{"id":"593b4362-7c40-4e7e-8f23-d0584de5e513","arxiv_id":"2501.01114","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":3.0,"correctness_risk":"high","formal_verification":"none","parameter_count":1,"one_line_summary":"GradProm gates the auxiliary recognition gradient by cosine similarity with the enhancement gradient, reporting improved PSNR, SSIM, and downstream accuracy on four medical image datasets.","lead":"This paper proposes GradProm, a training rule for medical image quality enhancement that adds the recognition model's gradient to the enhancement model's update only when the two gradients point in the same direction. The method is tested on denoising and super-resolution across four public medical datasets, reporting gains in PSNR, SSIM, and downstream recognition accuracy.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"As written, GradProm's update rule sums gradients with respect to disjoint parameter vectors θ and ϕ; the cosine similarity and the additive update are undefined, so the central theoretical claim and the algorithm itself are not well-posed.","rationale":"The reader's weakest_assumption correctly identifies the gradient-space mismatch. This is the single most load-bearing concern because the entire method consists of this update rule; if the update is undefined, there is no algorithm to test empirically. The proof has a second, independent false step, but even a corrected proof would not rescue the undefined operation. The empirical tables are extensive, but they cannot validate an ill-defined training procedure without code that shows the intended gradient computation. The lack of comparison with gradient-surgery baselines (e.g., PCGrad) is a further weakness, but it is secondary: the theoretical claim is already unsupported. Therefore the verdict REJECT stands. We recommend no change to the reader's verdict. Good faith: the authors may have intended G_VR = ∇_θ L_VR and wrote the notation sloppily; the test would reveal whether the intended algorithm is coherent. But as submitted, the central claim is not supported.","tokens_in":23267,"tokens_out":6026,"duration_ms":52565,"concrete_test":"Independently re-derive Lemma 3.1 from the formal definitions in §3.2. Write out the parameter spaces: let dim(θ) = d_1 and dim(ϕ) = d_2. Show that G_IP ∈ R^{d_1} and G_VR ∈ R^{d_2}; if d_1 ≠ d_2, cos(G_IP, G_VR) and the sum G_IP + G_VR are undefined. If d_1 = d_2 but the parameter sets are disjoint, the gradient of L_VR w.r.t. θ is not the same as ∇_ϕ L_VR; compute ∇_θ L_VR (via the chain rule through IP) and check whether the update matches the stated one. Finally, verify the proof step: if ⟨G_VR, ∇L_IP⟩ = 0 for two nonzero vectors, can we conclude ∇L_IP = 0? A single example, e.g., G_VR = (1,0), ∇L_IP = (0,1), settles that the inference is false.","verdict_should_be":"UNCHANGED","load_bearing_attack":"In Section 3.2, the paper defines G_IP = ∇_θ L_IP(θ) and G_VR = ∇_ϕ L_VR(ϕ), with θ and ϕ explicitly the parameters of IP and VR respectively. Eq. (5) writes L_total(θ_T) = L_IP(θ) + λ L_VR(ϕ), indicating θ and ϕ are separate blocks of the full parameter vector. In a typical task-driven IQE network, IP and VR have different architectures and therefore different parameter counts. The cosine similarity s = cos(G_IP, G_VR) and the update rule in Lemma 3.1, θ^{t+1}_T := θ^t_T − α_t(G^t_IP + G^t_VR max(0, cos(G^t_IP, G^t_VR))), require adding G_VR to a gradient vector for θ. Unless G_VR is redefined as ∇_θ L_VR (the gradient of the VR loss through the IP parameters), this addition is undefined: G_VR is not a tangent vector in the parameter space of θ. If one pads the gradients to the full parameter space, the cosine becomes 0 because the nonzero entries are disjoint, making the method identical to training IP alone, contradicting the reported gains. The paper never states that G_VR is ∇_θ L_VR; the notation L_VR(ϕ) and the definition in §3.2 point the other way. This makes Lemma 3.1 not merely unproven but ill-typed. Even setting that aside, the proof's inference that ⟨G_VR, ∇L_IP⟩=0 implies ∇L_IP=0 is false: orthogonality of two nonzero vectors does not force either to vanish. Both flaws hit the central claim that GradProm \"achieves the local minimum of the mainstream IP model\" without bias from VR.