{"id":"eb83a0a0-8c9f-4ce7-a9a6-22521c029afa","arxiv_id":"2504.20033","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"high","formal_verification":"none","parameter_count":7,"one_line_summary":"A continual learning system combining triplet loss, VAE-generated pseudo-replay, and knowledge distillation is reported to reduce forgetting on medical image datasets, though key loss terms are described inconsistently.","lead":"This paper tests a continual learning method for medical images that avoids storing old data by generating synthetic copies with a VAE and using knowledge distillation to stop forgetting. It reports accuracy gains over older methods on prostate MRI, OCT, pathology, and CIFAR-10, but the method details contain inconsistencies that currently make the reported medical results hard to trust.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The stated covariance \"alignment\" (Eq. 7) is not an alignment term, and the OCT protocol's one-class tasks make the sole current-task loss (triplet loss) undefined, so Table 2 cannot be reproduced from the equations given.","rationale":"The central claim is that teacher-guided VAE replay plus feature-attention/covariance KD prevents forgetting. For that claim to hold, the losses in Section 3.2 must be implementable and must do what the text says. Eq. (7) fails on both counts: a matrix sum cannot be added to scalar losses in Eq. (1), and no cross-covariance alignment term appears. Even with the natural correction to c(Z_k)+c(Z_{k-1}), the loss is not teacher-student alignment. This is a correctness risk, not a stylistic issue.\n\nThe OCT setup is independently decisive: the text assigns one-class tasks to tasks 2 and 3, making Ltri, the only current-task loss, undefined for those tasks. Since the generator replay only concerns past classes, the model has no supervised signal for DME/Drusen, so Table 2's OCT number cannot be produced by Algorithm 1 as stated.\n\nThe reader's verdict of REJECT is therefore appropriate; I would keep it. Agreement is partial because the reader identified the Eq. (7) mismatch but did not flag the OCT triplet contradiction, which I find equally load-bearing.","tokens_in":11447,"tokens_out":6652,"duration_ms":69460,"concrete_test":"Re-run OCT tasks 2 and 3 with Algorithm 1 exactly as written: construct Ltri from DME-only and Drusen-only batches and report per-task accuracy for those classes; also print the shape and value of LCov from Eq. (7) for one batch. If Ltri cannot be formed or LCov is a matrix, the implementation must deviate from the paper, and Table 2 is not evidence for the described method.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 3.2.2 defines LCov in Eqs. (5)-(7). The text states that LCov implements \"covariance between the Teacher and the Student\" and Figure 2 describes alignment of covariance matrices; however Eq. (7), LCov = C(Z_k) + C(Z_{k-1}), is a sum of two d x d covariance matrices. As written, it is not a scalar loss, and it contains no term comparing C(Z_k) with C(Z_{k-1}). If the intended scalar is c(Z_k)+c(Z_{k-1}) from Eq. (6), the loss only drives off-diagonal entries of each model's covariance toward zero independently; it does not transfer the teacher's correlation structure. The claimed mechanism for preserving previous knowledge is therefore not expressed by the equations.\n\nIndependently, Section 4.1 makes OCT task 2 contain only DME and task 3 only Drusen. Triplet loss Ltri in Eq. (1) and Algorithm 1 needs an anchor, a positive, and a negative of different classes; a single-class batch cannot form triplets, so the only current-task loss is undefined or identically zero for those tasks. The model would receive no gradient from the new-class real images, yet Table 2 reports 64.43% average accuracy on OCT. This is an internal contradiction in the central experimental evidence, not merely a missing comparison.