{"id":"867aee2a-1659-4a95-9e44-951d5520f4e6","arxiv_id":"2506.20306","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A deep learning-guided, per-patient selection of radiomic features for knee MRI matches deep learning accuracy while preserving interpretable logistic-regression decisions.","lead":"This paper proposes a per-patient 'radiomic fingerprint' where a deep learning model selects a personalized set of handcrafted MRI features for each knee scan, and a simple logistic classifier makes the diagnosis. The approach rivals end-to-end deep learning accuracy on three knee-injury tasks while keeping the decision features visible to clinicians.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Underspecified pairwise-interaction logistic regression over a ~2,640-feature radiomic pool; without regularization details, reported accuracy and interpretability may be overfitting artifacts.","rationale":"I focused on the weakest structural link: the under-specified classifier. If the classifier is actually high-dimensional and unregularized, both the accuracy comparison to MRNet/ELNet and the interpretability argument fail. This is more fundamental than the AUC gap, because it affects the validity of all reported metrics. The reader's weakest-assumption statement captures exactly this issue; my independent calculation of the feature-pool size reinforces it. The paper's own Sec. 2 vs Sec. 3.1 wording is internally inconsistent about the degree of the classifier. Since code is publicly available, the concern is directly testable; if resolved favorably, the conditional acceptance can proceed. Therefore I recommend keeping the reader's CONDITIONAL verdict unchanged.","tokens_in":152,"tokens_out":3887,"duration_ms":47479,"concrete_test":"Open the public GitHub repo, locate the logistic-regression layer, and count its trainable parameters; check whether pairwise interaction features are explicitly constructed and whether any L1/L2 regularization is applied. Then retrain the ACL model with L1-penalized logistic regression on first-order selected features only and evaluate on the same 120-case test set. If the simplified model matches the reported 0.94 accuracy / 0.92 AUC within confidence intervals, the pairwise-interaction term is not needed; if the unregularized high-dimensional model was used, the claim of a low-dimensional interpretable classifier is unsupported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim of comparable/superior accuracy and per-patient interpretability rests on the 'extended logistic regression' described in Sec. 3.1. The paper never states the number of features (K), patch count (J), or whether pairwise interactions are formed over all soft-weighted features during training. From Sec. 3.1, K = 19+16+24+16+16+5+14 = 110; with J = 8 and 3 views, the pool has 3*8*110 = 2,640 features. Unrestricted second-order interactions yield ~3.5 million coefficients, versus 1,130 training samples. No regularization, optimizer, or interaction-selection scheme is described, so the model could be severely overparameterized. This directly undermines both the 'low-dimensional' interpretability claim and the validity of test-set accuracy, especially since only 120 test cases are used and threshold T is tuned on validation. Sec. 2 even states the classifier has 'no higher-degree terms,' contradicting the second-order interactions in Sec. 3.1. If the released code confirms an unregularized large interaction model, the reported results are not a reliable demonstration of the method's stated advantages.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes a patient-specific radiomic fingerprint framework for knee MRI diagnosis. A 3D ResNet takes a registered multi-view ROI as input and predicts relevance weights for a large pool of hand-crafted 3D radiomic features extracted from local patches; the weighted features are thresholded into a sparse per-patient fingerprint and passed to an extended logistic regression classifier with second-order interactions. The method is validated on the MRNet dataset for three tasks (general abnormality, ACL tear, meniscus tear), with reported accuracy/AUC that the authors claim is comparable or superior to end-to-end deep learning models. The paper also reports ablations on feature composition, feature selection, patch configuration, and registration, plus qualitative case analyses intended to show clinical interpretability.","tokens_in":8303,"tokens_out":4610,"duration_ms":48839,"significance":"If the quantitative claims hold, the paper would demonstrate a useful middle ground between rigid, interpretable radiomics and accurate but opaque end-to-end deep learning, with the added benefit of per-patient feature-level explanations. The