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes GradProm, a training strategy for task-driven medical image quality enhancement (IQE) in which the enhancement model (IP) and a downstream visual recognition model (VR) are trained with a gradient gate: when the gradients of the two sub-models have non-negative cosine similarity, the update to the IP parameters is augmented by the VR gradient; otherwise only the IP gradient is used. The authors claim a theoretical guarantee that this update converges to a local minimum of the IP loss without being biased by the VR loss, and they report state-of-the-art denoising and super-resolution results on ISIC 2018, COVID-CT, Lizard, and CAMUS, together with downstream diagnosis and segmentation performance.","tokens_in":23724,"tokens_out":5047,"duration_ms":48197,"significance":"If the theoretical claim and the algorithm were correctly stated, the paper would offer a simple, architecture-agnostic training modification with potentially broad applicability to medical image enhancement, and the experimental study is extensive: four public datasets, multiple noise levels and degradation types, two auxiliary tasks, supervised/unsupervised settings, domain generalization, and comparisons with several recent baselines. The paper also reports repeated runs with standard deviations, which is a strength. However, the central formalism is ill-typed, the proof of Lemma 3.1 contains a false logical inference, and several claimed improvement values in the text do not match the tables. These are not presentation issues: they undermine the theoretical contribution and the reproducibility of the method as written.","major_comments":[{"comment":"The method is not well-posed as written. The paper defines G_IP = ∇_θ L_IP(θ) and G_VR = ∇_ϕ L_VR(ϕ), with θ and ϕ parameters of IP and VR respectively, and Eq. (5) writes L_total(θ_T) = L_IP(θ) + λ L_VR(ϕ), indicating that θ and ϕ are different parameter blocks. The cosine similarity s = cos(G_IP, G_VR) and the update rule in Lemma 3.1, θ^{t+1}_T := θ^t_T − α_t(G^t_IP + G^t_VR max(0, cos(G^t_IP, G^t_VR))), require G_VR to be a tangent vector in the same parameter space as G_IP. As defined, G_VR is not; it is a gradient with respect to a disjoint parameter vector. If one pads the two gradients to the full parameter space, their cosine similarity is zero because their nonzero entries live in disjoint coordinates, which would make the method identical to training IP alone and contradict the reported gains. The paper never states that G_VR is intended to mean ∇_θ L_VR, the gradient of the VR loss with respect to the IP parameters through the composite function VR(IP(X)); the sentence in §3.2 that says \"we can use both L_IP(θ) and L_VR(ϕ) to update G_IP of the mainstream IP, i.e., G_IP = ∇_θ(L_IP(θ)+L_VR(ϕ))\" reassigns G_IP inconsistently. This ill-typed definition makes the central algorithm and the lemma it supports unreproducible from the text.","section":"§3.2, Eq. (5)-(6) and Lemma 3.1"},{"comment":"The proof contains a false inference. After expanding the inner product, the paper considers ⟨G_VR max(0, cos(G_IP, G_VR)), ∇L_IP⟩ = 0 and then states: \"Since max(0, cos(·)) ≥ 0, we can infer ... either ⟨G_VR, ∇L_IP⟩ = 0 or ⟨∇L_IP, G_VR⟩ = 0. In either case, we can conclude that ∇L_IP = 0, since the cosine similarity between two non-zero vectors is positive.\" This is invalid: two nonzero vectors can be orthogonal, and even if G_VR were zero, the conclusion ∇L_IP = 0 would not follow. Consequently, the claimed implication that a zero inner product forces the IP gradient to vanish is false, and the central assertion that GradProm \"can achieve the local minimum of the mainstream IP model\" is not established by the given argument.","section":"§3.3, Lemma 3.1 proof, Eq. (9)"},{"comment":"The proof's overall strategy is insufficient for the claimed conclusion. Showing ⟨Δθ^t_T, ∇L_IP^t⟩ ≥ 0 only establishes that the update direction is not an ascent direction at the current point; it does not imply convergence to a local minimum. A local-minimum guarantee would require a sufficient-descent condition relative to the gradient norm and a suitable learning-rate schedule, not merely the phrase \"under the condition that α_t is as small as possible.