REVIEW 3 major objections 4 minor
In-hoc Concept Representations to Regularise Deep Learning in Medical Imaging
T0 review · 3 major / 4 minor · reviewed 2026-08-15 · deepseek-v4-flash
Pith's one-line read A concept-direction regularizer steers medical-image CNNs toward clinically meaningful features, improving accuracy when the input distribution shifts.
desk verdict The idea—using CAVs as a training-time regularizer without concept labels—is plausible and worth checking, but the version I have is corrupted, so I can only judge the abstract and the claims remain unverified. read the letter →
The pith
A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.
The reading
What carries the argument
The load-bearing mechanism is the LCR: a direction in the network's latent feature space that represents a clinically relevant concept. A small auxiliary dataset provides positive and negative examples of each concept, and a Concept Activation Vector-style linear probe identifies those directions. The training objective adds a regularization term that encourages the model's activations to stay in the subspaces spanned by the concept directions. This is what lets the main model use concepts without ever seeing concept labels in its main training set.
What would settle it
Train LCRReg on diabetic retinopathy with a small auxiliary concept set drawn from a different camera or population than the main training set, then measure accuracy on the out-of-distribution test shift; if the regularized model is no better than the unregularized baseline, or worse, concept-direction transfer is the point of failure.
Extended reading notes
Core claim
The paper's central claim is that deep learning models in medical imaging can be made to rely on clinically meaningful features rather than spurious correlations by adding a regularization term computed from Latent Concept Representations (LCRs). LCRReg first extracts concept directions, for example Concept Activation Vectors, from a small auxiliary dataset with concept examples, then trains a CNN on the main dataset with no concept labels while guiding activations to lie in the latent subspaces associated with those concepts. The authors report that this improves robustness against injected spurious correlations on a toy dataset and improves performance on diabetic retinopathy binary classification under synthetic spurious perturbations and out-of-distribution generalization, compared with baselines including multitask learning, linear probing, and post-hoc concept-based models.
Load-bearing premise
The small auxiliary concept set has to produce directions that still point at the same clinically meaningful features in the main training data and in the shifted test data; if those directions do not transfer, the regularizer steers the model toward features that are irrelevant or misleading.
Editorial extensions
If this is right
- LCRReg should let medical-image models keep higher accuracy under input shifts, such as a new scanner or different population, than the same architecture trained with a plain classification loss.
- The main training set does not need dense concept labels; only a small auxiliary set with concept examples is required.
- The same architecture can be used unchanged because the regularizer is a single additional term, and the paper shows it works in multi-concept and multiclass settings.
- The method gives a way to compare against stronger baselines such as multitask learning, linear probing, and post-hoc concept-based models without adding heavy supervision.
- If the concept directions are well chosen, the resulting representations are semantically grounded rather than driven purely by spurious correlations.
Reading between the lines
- Inference: The same auxiliary concept set could be reused as an audit probe at deployment time, by projecting test activations onto the LCRs to check whether the model still relies on the intended concepts when the input shifts.
- Inference: The method's dependence on one small auxiliary set suggests a practical recipe: a shared concept set per imaging modality could inject domain knowledge into many downstream models without re-annotating every dataset.
- Inference: A natural next test is whether the same latent-direction steering transfers to vision transformer architectures, which have comparable latent spaces but were not evaluated here.
- Inference: Deliberately omitting a known confounder from the concept set should make LCRReg fail to protect against that confounder, providing a direct way to measure how much concept coverage the method needs.
Signed reviews
Editorial analysis
A structured set of objections, weighed in public.
Referee Report
Summary. The paper introduces LCRReg, a regularisation method for convolutional neural networks in medical imaging. The method first extracts Latent Concept Representations (LCRs), such as Concept Activation Vectors (CAVs), from a small auxiliary dataset that has concept labels, and then adds a regularisation term that encourages the main network to activate within latent subspaces associated with these concepts. The authors claim that LCRReg improves robustness to synthetic spurious correlations and to out-of-distribution (OOD) shifts on a controlled toy dataset and on a diabetic retinopathy binary classification task, in comparison with baselines such as multitask learning, linear probing, and post-hoc concept-based models, without requiring concept labels in the main training set. A code repository link is provided in the abstract.
Significance. If the empirical claims hold, LCRReg would be a practically useful addition to the robustness toolbox: it is architecture-agnostic, requires only a small concept-labelled auxiliary set, and is compared against several relevant baselines. The provision of a public code link is a reproducibility strength, and the design gives rise to a falsifiable prediction, namely that concept-aligned regularisation, rather than generic regularisation, is responsible for the reported gains. However, the central evidence is currently inaccessible: the supplied full text is severely corrupted, so no experimental protocol, numeric table, or equation can be checked. The abstract alone reports qualitative claims without confidence intervals or protocol details. The decisive transfer premise of the method is asserted but not demonstrated in the available material.
major comments (3)
- [Full text / Abstract] The full text supplied to the referee is undecodable mojibake; no section, equation, table, or experimental protocol can be checked. The abstract reports qualitative claims ("significantly improves robustness", "enhances performance") without numeric results, confidence intervals, dataset splits, or baseline specifications. Because the paper's central claim is empirical, the unavailability of the body text is a load-bearing gap: I cannot verify that the experiments support the stated conclusion.