manuscript has concrete strengths: it uses a public dataset (MRNet), provides a code repository, defines a joint training objective for feature weighting and classification, and includes an ablation study of core components. The clinical case analysis provides a plausible narrative for how selected radiomic features correspond to known ACL pathology. However, the central quantitative claims are currently difficult to verify because the classifier is underspecified, a conventional fixed-radiomic-signature baseline is absent, and the reported statistical comparisons do not cover all baselines. These issues are load-bearing for the abstract's 'comparable or superior' claim and for the interpretability-vs-performance trade-off the paper emphasizes.","major_comments":[{"comment":"The extended logistic regression with second-order pairwise interactions is underspecified in a way that directly affects the paper's low-dimensionality and generalization claims. With K=110 (the sum of the listed feature families), J=8 patches, and three views, the feature pool has 3JK=2,640 entries; unrestricted pairwise interactions would produce roughly 3.5 million coefficients, compared with 1,130 training samples. The manuscript does not state whether interactions are formed over the full soft-weighted pool, the thresholded binary selection, or a restricted subset, nor does it give any regularization, optimizer, or interaction-selection scheme. The claim that this classifier is 'low-dimensional' is therefore unsupported, and the reported test accuracy could reflect overfitting. This also conflicts with Sec. 2, where the classifier is described as having 'no higher-degree terms.' Please specify the exact number of trainable parameters, the regularization/selection procedure for interactions, and how the model complexity is controlled relative to the training-set size.","section":"Sec. 3.1, Classification Model"},{"comment":"The abstract and Sec. 3.2 claim 'comparable or superior diagnostic accuracy' relative to state-of-the-art end-to-end DL models, but the reported AUC is lower than MRNet for general abnormality (0.85±0.17 vs 0.94±0.02) and ACL tear (0.92±0.04 vs 0.97±0.02). The only significance tests reported are against ELNet (p values in Sec. 3.2), with no statistical test against MRNet or SKID. Because accuracy, sensitivity, and specificity depend on the threshold T, which is tuned on the validation set, accuracy-based comparisons are not a satisfactory substitute for AUC comparisons. Please report AUC differences with confidence intervals or a statistical test against all baselines, and state the primary metric used for model selection.","section":"Tables 1 and 2, Quantitative comparison"},{"comment":"The paper's motivation is that fixed population-level radiomic signatures are too constrained and underperform, but no conventional radiomic-signature baseline is evaluated; all comparators are end-to-end DL models. The NoFS ablation is not such a baseline, and its results do not support the text's claim that removing the feature-weighting network leads to 'a noticeable decline in performance.' In Table 2, NoFS has higher AUC than Ours for all three tasks (0.89 vs 0.85 for abnormalities, 0.94 vs 0.92 for ACL, 0.85 vs 0.82 for meniscus), while accuracy is comparable (0.88 vs 0.92, 0.92 vs 0.94, 0.83 vs 0.84). Please include a standard radiomic-signature baseline (e.g., features selected on the training set only plus logistic regression, with no per-patient weighting) and reconcile the NoFS numbers with the claimed benefit of adaptive selection.","section":"Sec. 3.1 / Table 2, Baseline and NoFS ablation"},{"comment":"The manuscript reports means and standard deviations but does not state the number of training/validation splits, random seeds, or how the 1,130 development cases are divided. The 120-case test set is small for three binary tasks, and several metrics have large standard deviations (e.g., specificity 0.71±0.33 for abnormalities and 0.67±0.34 for meniscus). Please specify the cross-validation or repeated-split protocol, the number of runs, and the exact use of the validation set for threshold T and any hyperparameter selection, so that the reported error bars are interpretable.","section":"Sec. 3.1, Evaluation protocol"}],"minor_comments":[{"comment":"There are typos such as 'brievity' and 'intepretability'; please proofread the manuscript.","section":"Sec. 2"},{"comment":"The column headers in Table 1 and Table 2 are misaligned; the metric values for 'Ours' appear under the 'Reg Type' columns, making the tables difficult to read. Please restructure them so that Acc, Sen, Spe, and AUC are clearly associated with each method.","section":"Tables 1 and 2"},{"comment":"The dimension K