Relevance analysis of MRI sequences for automatic liver tumor segmentation

Andrea Schenk; Grzegorz Chlebus; Hans Meine; Nasreddin Abolmaali

Relevance analysis of MRI sequences for automatic liver tumor segmentation

T0 review · 1 major / 0 minor · reviewed 2026-05-24 · grok-4.3

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

Pith's one-line read Layer-wise relevance propagation attributes segmentation decisions back to individual MRI input sequences for liver tumor models.

desk verdict Applies LRP to rank MRI sequences for liver tumor segmentation but provides no checks on whether the attributions are faithful. read the letter →

arxiv 1907.11773 v1 pith:5TMCMGHH submitted 2019-07-26 eess.IV

Grzegorz Chlebus , Nasreddin Abolmaali , Andrea Schenk , Hans Meine This is my paper

classification eess.IV

keywords MRIsequenceslivertumorsegmentationlayer-wiserelevancepropagationexplainableAIsemanticdeepneuralnetworksanalysismedicalimaging

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

The paper proposes applying layer-wise relevance propagation to semantic segmentation networks in order to explain their decisions in terms of the input data. It demonstrates this on the task of liver tumor segmentation from multi-sequence MRI, with the goal of determining the contribution of each sequence to the final output. A reader would care because the method offers a way to validate whether the network relies on clinically sensible inputs and potentially to simplify acquisition protocols. The approach works by propagating relevance scores backward through the network layers to the input channels without requiring changes to the trained model.

What carries the argument

Layer-wise relevance propagation applied channel-wise to the multi-sequence MRI input of a segmentation network to produce relevance maps and aggregate scores per sequence.

What would settle it

An ablation study in which sequences ranked highest by relevance propagation are removed and the segmentation Dice score drops no more than when low-ranked sequences are removed would falsify the attribution faithfulness.

Watch

Extended reading notes

Core claim

The central claim is that layer-wise relevance propagation can be extended from classification to semantic segmentation networks and used to compute per-sequence relevance scores that indicate which MRI sequences drive the automatic liver tumor segmentation output.

Load-bearing premise

Layer-wise relevance propagation produces faithful and unbiased attributions from the segmentation output to the individual MRI input sequences.

Editorial extensions

If this is right

MRI sequences can be ranked by their contribution to liver tumor segmentation accuracy.
Low-relevance sequences may be omitted from future scans or model inputs with limited impact on performance.
Model validation becomes possible by checking whether high-relevance sequences align with radiological expectations.
The same relevance scores can guide targeted improvements to the network architecture or training data.

Reading between the lines

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

The method could be applied to other multi-modal medical segmentation tasks to optimize the number of required imaging sequences.
Clinical workflows might use the rankings to shorten MRI protocols while preserving diagnostic utility for tumor segmentation.
If relevance attributions prove stable across different network architectures, they could serve as a general tool for comparing input modalities.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Desk Editor's Note

read the letter

The punchline is that this paper runs standard layer-wise relevance propagation on a U-Net trained for liver tumor segmentation from four MRI sequences and reports which sequences get the highest relevance scores. It is an application note, not a new method or framework. The work does a clean job of adapting LRP rules to multi-channel input and framing the output as a practical check on what the model is using. That kind of concrete example can help people who already train segmentation networks and want a quick way to inspect input channels. The citation pattern is standard and appropriate for LRP work. The central claim, however, rests on an untested assumption that the relevance scores actually reflect what drives the predictions. There are no perturbation tests, no ablation that removes high-relevance sequences and measures the performance drop, and no comparison against other attribution techniques. Without those steps the rankings remain unvalidated numbers. The evaluation section appears to stop once the scores are produced. This paper is mainly for readers already working on explainability in medical image segmentation who need a short case study. It does not advance the underlying technique or supply the evidence needed to trust the rankings for model validation. I would not bring it to a reading group and would not cite it. It does not deserve peer review in its current form because the load-bearing claim about relevance lacks supporting experiments.

Referee Report

1 major / 0 minor

Summary. The manuscript proposes an approach to explain semantic segmentation networks by means of layer-wise relevance propagation (LRP). As an exemplary application, it investigates which MRI sequences are most relevant for liver tumor segmentation.

Significance. If LRP attributions prove faithful in this multi-sequence setting, the work could contribute to interpretability of medical segmentation models and inform sequence selection. However, the absence of any quantitative checks on attribution quality limits the potential impact, as the central claim rests on unverified faithfulness of the relevance scores.

major comments (1)

The central claim requires that LRP attributions from the segmentation output back to the four input MRI sequences are faithful. The manuscript applies standard LRP rules to a U-Net-like architecture but provides no ablation, perturbation, or comparison against other attribution methods to test whether the resulting per-sequence relevance scores are unbiased or predictive of actual model behavior. This is load-bearing for the exemplary application.

Simulated Author's Rebuttal

1 responses · 0 unresolved

We thank the referee for their review. We address the major comment regarding validation of LRP attributions below.

read point-by-point responses

Referee: The central claim requires that LRP attributions from the segmentation output back to the four input MRI sequences are faithful. The manuscript applies standard LRP rules to a U-Net-like architecture but provides no ablation, perturbation, or comparison against other attribution methods to test whether the resulting per-sequence relevance scores are unbiased or predictive of actual model behavior. This is load-bearing for the exemplary application.

Authors: We recognize the referee's point but note that the manuscript presents LRP as a tool for an exemplary analysis of sequence relevance rather than a validation study of attribution methods. Standard LRP rules are applied to a U-Net-like model following established implementations; the work does not assert that the resulting scores have been quantitatively verified as faithful or unbiased. The central contribution is the demonstration of the approach on multi-sequence MRI data and the resulting per-sequence relevance insights. We maintain that additional ablations or comparisons fall outside the stated scope. revision: no

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: standard LRP application to segmentation network

full rationale

The paper applies layer-wise relevance propagation (a pre-existing method) to attribute relevance from a U-Net segmentation output back to input MRI sequences. No equations or derivations are presented that reduce by construction to fitted inputs, self-definitions, or self-citation chains. The central claim (which sequences are most relevant) is an empirical output of the attribution process rather than a tautological renaming or forced prediction. The method is self-contained against external LRP literature and does not rely on load-bearing self-citations for its validity.

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

Based solely on the abstract, no free parameters, axioms, or invented entities can be identified.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Relevance analysis of MRI sequences for automatic liver tumor segmentation." pith.science (2026). https://pith.science/paper/5TMCMGHH

@misc{pith2026190711773,
  author       = {Pith},
  title        = {Pith review of: Relevance analysis of MRI sequences for automatic liver tumor segmentation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/5TMCMGHH}},
  note         = {Machine review of arXiv:1907.11773}
}

read the original abstract

Explainability of decisions made by deep neural networks is of high value as it allows for validation and improvement of models. This work proposes an approach to explain semantic segmentation networks by means of layer-wise relevance propagation. As an exemplary application, we investigate which MRI sequences are most relevant for liver tumor segmentation.

Discussion (0). Sign in to comment.

Pith tools

Reviewed May 24, 2026 · model on record in the stance chip above.