- [Abstract, transfer premise] The method's success depends on the premise that LCRs estimated from a small auxiliary concept dataset remain stable and aligned with the corresponding features in the main training distribution and under the OOD shifts tested. The abstract asserts this premise through the phrase "high-quality, disentangled concept examples" but provides no supporting evidence, such as a measure of CAV stability across domains, a comparison with random concept directions, or a control with shuffled concept labels. Without such a control, the reported gains could be attributable to generic regularisation rather than to semantically grounded concept alignment.
- [Hyperparameters / experimental protocol] The abstract does not report the regularisation weight or the dimensionality of the concept subspaces, and the visible fragments of the manuscript do not reveal how these were selected. If these hyperparameters were tuned on the OOD test sets, the claimed robustness improvement would be weakened by selection on the test distribution. The authors should state the validation-based selection procedure and, preferably, provide a sensitivity analysis over the regularisation weight and subspace dimension.
minor comments (4)
- [Abstract / Notation] The acronym CAV is defined in the abstract, but LCR is not; please define the acronym at first use and state explicitly the form of the regularisation term (e.g., projection loss, cosine-similarity penalty, or CAV-alignment loss) in the main text.
- [Experimental description] Please state the size of the auxiliary concept dataset, the number and type of concepts used, and any measures taken to assess domain mismatch between the auxiliary dataset and the main training set; this information is essential for assessing the transfer claim.
- [Reproducibility] The code repository link should include a version or commit identifier and a license statement; this would make the reported experiments reproducible from the archival record.
- [Figures/Tables] The corrupted text includes unreadable table and figure fragments; the final PDF must be re-rendered so that tables show units, sample sizes, and confidence intervals or error bars, and so that figure axes and legends are legible.
Circularity Check
No circularity: LCRReg is an empirical regularization method whose concept vectors are external auxiliary inputs and whose robustness claims are evaluated on held-out tasks; no derivation reduces to its own inputs.
full rationale
The paper's central claim is empirical: regularization via latent concept representations (CAVs) improves robustness under synthetic spurious perturbations and OOD generalization. The concept directions are extracted from a small auxiliary dataset and are therefore inputs to the method, not quantities derived from the prediction target. The reported gains are compared against external baselines on held-out tasks, so the robustness numbers are not forced by construction. The abstract states this design openly: 'uses a small auxiliary dataset to synthesise high-quality, disentangled concept examples' and 'incorporate[s] a regularisation term that guides a CNN to activate within latent subspaces associated with those concepts.' There is no equation in the readable text where a predicted robustness value equals a fitted parameter, nor any load-bearing self-citation chain; the method is presented as a new regularization strategy built on the standard CAV notion. Concerns about auxiliary-domain transfer, hyperparameter selection, or shared features between synthetic perturbations and concept examples are empirical validity risks, not circularity, and no quoted passage exhibits such sharing. The supplied full text is heavily corrupted, so I could not extract additional derivations, but the available abstract and readable fragments do not support a circularity finding. Therefore the paper receives a score of 0.
Assumptions & free parameters
free parameters (2)
- Regularization weight for LCRReg term =
Unknown, not stated in abstract
- Concept set and subspace dimensionality =
Predefined clinically relevant features, unspecified
assumptions (3)
- domain assumption The auxiliary dataset can generate high-quality, disentangled concept examples whose CAV directions transfer to the main training distribution.
- domain assumption Aligning CNN activations with CAV subspaces improves robustness to spurious correlations and OOD shifts.
- domain assumption CAVs computed from the auxiliary set are stable and semantically meaningful for the predefined concepts.
Cite this review
Pith. "Pith review of In-hoc Concept Representations to Regularise Deep Learning in Medical Imaging." pith.science (2026). https://pith.science/paper/JSUXLPK6
@misc{pith2026250813880,
author = {Pith},
title = {Pith review of: In-hoc Concept Representations to Regularise Deep Learning in Medical Imaging},
year = {2026},
howpublished = {\url{https://pith.science/paper/JSUXLPK6}},
note = {Machine review of arXiv:2508.13880}
}
read the original abstract
Deep learning models in medical imaging often achieve strong in-distribution performance but struggle to generalise under distribution shifts, frequently relying on spurious correlations instead of clinically meaningful features. We introduce LCRReg, a novel regularisation approach that leverages Latent Concept Representations (LCRs) (e.g., Concept Activation Vectors (CAVs)) to guide models toward semantically grounded representations. LCRReg requires no concept labels in the main training set and instead uses a small auxiliary dataset to synthesise high-quality, disentangled concept examples. We extract LCRs for predefined relevant features, and incorporate a regularisation term that guides a Convolutional Neural Network (CNN) to activate within latent subspaces associated with those concepts. We evaluate LCRReg across synthetic and real-world medical tasks. On a controlled toy dataset, it significantly improves robustness to injected spurious correlations and remains effective even in multi-concept and multiclass settings. On the diabetic retinopathy binary classification task, LCRReg enhances performance under both synthetic spurious perturbations and out-of-distribution (OOD) generalisation. Compared to baselines, including multitask learning, linear probing, and post-hoc concept-based models, LCRReg offers a lightweight, architecture-agnostic strategy for improving model robustness without requiring dense concept supervision. Code is available at the following link: https://github.com/Trustworthy-AI-UU-NKI/lcr\_regularization
Reviewed August 15, 2026 · model on record in the stance chip above.
Discussion (0). Continue with ORCID to comment.