of the radiomic feature set is not stated explicitly; the reader has to sum the listed feature counts to obtain K=110. Please state K directly, as the feature-pool dimension 3JK is used throughout the method description.","section":"Sec. 3.1"},{"comment":"The ROI is defined as 50% depth, 30% height, and 50% width of the registered volume; please specify whether these percentages are relative to the resized 32×128×128 volume and provide the anatomical landmarks used to define the subvolume, since this affects reproducibility.","section":"Sec. 3.1, ROI definition"}],"recommendation":"major_revision","confidential_remarks":"For the editor: the manuscript's self-citation to prior work (ref. [5]) is appropriate and not circular. The main risk is that the released code, once inspected, may confirm an unregularized large interaction model; in that case the central claims would need substantial revision. The paper fits the scope of the journal as a medical-imaging methods paper, but the quantitative evidence currently does not fully support the abstract's 'comparable or superior' claim without added baselines and statistical testing."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: the paper has a real idea—per-patient dynamic radiomic selection via a DL weighting network—and the ablations suggest the selection matters. But the central accuracy claim is weaker than it appears, and the classifier that does the heavy lifting is underspecified to the point of not being reproducible as written.\n\nWhat's genuinely new: instead of one global radiomic signature, they learn a relevance vector per case, threshold it, and feed the sparse fingerprint to a logistic classifier. That is a legitimate extension of their prior healthy-persona work and of attention-style weighting, and the knee MRI experiments are new. The feature analysis (sagittal view dominance, texture features mattering more in non-tear cases) is the kind of interpretability you want to see. Code is released, the ablations cover feature composition, patch size, feature selection, and registration, and the ablation evidence that feature selection helps is convincing.\n\nSoft spots, in proportion. First, the classifier: Sec. 3.1 says 'extended logistic regression' with second-order pairwise interactions over a pool I make as 3 views × 8 patches × 110 features = 2,640 features. That's millions of interaction terms without any stated regularization, optimizer, or interaction-selection scheme, and it's trained on 1,130 samples. Sec. 2 even says the classifier has 'no higher-degree terms.' That contradiction is not cosmetic—it directly bears on whether the reported test accuracy and the interpretability claims are artifacts of overfitting. Second, the headline claim. Table 1 shows your AUC is numerically below MRNet on all three tasks (e.g., 0.85 vs 0.94 for general abnormalities), and the paper's 'comparable or superior' rests on accuracy, which is threshold-dependent with the threshold tuned on validation. No significance tests against MRNet are given. Third, the motivating claim that fixed radiomic signatures are limited is never tested—there's no fixed-signature baseline, and the NoFS ablation is not equivalent to one. These are fixable: report the regularization and interaction construction, add the baseline, run proper significance tests on AUC, and tone down the abstract.\n\nWho's this for? Someone working on interpretable medical imaging, particularly knee MRI. It deserves a serious referee—the idea is sound enough to warrant careful scrutiny, and the issues are addressable in revision. I would not desk-reject it.","headline":"A real idea—per-patient dynamic radiomic selection—with convincing ablations, but the underspecified pairwise logistic regression and the threshold-dependent accuracy claims need tightening before the results can be trusted.","tokens_in":8798,"tokens_out":2146,"would_cite":false,"duration_ms":21936,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper claims that a patient-specific radiomic fingerprint—a sparse set of hand-crafted MRI features selected per patient by a neural network and read out by a logistic regression—can match or beat end-to-end deep learning on knee MRI…","keywords":["knee MRI","radiomics","radiomic fingerprint","patient-specific feature selection","clinical interpretability","deep learning","ACL tear","meniscus tear"],"falsifier":"Retrain the same feature-weighting network with the classifier restricted to first-order terms only, and again with all pairwise interactions but with explicit L1 regularization, on the same train/test split, and compare test accuracy plus the overlap of selected features across random seeds. If the interaction-based classifier's accuracy