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a class-incremental learning method for medical image classification. For each new task, the student model is trained with triplet loss on the current real data and a knowledge-distillation loss applied to synthetic images of previous tasks generated by a shallow VAE. The distillation loss combines feature-attention matching (Eq. 4) and a covariance term (Eq. 7) that the text describes as aligning teacher and student embedding statistics. The method is evaluated on PI-CAI, OCT, PathMNIST, and CIFAR-10, and the authors report that it outperforms baselines on OCT and CIFAR-10 while avoiding storage of real past samples.","tokens_in":11802,"tokens_out":5202,"duration_ms":48462,"significance":"The target problem is relevant: privacy-preserving continual learning for medical imaging has practical value, and the proposed pipeline (VAE-based pseudo-replay plus distillation) is a plausible design direction. The paper evaluates on multiple datasets, includes a comparison with several baselines, and provides an ablation of the two distillation terms. However, the central mechanism as written is not what is claimed: Eq. (7) is not a teacher-student covariance alignment term, and the OCT protocol contains single-class tasks for which the only current-task loss (triplet loss) is undefined. These are load-bearing inconsistencies, not presentation issues, so the empirical claims are not currently supported.","major_comments":[{"comment":"The covariance loss as written is LCov = C(Zk) + C(Zk-1), a sum of two d x d covariance matrices, yet LKD and LMk in Eq. (1) are scalars. If the intended scalar is c(Zk) + c(Zk-1) using Eq. (6), then the loss only pushes the off-diagonal entries of each model's embedding covariance toward zero independently; there is no cross-model term comparing C(Zk) with C(Zk-1). The text and Figure 2 state that the loss implements covariance alignment between Teacher and Student, but the equations express no such alignment. The claimed mechanism for preserving previous knowledge through aligned feature correlations is therefore not implemented by the given loss.","section":"Section 3.2.2, Eq. (7)"},{"comment":"The OCT protocol assigns task 2 to the single class DME and task 3 to the single class Drusen. The triplet loss Ltri in Eq. (1) and Algorithm 1 requires an anchor, a positive, and a negative from a different class, so with a one-class task it is undefined or identically zero for those tasks. The student would receive no gradient from the current real images, yet Table 2 reports 64.43% average OCT accuracy. This is an internal contradiction in the central experimental evidence and makes the OCT results unreproducible from the described algorithm.","section":"Section 4.1, OCT benchmark"},{"comment":"The generator objective is described inconsistently. The text says the generator minimizes LG in ng iterations, with LG = -DE(Mk(xg), Mk-1(xg)); gradient descent on this LG maximizes the teacher-student distance, which is an adversarial objective, not one that 'adheres closely' to the teacher's distribution. It is also unclear whether the generator update is performed once per epoch over all tasks or per mini-batch. Clarify the exact optimization roles of the generator, student, and teacher, because this determines whether the pseudo-replay mechanism is actually as claimed.","section":"Section 3.2.1, Eq. (2) and Algorithm 1"}],"minor_comments":[{"comment":"The abstract describes the work solely in terms of PI-CAI prostate MRI detection, but the experiments also cover OCT, PathMNIST, and CIFAR-10; please align the abstract with the actual scope.","section":"Abstract"},{"comment":"The symbol n is used both for the number of samples in the covariance estimate in Eq. (5) and for the KD mini-batch size in the implementation details; this makes the formulas ambiguous.","section":"Eq. (5) and Section 4.1"},{"comment":"'The adding of the squared-off diagonal coefficients' should read 'the sum of the squared off-diagonal coefficients'; the current phrasing is a typo.","section":"Eq. (6)"},{"comment":"The tables report averages over ten trials but no standard deviations or statistical significance tests, which makes it hard to judge whether the differences from the nearest baselines are reliable.","section":"Tables 2 and 3"},{"comment":"The PI-CAI experiment is reported only as a single number in Table 1 without comparison to other IL baselines, so the claim that the method is effective on MRI data is not supported by the same evidence as the other datasets.","section":"Section 4.1, PI-CAI"}],"recommendation":"reject","confidential_remarks":null},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper combines triplet metric learning, feature attention matching, a Barlow-style covariance penalty, and VAE-based adversarial pseudo-replay into a privacy-preserving continual learning recipe for medical images. The combination is new and the medical motivation is legitimate, but two load-bearing inconsistencies in the method section make the main results not reproducible as stated.