\" Without such a condition, the theorem is at best a statement about non-increase of a linearized objective. This is a load-bearing gap because the paper's advertised contribution includes the theoretical guarantee of unbiased convergence.","section":"§3.3, convergence claim"},{"comment":"Several claimed improvement magnitudes in the text do not match the reported tables. In the A-1 paragraph, after correctly reporting gains over Frozen-params for Table 1, the text says \"GradProm improves ESTR by 5.544/0.082, 7.085/0.108, 4.534/0.240, and 5.898/0.197 PSNR/SSIM\" for Table 2. These differences do not correspond to GradProm versus Frozen-params in Table 2; for example, at σ=0.05 the frozen-params unsupervised PSNR is 32.826 and GradProm is 33.525, a difference of 0.699, while 5.544 equals 36.825 − 31.281, i.e., the difference between GradProm (supervised) and Benchmark-i. This makes the strength of the reported improvements unclear and needs correction.","section":"§4.4, A-1 paragraph"}],"minor_comments":[{"comment":"Figure 3 and the surrounding text are inconsistent about whether VR parameters are updated: the Abstract says GradProm \"updates only parameters of the image enhancement model,\" but §3.2 says \"L_VR(ϕ) of the auxiliary V_R is only used to updated by its own gradient G_VR\" and Eq. (6) explicitly includes ∇_ϕ L_VR in the total gradient. Please clarify whether VR parameters are frozen, updated, or jointly trained.","section":"§3.2 and Figure 3"},{"comment":"The notation in the two-case display for G_T is inconsistent with the definitions: the first case writes ∇_θ(L_IP(θ)+λ L_VR(ϕ)) for the IP block, but earlier G_IP was defined as ∇_θ L_IP(θ), so the left-hand side should be a total update direction, not G_IP. Please introduce separate symbols for the update direction and the individual gradients.","section":"§3.2, equation after Eq. (6)"},{"comment":"The paper states that λ is set to 10^-4 but provides no sensitivity analysis or ablation over λ; since λ controls the contribution of the VR loss, a small study or at least a discussion of its role would strengthen the empirical claims.","section":"§4.1, Experiment Setup"},{"comment":"The text in A-5 says the performance gain of GradProm in the challenging scenario is reduced compared to Tables 1 and 2, but Table 3 shows GradProm outperforming Frozen-params by much larger relative SSIM margins (e.g., 0.088 and 0.144) than many entries in Tables 1 and 2; this qualitative statement should be checked against the actual numbers.","section":"§4.4, Table 2 vs Table 3"},{"comment":"There are several typographical errors and unclear phrases, including \"the the image enhancement model\" (Abstract), \"mainstream mainstream\" (§4.4 qualitative results), \"Experimrntal\" (Table 4 caption), and duplicated SSIM header in Table 7. These do not affect the technical content but should be corrected.","section":"Throughout"}],"recommendation":"reject","confidential_remarks":"The core problem is not a disagreement with the field's consensus but an internally inconsistent formalism: the algorithm as written operates on gradients in different parameter spaces, and the proof of the central lemma contains a false orthogonality inference. These flaws are load-bearing for the paper's main claimed contribution, so the manuscript cannot be accepted in its current form. If the authors re-define G_VR as ∇_θ L_VR, correct the proof to a weaker statement (e.g., the gated update avoids acute-angle conflict with the IP gradient), and fix the inconsistent comparison numbers in §4.4, a substantially revised version could merit re-review. The empirical results are extensive and may be salvageable, but the theory and the algorithm specification must be repaired first."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nYou should know two things about this paper before deciding what to do with it. First, the empirical part is more solid than the theory: GradProm consistently improves over ESTR and other baselines on four medical datasets, with ablations that are internally consistent. Second, the theoretical claim that anchors the paper does not hold as written, and the gradient definitions are inconsistent.\n\nWhat is actually new: applying gradient-similarity gating to task-driven medical IQE, where the auxiliary recognition loss only contributes to the enhancement update when its gradient is aligned with the main gradient. That is a reasonable training trick, and the paper shows it can improve PSNR/SSIM and downstream metrics. The experiments cover denoising and super-resolution, supervised and unsupervised auxiliary tasks, and a domain generalization check. I believe the empirical gains; the ablations are extensive and the margins over the frozen-params baseline are consistent across noise levels and datasets.