advantage disappears, or if the selected fingerprints are unstable across retraining runs, the claim that per-patient fingerprints are both accurate and interpretable would be falsified.","tokens_in":7852,"feed_emoji":"🦵","tokens_out":13034,"duration_ms":124904,"temperature":0.7,"pith_summary":"The paper argues that the weakness of classical radiomics for knee MRI is not the use of hand-crafted features but the use of one fixed feature set for every patient. It proposes a radiomic fingerprint framework in which a neural network examines each MRI and assigns an importance score to each feature in a large hand-crafted radiomic pool, producing a sparse, patient-specific set of features that is then classified by a low-dimensional logistic regression. On a public knee MRI benchmark of 1,370 scans, the authors report diagnostic accuracy and area under the curve comparable to or better than established end-to-end deep learning models for general abnormalities, ACL tears, and meniscus tears. If correct, this means interpretable radiomics can retain a transparent decision process without giving up deep-learning-level accuracy, and the selected features may serve as candidate clinical biomarkers.","feed_headline":"Radiomic fingerprints match or beat deep learning on knee MRI","feed_subtitle":"A network picks a sparse patient-specific feature set; logistic regression then makes a transparent knee diagnosis.","key_machinery":"The central mechanism is the radiomic fingerprint: a patient-specific sparse selection from a large pool of hand-crafted radiomic features, produced by a neural network's predicted relevance scores. A 3D residual network takes the full knee ROI and outputs a weight $q_i \\in [0,1]$ for each of the $3JK$ features computed over $J$ non-overlapping patches in axial, coronal, and sagittal views. The weighted vector $f \\odot q$ is passed to an extended logistic regression that includes first-order terms and second-order pairwise interactions, and both models are trained jointly by minimizing cross-entropy loss; at inference the weights are thresholded at $T$ to form a binary selection mask, yielding the sparse fingerprint $f_s = f \\odot q^{(b)}$. The central property is that selection happens per patient, so the effective radiomic space is far larger than any fixed signature, while the final classifier remains a transparent low-dimensional model.","core_discovery":"On its own terms, the paper's central claim is that a per-patient radiomic fingerprint—a sparse subset of a large hand-crafted radiomic feature pool, chosen by a deep network for each individual case—can provide both the accuracy of end-to-end deep learning and the interpretability of classical radiomics. The framework extracts features from multiple patches and three MRI views, predicts a weight for every feature with a 3D residual network, weights the features by those predictions, and trains the network jointly with an extended logistic regression classifier, which includes pairwise interactions, via cross-entropy loss. At inference the continuous weights are thresholded into a binary mask, and the resulting sparse fingerprint is passed to the same logistic regression for the final diagnosis. The authors validate this on three diagnostic tasks and report comparable or superior accuracy relative to the deep learning baselines, then analyze specific ACL cases in which the top-ranked features align with radiological findings such as ligament discontinuity, joint effusion, and bone marrow edema.","pith_inferences":["Beyond the paper: replacing the pairwise logistic regression with a strictly first-order classifier, or adding explicit sparsity regularization, would test whether the reported accuracy comes from the interaction terms or from the per-patient selection mechanism itself.","Beyond the paper: computing fingerprints from repeated or slightly perturbed scans of the same knee would test the stability of the selection, a property the paper does not assess but that any biomarker use would require.","Beyond the paper: correlating the top-ranked fingerprint features with independent radiological grading scores, rather than binary labels alone, could turn the observed case-study findings into a formal biomarker validation study.","Beyond the paper: applying the same selection-plus-transparent-classifier design to other anatomies or imaging modalities, or extending the selected pool to include deep features, would show whether the per-patient fingerprint idea generalizes beyond knee MRI."],"forward_implications":["A fixed population-level radiomic signature is not required for interpretability; selecting features per patient can match or exceed the accuracy of end-to-end deep learning on this task.","The