\n\nWhat it does well: The specific mix of components is not present in the cited prior work, and the ablation shows both KD terms contribute to the final accuracy. Evaluating on PI-CAI, OCT, PathMNIST, and CIFAR-10 is a reasonable, if somewhat scattershot, way to demonstrate versatility. The core idea--using a fixed teacher and a generated pseudo-replay set rather than storing real patient data--is practically relevant.\n\nSoft spots: First, Eq. 7 defines LCov = C(Zk) + C(Zk-1), which is a sum of two covariance matrices, not a scalar loss and not an alignment term. The text and Figure 2 describe covariance alignment between teacher and student, but the equation contains no cross-model term. If the intended scalar is c(Zk) + c(Zk-1) from Eq. 6, then the loss only drives off-diagonal entries of each model's covariance toward zero independently; it does not transfer the teacher's correlation structure. The mechanism as written is not the mechanism described.\n\nSecond, the OCT benchmark defines task 2 as only DME and task 3 as only Drusen. The triplet loss in Eq. 1 and Algorithm 1 requires an anchor, positive, and negative from different classes. A single-class batch cannot form triplets, so the only current-task loss is undefined or identically zero on those tasks. Yet Table 2 reports 64.43% average accuracy on OCT. This is an internal contradiction in the central experimental evidence, not a missing comparison.\n\nMinor concerns: no code or standard deviations are released, and hyperparameters (lambda, n_g, n_s, margins, etc.) are hand-picked on the same datasets. These are secondary.\n\nWho this is for: researchers in continual learning for medical imaging, especially those interested in privacy-preserving replay. The recipe is worth understanding, but the paper needs a careful rewrite that untangles the loss definitions and either fixes the OCT protocol or explains how triplets are formed in one-class tasks.\n\nRecommendation: I would not accept this version, but I would not desk-reject the idea. Send it to peer review with the expectation of heavy revision, a request for code, and a demand that the covariance loss and the OCT protocol be clarified.","headline":"A privacy-preserving continual learning recipe with a novel component mix, but two load-bearing inconsistencies--the covariance loss is not an alignment term and the OCT protocol makes triplet loss undefined--make the results unreproducible as stated.","tokens_in":12267,"tokens_out":4269,"would_cite":false,"duration_ms":40808,"reading_group":"maybe","serious_thinker":"no","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper argues that teacher-guided pseudo-replay plus feature- and covariance-distillation lets medical image classifiers add new tasks without storing old data, reporting final-task accuracy of 64.43% on OCT and 67.23% on CIFAR-10.","keywords":["catastrophic forgetting","incremental learning","knowledge distillation","medical image classification","generative replay","variational autoencoder","triplet loss","covariance regularization"],"falsifier":"Track the cross-covariance between Teacher and Student embeddings, $\\frac{1}{n-1}\\sum_i (z_i^k-m_k)(z_i^{k-1}-m_{k-1})^\\top$, during training on a fixed task sequence. If this quantity does not respond while $L_{Cov}=C(Z_k)+C(Z_{k-1})$ decreases, the loss is not aligning the two models' covariances; equivalently, replace $L_{Cov}$ by $C(Z_k)$ alone and check whether final-task accuracy changes. If replacing it leaves accuracy essentially unchanged, the claimed teacher-student covariance mechanism is not the active ingredient.","tokens_in":11212,"feed_emoji":"🩺","tokens_out":9871,"duration_ms":87437,"temperature":0.7,"pith_summary":"This paper sets out to show that an image classifier can keep learning new medical-imaging tasks without forgetting earlier ones, even when no real examples from past tasks are stored or replayed. The recipe is to train a shallow variational autoencoder against the fixed previous model so it generates synthetic images that mimic the old task's distribution, then to train the new model on the current task while distilling the previous model's normalized feature maps and embedding covariance structure. The authors report that this keeps accuracy high across tasks on prostate MRI, OCT, and histopathology data and on CIFAR-10, with final-task average accuracy of 64.43% on OCT and 67.23% on CIFAR-10, above the compared continual-learning baselines on those two datasets. If the claim holds, it offers a privacy-compatible way for medical centers to update models incrementally