\n\nWhere it falls apart: the method as formulated is not well-posed. Section 3.2 defines G_IP = ∇_θ L_IP(θ) and G_VR = ∇_ϕ L_VR(ϕ), with θ and ϕ the parameters of IP and VR. The update rule in Lemma 3.1 adds G_VR to a gradient for θ. Unless G_VR is redefined as ∇_θ L_VR, the cosine similarity is between vectors in different spaces and the sum is undefined. The paper never says that. This is not a cosmetic issue; it is the core update. The proof is also wrong: it argues that if the inner product of the update with ∇L_IP is zero, then ∇L_IP = 0, because two non-zero vectors have positive cosine. That is false; orthogonal non-zero vectors have zero cosine. The proof at most establishes a descent direction property, not convergence to a local minimum.\n\nThe novelty is thin but honest: the update is a hard-gated variant of gradient surgery (PCGrad) and Du et al.'s adaptive auxiliary loss weighting, both of which the paper cites. I would not call it a new result, but the application to medical IQE is not something I have seen done this way. The bigger problem is that the paper never compares against these gradient-surgery baselines, and no code is provided, so the empirical advantage over ESTR could be coming from the gating, but we cannot check against the closest methods.\n\nVerdict: the paper deserves a serious referee, not a desk reject, because the empirical work is substantial and the method is likely salvageable. But as submitted, the central theoretical guarantee is invalid and the algorithm needs redefinition. A referee should ask for a corrected formulation, a valid proof or a clear statement of the descent property, and comparisons with PCGrad/Du et al., plus code. I would not cite it in its current form.","headline":"Solid empirical story undermined by an ill-typed update rule and an invalid proof; salvageable but not as written.","tokens_in":24198,"tokens_out":3518,"would_cite":false,"duration_ms":33636,"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":"GradProm adds a recognition model's gradient to medical image enhancement training only when the two gradients point the same way, preventing the recognition task from biasing the enhancer.","keywords":["image quality enhancement","medical image processing","task-auxiliary learning","multi-task learning","gradient promotion","cosine similarity","medical image denoising","medical image super-resolution"],"falsifier":"Take any task-driven IQE network in which the enhancement and recognition sub-models have disjoint parameter sets and compute $\\cos(\\nabla_\\theta L_{IP}, \\nabla_\\phi L_{VR})$; because the two gradients live in different vector spaces, the cosine similarity is undefined, so Lemma 3.1's update rule cannot be executed as written unless the paper explicitly redefines $G_{VR}$ as $\\nabla_\\theta L_{VR}$.","tokens_in":23087,"feed_emoji":"🩻","tokens_out":12891,"duration_ms":104049,"temperature":0.7,"pith_summary":"This paper proposes GradProm, a training strategy for task-driven medical image quality enhancement in which an image enhancement model and a visual recognition model are trained together, but the recognition model's gradient is used to update the enhancement model only when the two gradients point in the same direction, as measured by cosine similarity. The paper's central claim is that this gating prevents the auxiliary recognition task from biasing the enhancement model's optimization direction, while still allowing recognition information to help when it is aligned with the enhancement objective. The authors prove that with a sufficiently small learning rate, the GradProm update keeps a non-negative inner product with the enhancement gradient and can reach a local minimum of the enhancement loss. Across four medical datasets (skin lesion, COVID CT, histopathology, and ultrasound), they report consistent gains in PSNR and SSIM over joint training, frozen-parameter training, and several recent denoising and super-resolution baselines, together with improved diagnosis and segmentation accuracy. If the claim is right, medical images can be restored to be both more human-perceivable and more machine-usable without changing the network architecture or adding any inference cost.","feed_headline":"Cosine-gated gradients boost medical image quality and recognition","feed_subtitle":"Adding only aligned recognition gradients boosts measured quality and accuracy on four medical datasets.","key_machinery":"The mechanism is the gradient gate $\\max(0, \\cos(G_{IP}, G_{VR}))$, a scalar computed from the cosine similarity between the enhancement-model gradient and the recognition-model gradient. At every training step the enhancement update is $G_{IP} + G_{VR}\\max(0,\\cos(G_{IP},G_{VR}))$, so the auxiliary gradient is switched on when the two gradients agree and switched off when they disagree. This gate is the object the theoretical analysis works with: the non-negativity