases.
Sections 3.2 and 4.1 describe these exclusions as heuristic; the paper provides no analysis showing the excluded slices are randomly distributed across difficulty.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Machine-agnostic Automated Lumbar MRI Segmentation using a Cascaded Model Based on Generative Neurons." pith.science (2026). https://pith.science/paper/C7WXDFM2

@misc{pith2026241115656,
  author       = {Pith},
  title        = {Pith review of: Machine-agnostic Automated Lumbar MRI Segmentation using a Cascaded Model Based on Generative Neurons},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/C7WXDFM2}},
  note         = {Machine review of arXiv:2411.15656}
}

read the original abstract

Automated lumbar spine segmentation is very crucial for modern diagnosis systems. In this study, we introduce a novel machine-agnostic approach for segmenting lumbar vertebrae and intervertebral discs from MRI images, employing a cascaded model that synergizes an ROI detection and a Self-organized Operational Neural Network (Self-ONN)-based encoder-decoder network for segmentation. Addressing the challenge of diverse MRI modalities, our methodology capitalizes on a unique dataset comprising images from 12 scanners and 34 subjects, enhanced through strategic preprocessing and data augmentation techniques. The YOLOv8 medium model excels in ROI extraction, achieving an excellent performance of 0.916 mAP score. Significantly, our Self-ONN-based model, combined with a DenseNet121 encoder, demonstrates excellent performance in lumbar vertebrae and IVD segmentation with a mean Intersection over Union (IoU) of 83.66%, a sensitivity of 91.44%, and Dice Similarity Coefficient (DSC) of 91.03%, as validated through rigorous 10-fold cross-validation. This study not only showcases an effective approach to MRI segmentation in spine-related disorders but also sets the stage for future advancements in automated diagnostic tools, emphasizing the need for further dataset expansion and model refinement for broader clinical applicability.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

12 extracted references · 11 canonical work pages

[1]

vert ebrae

Introduction The spine is regarded as one of the most critical parts of the body , as it is responsible for upright posture, movement, and structural support. Besides its mechanical tasks, the spine protects the spinal cord, a neurological conduit that connects the brain and body. This complicated network of nerves allows voluntary control of physical mot...

work page 2005
[2]

Due to the low contrast of MRI images, the boundary between the spine and surrounding tissues is usually indistinct

work page
[3]

The shape of the vertebrae or IVDs varies greatly and may fluctuate dramatically across patients

work page
[4]

Variation among slices of the same MRI scan is too significant to create a single generalized approach for all distinct slices

work page
[5]

Because of technology differences, inter-scanner variability may arise

work page
[6]

Finally, the presence of foreign bodies, irrelevant organs, and noise adds another layer of challenge to this task

Intra-scanner variability can occur due to mobility, physiological state, medication, time of day, and other factors (Wittens, et al., 2021). Finally, the presence of foreign bodies, irrelevant organs, and noise adds another layer of challenge to this task. To address the problems of spinal MRI interpretation, a range of computer-aided diagnostic strategi...

work page 2021
[7]

Nagel et al.(Naegel, 2007) proposed a mathematical morphology-based method for anatomically labeling vertebrae from 3D CT-scan images

Related Works Researchers have been working on lumber vertebrae and intervertebral disc segments for quite some time. Nagel et al.(Naegel, 2007) proposed a mathematical morphology-based method for anatomically labeling vertebrae from 3D CT-scan images. Ghosh et al.(Ghosh & Chaudhary, 2014) used a two-stage algorithm to detect the spinal cord with the Houg...

work page 2007
[8]

Methodology 3.1 Overview of the framework One of the significant properties of the dataset is the annotation is done only on lumbar vertebrae and IVDs even though other parts of the spine, including the thoracic, sacrum, or coccyx, might be present in the scan. To address this issue, we propose a cascaded approach, including a region of interest (ROI) ext...

work page 2023

Show all 12 references

[10]

Multi -scanner and multi -modal lumbar vertebral body and intervertebral disc segmentation database

Result 4.1 Dataset description The "Multi -scanner and multi -modal lumbar vertebral body and intervertebral disc segmentation database" (Khalil, et al., 2022) stands out as one of the few publicly available datasets that integrates data from a variety of machines and modaliti...

2022
[12]

Conclusion In this study, we have taken a machine -agnostic approach to automatically segment lumbar vertebrae and intervertebral discs from MR images using our proposed model. Traditional segmentation methods are often found to underperform in this task because of the extreme...

2017 arXiv
[2004]

After modality selection and slice removal process, the number of samples became too small

to improve the equalization of the histogram of the MR images. After modality selection and slice removal process, the number of samples became too small. So, we opted for data augmentation in both stages of the pipeline. However, the augmentation types were different dependin...

2016
[2018]

Zhang, et al., 2020) Not mentioned 8: L1-L5, S1, T11, T12 (only vertebrae) Adversarial Network, LSTM 95% 2020 (J

Multiple 6: lumber IVD, other IVD, lumbar vertebrae, other vertebrae, NF, NFS GAN DSC: 87.1% 86% Specificity: 89.1% 2020 (R. Zhang, et al., 2020) Not mentioned 8: L1-L5, S1, T11, T12 (only vertebrae) Adversarial Network, LSTM 95% 2020 (J. Zhou, et al., 2020) 1 1: vertebrae U-n...

2020

Pith tools

Reviewed August 12, 2026 · model on record in the stance chip above.

[1] [1]

vert ebrae

Introduction The spine is regarded as one of the most critical parts of the body , as it is responsible for upright posture, movement, and structural support. Besides its mechanical tasks, the spine protects the spinal cord, a neurological conduit that connects the brain and body. This complicated network of nerves allows voluntary control of physical mot...

work page 2005

[2] [2]

Due to the low contrast of MRI images, the boundary between the spine and surrounding tissues is usually indistinct

work page

[3] [3]

The shape of the vertebrae or IVDs varies greatly and may fluctuate dramatically across patients

work page

[4] [4]

Variation among slices of the same MRI scan is too significant to create a single generalized approach for all distinct slices

work page

[5] [5]

Because of technology differences, inter-scanner variability may arise

work page

[6] [6]

Finally, the presence of foreign bodies, irrelevant organs, and noise adds another layer of challenge to this task

Intra-scanner variability can occur due to mobility, physiological state, medication, time of day, and other factors (Wittens, et al., 2021). Finally, the presence of foreign bodies, irrelevant organs, and noise adds another layer of challenge to this task. To address the problems of spinal MRI interpretation, a range of computer-aided diagnostic strategi...

work page 2021

[7] [7]

Nagel et al.(Naegel, 2007) proposed a mathematical morphology-based method for anatomically labeling vertebrae from 3D CT-scan images

Related Works Researchers have been working on lumber vertebrae and intervertebral disc segments for quite some time. Nagel et al.(Naegel, 2007) proposed a mathematical morphology-based method for anatomically labeling vertebrae from 3D CT-scan images. Ghosh et al.(Ghosh & Chaudhary, 2014) used a two-stage algorithm to detect the spinal cord with the Houg...

work page 2007

[8] [8]

Methodology 3.1 Overview of the framework One of the significant properties of the dataset is the annotation is done only on lumbar vertebrae and IVDs even though other parts of the spine, including the thoracic, sacrum, or coccyx, might be present in the scan. To address this issue, we propose a cascaded approach, including a region of interest (ROI) ext...

work page 2023