without exchanging or storing patient images.","feed_headline":"Teacher-guided synthetic images cut forgetting in medical classifiers","feed_subtitle":"A shallow VAE recreates past-task images so incremental models keep old knowledge without storing patient data.","key_machinery":"The central object is the distillation loss $L_{KD}=L_{FAM}+L_{Cov}$. $L_{FAM}$ is an $L^2$-norm feature-attention matching term that compares normalized feature maps at every convolutional layer except the embedding layer, so the Student keeps the Teacher's spatial and structural patterns. $L_{Cov}$ is a Barlow-Twins-style covariance regularizer: it penalizes the squared off-diagonal coefficients of the embedding covariance matrices, written as $L_{Cov}=C(Z_k)+C(Z_{k-1})$, where $C(Z)=\\frac{1}{n-1}\\sum_i (z_i-m)(z_i-m)^t$. A shallow VAE supplies the pseudo-replay data; its generator is trained adversarially to maximize $D_E(M_k(x_g),M_{k-1}(x_g))$, pushing the synthetic images toward the previous task's distribution. The triplet loss on real current-task data, together with a nearest-class-mean classifier, supplies the metric-learning backbone that keeps new classes separable and discriminative.","core_discovery":"The paper's central claim is that catastrophic forgetting in zero-shot class-incremental learning can be mitigated without using any real data from past tasks by combining two mechanisms: (1) a lightweight VAE generator trained adversarially against the fixed previous model produces synthetic images $\\hat{D}_k$ that approximate the previous task's distribution; (2) a knowledge-distillation loss $L_{KD}=L_{FAM}+L_{Cov}$ transfers the previous model's representation to the current one. The feature-attention term $L_{FAM}$ aligns $L^2$-normalized feature maps across intermediate layers, and the covariance term $L_{Cov}$ penalizes off-diagonal entries of the embedding covariance matrices of the Teacher and the Student. Training uses the total loss $L_{M_k}=L_{tri}(x_k)+\\lambda L_{KD}(x_g)+D_E(M_k,M_{k-1})$. On the PI-CAI prostate MRI dataset the method reaches 68.73% average accuracy versus a non-incremental upper bound of 83.21% and a fine-tuning lower bound of 26.25%; on OCT and CIFAR-10 it reports the best final-task average accuracy among the baselines (64.43% and 67.23%). On PathMNIST it reaches 53.75%, below the EFT baseline's 66.82%, a gap the paper attributes to that baseline's dynamic architecture. If these results hold, the method provides a privacy-compatible route to continual medical image analysis.","pith_inferences":["This goes beyond the paper: as written, $L_{Cov}$ is a sum of two within-model penalties with no explicit cross-model term; an explicit cross-covariance alignment term would be a sharper test and a plausible refinement of the stated mechanism.","This goes beyond the paper: the pseudo-replay design could be carried over to segmentation or detection tasks, where preserving past distributions without storing original annotations is even more valuable.","This goes beyond the paper: the PathMNIST result suggests the advantage of synthetic replay may shrink when tasks have many classes or strong visual similarity; varying class counts and task order would map where the method helps."],"forward_implications":["Medical-image models could be updated across new sites or disease categories without retaining patient scans, lowering storage and privacy barriers.","On the PI-CAI prostate MRI benchmark, the reported 68.73% average accuracy sits between the 83.21% joint-training upper bound and the 26.25% fine-tuning lower bound, indicating that most old-task knowledge can survive the transition.","Ablation results imply both distillation terms matter: using either feature-attention matching or covariance regularization alone drops OCT accuracy from 64.43% to roughly 47–50% and CIFAR-10 accuracy from 67.23% to roughly 44–46%.","Because the same framework improves a natural-image benchmark (CIFAR-10), the mechanism is not tied to medical-image statistics and may transfer to other class-incremental learning settings."],"supporting_citations":[{"why":"Defines the dataset splits and baseline comparison setup for OCT, PathMNIST, and CIFAR-10.","marker":"[55]"},{"why":"Supplies the adversarial teacher-student training idea used to guide the VAE generator.","marker":"[32]"},{"why":"Source of the Barlow-Twins covariance regularizer $c(Z)$ adapted into $L_{Cov}$.","marker":"[66]"},{"why":"Provides the PI-CAI prostate MRI dataset used in the main two-task