of the inner product between the total update and $G_{IP}$ is what the local-minimum proof relies on.","core_discovery":"GradProm partitions a task-driven IQE system into a mainstream image enhancement model with parameters $\\theta$ and an auxiliary visual recognition model with parameters $\\phi$, and updates only the enhancement model's parameters. With the paper's notation $G_{IP}=\\nabla_\\theta L_{IP}(\\theta)$ and $G_{VR}=\\nabla_\\phi L_{VR}(\\phi)$, GradProm's update for the enhancement parameters is $\\theta^{t+1} = \\theta^t - \\alpha^t\\big(G_{IP}^t + G_{VR}^t \\max(0, \\cos(G_{IP}^t, G_{VR}^t))\\big)$; when the cosine similarity is negative, the auxiliary gradient term is dropped. The paper claims that this rule guarantees the inner product of the total update with the enhancement gradient $\\nabla_\\theta L_{IP}$ is non-negative, so the enhancement model's descent direction is never biased toward the recognition model's direction, and that GradProm can achieve a local minimum of the mainstream model. Empirically, the authors report state-of-the-art PSNR/SSIM for denoising and super-resolution on ISIC 2018, COVID-CT, Lizard, and CAMUS, with better downstream accuracy and mIoU than the compared training strategies, and show that the method generalizes across datasets and image modalities without extra data or architecture changes.","pith_inferences":["A natural extension is to replace the hard on/off threshold at $\\cos=0$ with a continuous weight, for example scaling $G_{VR}$ by a smooth function of the cosine similarity; the paper explicitly declines to explore intermediate alignments, so whether finer gating helps is an open empirical question.","The proof only establishes a non-negative inner product with the enhancement gradient, which is a much weaker condition than a descent inequality on a nonconvex network; a stronger convergence guarantee would require bounding how much the auxiliary term can increase $L_{IP}$.","Part of the gain might come from the gate acting as a denoiser of the enhancement gradient itself, rather than from the recognition task's semantic content; a control experiment replacing $G_{VR}$ with random vectors whose alignment distribution matches the real one would isolate that effect.","The same gating logic applies to any cascade of a main model and an auxiliary model, such as detection or compression branches, because the rule only needs two gradient vectors and a cosine similarity."],"forward_implications":["Task-driven medical image enhancement can be improved without extra data, new architectures, or added inference time; the only extra training-time cost is computing one cosine similarity per update.","The gating rule should transfer across medical modalities and task combinations, since the paper shows gains on skin lesion, CT, histology, and ultrasound data with both classification and segmentation as the auxiliary task.","If the proof holds, GradProm converges to a local minimum of the enhancement model even with a recognition model attached, so the recognition task can assist enhancement without pulling it toward the recognition model's own optimum.","Cross-dataset generalization improves: training on one medical dataset and testing on another yields higher PSNR/SSIM with GradProm than with the frozen-parameter baseline.","Combining two auxiliary recognition tasks can hurt rather than help, so the method is best used with auxiliary tasks whose feature requirements are similar to each other and to the enhancement task."],"supporting_citations":[{"why":"Defines the task-driven IQE paradigm of jointly training enhancement and recognition, and provides the frozen-parameter baseline that GradProm is compared against.","marker":"[1]"},{"why":"SR-ResNet is the mainstream image enhancement architecture used for the enhancement branch and as the Benchmark-i baseline.","marker":"[38]"},{"why":"ResNet serves as the auxiliary diagnosis model and as the backbone of the recognition branches.","marker":"[6]"},{"why":"UNet with a ResNet backbone is the auxiliary semantic segmentation model used throughout the experiments.","marker":"[39]"},{"why":"Provides the gradient-similarity and cosine-alignment idea used to decide when the auxiliary gradient is beneficial.","marker":"[57]"},{"why":"Another source for measuring the agreement of two optimization directions by cosine similarity in the update rule.","marker":"[37]"},{"why":"ISIC 2018 is the primary dataset for the ablation studies and the main denoising comparison.","marker":"[40]"},{"why":"COVID-CT supplies the