jointly trained network and logistic regression make feature selection task-specific, so the chosen features are optimized for diagnostic accuracy rather than chosen independently of the classifier.","The selected features carry usable clinical meaning: in ACL tear cases top features concentrate near the ligament in the sagittal view, and features like Large Dependence Low Gray Level Emphasis track the extent of tissue disruption.","Ablations show that the full radiomic feature set, a 2×2×2 patch decomposition, per-patient feature selection, and image registration each contribute to the reported performance.","The framework produces a transparent final decision from a sparse patient-specific fingerprint, giving clinicians a feature-by-feature explanation of each prediction."],"supporting_citations":[{"why":"Supplies the knee MRI dataset used in all experiments and one of the end-to-end deep learning baselines.","marker":"[3]"},{"why":"Serves as another end-to-end deep learning baseline reproduced for comparison.","marker":"[12]"},{"why":"Serves as a self-supervised representation learning baseline in the comparison table.","marker":"[9]"},{"why":"Defines the hand-crafted radiomic feature types that make up the large feature pool.","marker":"[13]"},{"why":"Performs the affine registration aligning the three MRI views; the no-registration ablation shows this step matters for the reported results.","marker":"[10]"},{"why":"Supplies the residual network architecture used as the feature-weighting network.","marker":"[6]"},{"why":"The authors' earlier patient-specific radiomics study that this work builds on and contrasts with.","marker":"[5]"}],"fun_headline_variants":["Knee MRI diagnosis gets a personalized radiomics spin","Per-patient radiomic fingerprints rival deep learning on knee MRI","Sparse per-patient fingerprints match deep nets for knee MRI","Dynamic radiomic fingerprints bring interpretability to knee MRI","Personalized radiomics: accuracy meets transparency in knee MRI"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The method assumes that the extended logistic regression with pairwise interactions can be trained on 1,130 exams without overfitting, even though the paper does not state the number of interactions, any regularization, or the optimization details.","fun_headline_variants_meta":{"raw":{"variants":["Knee MRI diagnosis gets a personalized radiomics spin","Per-patient radiomic fingerprints rival deep learning on knee MRI","Sparse per-patient fingerprints match deep nets for knee MRI","Dynamic radiomic fingerprints bring interpretability to knee MRI","Personalized radiomics: accuracy meets transparency in knee MRI"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000611,"raw_usage":{"total_tokens":2886,"prompt_tokens":1031,"completion_tokens":1855,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":647,"completion_tokens_details":{"reasoning_tokens":1776}},"tokens_in":647,"tokens_out":1855,"duration_ms":13764,"temperature":1.0,"reasoning_tokens":1776,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T22:52:03.453422+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Retrain the same feature-weighting network with the classifier restricted to first-order terms only, and again with all pairwise interactions but with explicit L1 regularization, on the same train/test split, and compare test accuracy plus the overlap of selected features across random seeds. If the interaction-based classifier's accuracy advantage disappears, or if the selected fingerprints are unstable across retraining runs, the claim that per-patient fingerprints are both accurate and interpretable would be falsified.","supporting_citations":[{"cited_title":"PLoS medicine15(11), e1002699 (2018)","cited_arxiv_id":null,"evidence_quote":"Supplies the knee MRI dataset used in all experiments and one of the end-to-end deep learning baselines."},{"cited_title":"In: Medical Imaging with Deep Learning","cited_arxiv_id":null,"evidence_quote":"Serves as another end-to-end deep learning baseline reproduced for comparison."},{"cited_title":"IEEE Transactions on Arti- ficial Intelligence (2023)","cited_arxiv_id":null,"evidence_quote":"Serves as a self-supervised representation learning baseline in the comparison table."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Performs the affine registration aligning the three MRI views; the no-registration ablation shows this step matters for the reported results."},{"cited_title":"Patient-specific radiomic feature selection with reconstructed healthy persona of knee MR images","cited_arxiv_id":"2503.13131","evidence_quote":"The authors' earlier patient-specific radiomics study that this work builds on and contrasts with."}],"review_version":1}