evaluation.","marker":"[42]"},{"why":"Provides the OCT retinal image dataset used in the three-task comparison.","marker":"[21]"},{"why":"Provides the PathMNIST histopathology dataset used in the three-task comparison.","marker":"[62]"},{"why":"Provides the CIFAR-10 benchmark showing the method transfers beyond medical images.","marker":"[23]"},{"why":"Supplies the triplet loss that forms the metric-learning backbone of the Student objective.","marker":"[44]"},{"why":"One of the continual-learning baselines (LwF) that the method is compared against in Table 2.","marker":"[29]"},{"why":"Brain-inspired replay, a storage-free generative-replay baseline the method must beat in the comparison.","marker":"[53]"}],"fun_headline_variants":["No old data needed: VAE ghosts fight forgetting in MRI models","Synthetic past images let medical AI learn without forgetting","Forgetting-free incremental learning via generated training ghosts","VAE recreates old tasks so medical classifiers keep their memory","Privacy-safe continual learning: recycle old knowledge via synthetic data"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing assumption is that the covariance term, written as the sum of each model's own covariance penalty, actually transfers the previous model's feature-correlation structure to the current model; if it does not couple the two models' correlations, the reported forgetting protection may come from elsewhere in the training objective.","fun_headline_variants_meta":{"raw":{"variants":["No old data needed: VAE ghosts fight forgetting in MRI models","Synthetic past images let medical AI learn without forgetting","Forgetting-free incremental learning via generated training ghosts","VAE recreates old tasks so medical classifiers keep their memory","Privacy-safe continual learning: recycle old knowledge via synthetic data"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000635,"raw_usage":{"total_tokens":2980,"prompt_tokens":1049,"completion_tokens":1931,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":665,"completion_tokens_details":{"reasoning_tokens":1850}},"tokens_in":665,"tokens_out":1931,"duration_ms":11558,"temperature":1.0,"reasoning_tokens":1850,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T05:36:50.022957+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Track the cross-covariance between Teacher and Student embeddings, $\\frac{1}{n-1}\\sum_i (z_i^k-m_k)(z_i^{k-1}-m_{k-1})^\\top$, during training on a fixed task sequence. If this quantity does not respond while $L_{Cov}=C(Z_k)+C(Z_{k-1})$ decreases, the loss is not aligning the two models' covariances; equivalently, replace $L_{Cov}$ by $C(Z_k)$ alone and check whether final-task accuracy changes. If replacing it leaves accuracy essentially unchanged, the claimed teacher-student covariance mechanism is not the active ingredient.","supporting_citations":[{"cited_title":"Bosma, Jasper J","cited_arxiv_id":null,"evidence_quote":"Provides the PI-CAI prostate MRI dataset used in the main two-task evaluation."},{"cited_title":"Large dataset of labeled optical coherence tomography (oct) and chest x-ray images, 2018","cited_arxiv_id":null,"evidence_quote":"Provides the OCT retinal image dataset used in the three-task comparison."},{"cited_title":"Ting, et al","cited_arxiv_id":null,"evidence_quote":"Defines the dataset splits and baseline comparison setup for OCT, PathMNIST, and CIFAR-10."},{"cited_title":"Barlow twins: Self-supervised learning via redundancy reduction","cited_arxiv_id":null,"evidence_quote":"Source of the Barlow-Twins covariance regularizer $c(Z)$ adapted into $L_{Cov}$."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the PathMNIST histopathology dataset used in the three-task comparison."},{"cited_title":"Krizhevsky","cited_arxiv_id":null,"evidence_quote":"Provides the CIFAR-10 benchmark showing the method transfers beyond medical images."},{"cited_title":"Facenet: A unified embedding for face recognition and clustering","cited_arxiv_id":null,"evidence_quote":"Supplies the triplet loss that forms the metric-learning backbone of the Student objective."},{"cited_title":"Li and D","cited_arxiv_id":null,"evidence_quote":"One of the continual-learning baselines (LwF) that the method is compared against in Table 2."},{"cited_title":"Brain-inspired replay for continual learning with artificial neural networks","cited_arxiv_id":null,"evidence_quote":"Brain-inspired replay, a storage-free generative-replay baseline the method must beat in the comparison."}],"review_version":1}