CT-modality denoising results that demonstrate cross-modality generalization.","marker":"[41]"},{"why":"Lizard supplies the super-resolution evaluations and the cross-dataset generalization test.","marker":"[42]"},{"why":"CAMUS provides the ultrasound denoising results, including genuine low-quality images with speckle noise.","marker":"[43]"}],"fun_headline_variants":["Cosine-gated gradients improve medical image enhancement and recognition","GradProm: only aligned gradients train image quality and vision","Bias-free gradient gating sharpens medical images and task accuracy","Align gradients to boost both image quality and recognition"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole method depends on comparing and adding two gradients that are defined with respect to different models' parameters; if those parameter sets do not overlap, the cosine similarity and the update are not well defined.","fun_headline_variants_meta":{"raw":{"variants":["Cosine-gated gradients improve medical image enhancement and recognition","GradProm: only aligned gradients train image quality and vision","Bias-free gradient gating sharpens medical images and task accuracy","Align gradients to boost both image quality and recognition"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000652,"raw_usage":{"total_tokens":3052,"prompt_tokens":1074,"completion_tokens":1978,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":690,"completion_tokens_details":{"reasoning_tokens":1911}},"tokens_in":690,"tokens_out":1978,"duration_ms":12771,"temperature":1.0,"reasoning_tokens":1911,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T22:34:55.221493+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take any task-driven IQE network in which the enhancement and recognition sub-models have disjoint parameter sets and compute $\\cos(\\nabla_\\theta L_{IP}, \\nabla_\\phi L_{VR})$; because the two gradients live in different vector spaces, the cosine similarity is undefined, so Lemma 3.1's update rule cannot be executed as written unless the paper explicitly redefines $G_{VR}$ as $\\nabla_\\theta L_{VR}$.","supporting_citations":[{"cited_title":"Exploring simple and transferable recognition-aware image processing,","cited_arxiv_id":null,"evidence_quote":"Defines the task-driven IQE paradigm of jointly training enhancement and recognition, and provides the frozen-parameter baseline that GradProm is compared against."},{"cited_title":"Photo- realistic single image super-resolution using a generative adver- sarial network,","cited_arxiv_id":null,"evidence_quote":"SR-ResNet is the mainstream image enhancement architecture used for the enhancement branch and as the Benchmark-i baseline."},{"cited_title":"Deep residual learning for image recognition,","cited_arxiv_id":null,"evidence_quote":"ResNet serves as the auxiliary diagnosis model and as the backbone of the recognition branches."},{"cited_title":"U-net: Convolutional networks for biomedical image segmentation,","cited_arxiv_id":null,"evidence_quote":"UNet with a ResNet backbone is the auxiliary semantic segmentation model used throughout the experiments."},{"cited_title":"Gradient surgery for multi-task learning,","cited_arxiv_id":null,"evidence_quote":"Provides the gradient-similarity and cosine-alignment idea used to decide when the auxiliary gradient is beneficial."},{"cited_title":"Prompt-aligned gradient for prompt tuning,","cited_arxiv_id":null,"evidence_quote":"Another source for measuring the agreement of two optimization directions by cosine similarity in the update rule."},{"cited_title":"Skin lesion analysis toward melanoma detection 2018: A chal- lenge hosted by the international skin imaging collaboration (isic),","cited_arxiv_id":null,"evidence_quote":"ISIC 2018 is the primary dataset for the ablation studies and the main denoising comparison."},{"cited_title":"Covid-ct- dataset: a ct scan dataset about covid-19,","cited_arxiv_id":null,"evidence_quote":"COVID-CT supplies the CT-modality denoising results that demonstrate cross-modality generalization."},{"cited_title":"Lizard: a large-scale dataset for colonic nuclear instance segmentation and classification,","cited_arxiv_id":null,"evidence_quote":"Lizard supplies the super-resolution evaluations and the cross-dataset generalization test."},{"cited_title":"Deep learning for segmentation using an open large-scale dataset in 2d echocardiography,","cited_arxiv_id":null,"evidence_quote":"CAMUS provides the ultrasound denoising results, including genuine low-quality images with speckle noise."}],"review_version":1}