{"id":"73ef03f9-944a-44d5-b29e-84d5da94fbbb","arxiv_id":"2501.01392","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"A GAN with feature pyramid encoder and cross-channel mixing produces lower-FID counterfactual explanations for prostate MRI classification than the StylEx baseline.","lead":"ProjectedEx is a modified GAN framework that creates counterfactual images to explain prostate cancer MRI classifiers. On the PI-CAI dataset it reports lower FID than the [Re]StylEx baseline for four classifiers, but with minimal experimental details.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The FID advantage over [Re]StylEx is unverifiable because the paper does not state what images were scored or how the baseline was trained; if FID protocols differ, the central comparison collapses.","rationale":"The paper's strongest claim is an empirical superiority result, so the load-bearing condition is that the comparison is apples-to-apples. The reader identified this as the weakest assumption, and I agree: Section IV.B describes only ProjectedEx's training, Section IV.C gives no FID computation details, and no code or seed-level results are provided. My reading sharpens the concern by noting that even the image population used for FID is unspecified; because ProjectedEx has strong reconstruction losses, a FID computed on reconstructed inputs would be incomparable to one computed on attribute-modified counterfactuals. This is not an internal contradiction in the method, and the visualizations do show that adjusting discovered attributes moves classifier logits as intended, so I do not consider the claim refuted. The absence of a controlled protocol makes the central comparison unverified rather than false. Given the reader's CONDITIONAL verdict already reflects this gap, no verdict change is needed. If the protocol were supplied and the comparison survived, the architectural contribution would be credible; if the protocol is not supplied, the paper should be revised to state it or the claim downgraded.","tokens_in":9419,"tokens_out":3915,"duration_ms":39428,"concrete_test":"Request the released code and the exact [Re]StylEx configuration; then rerun both methods on the same PI-CAI split with identical preprocessing, training steps, and attribute-selection procedure, and compute FID for both models on the same generated image set (e.g., the same number of attribute-modified counterfactual images) against the same real-image reference set and Inception-V3 feature layer, repeated over at least three seeds. If ProjectedEx's FID remains lower by more than the seed-level standard deviation, the claim is supported; if the gap shrinks or reverses, the headline comparison was an artifact of protocol differences.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that ProjectedEx 'consistently outperforms [Re]StylEx across all classifiers' on FID (Section IV.C, Table III). For this claim to hold, the two models must be trained and evaluated under conditions different only in architecture. The paper supplies training details for ProjectedEx only (Section IV.B) and gives no training configuration, attribute budget, or image-set definition for [Re]StylEx. More seriously, no FID protocol is reported: it is unclear whether FID is computed on reconstructions, on counterfactual images after attribute modification, on which real reference set, with which Inception feature layer, or with how many samples. Because ProjectedEx is explicitly trained to reconstruct x with L1, LPIPS, and latent-space losses, its outputs will trivially resemble the input; a FID computed on such near-identity reconstructions would be artificially low compared with a FID computed on attribute-edited counterfactuals. Without a shared, specified protocol, the reported 108.63 vs. 127.08 gap cannot be attributed to the architectural contributions.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes ProjectedEx, a generative counterfactual explanation framework for prostate cancer MRI classification. The method adapts the StylEx / [Re]StylEx paradigm to medical images by introducing a feature pyramid encoder with multiscale discriminators, differentiable random projections via cross-channel mixing, and a StyleSpace construction that concatenates multiscale features with classifier logits. The authors report classifier performance (accuracy, precision, recall, F1) for four classifiers on the PI-CAI dataset and claim that ProjectedEx consistently outperforms [Re]StylEx in FID across all classifiers, with the best FID of 108.63 paired with EfficientNet-B. The paper includes qualitative visualizations of attribute modifications and discussion of clinical implications.","tokens_in":9588,"tokens_out":3273,"duration_ms":32586,"significance":"If the FID improvement is real and reproducible, the paper offers a useful architectural contribution to medical-image counterfactual explanation: a domain-specific adaptation of StyleGAN-based explainability that improves generation quality for prostate MRI. The use of a public dataset (PI-CAI) and the promise of code release are positive factors. However, the current experimental reporting is insufficient to establish the central quantitative claim, and the paper does not yet demonstrate that the improved FID translates into more interpretable or clinically usable explanations beyond qualitative examples.","major_comments":[{"comment":"The central claim that ProjectedEx consistently outperforms [Re]StylEx in FID cannot be verified because the FID computation protocol is not specified. It is unclear whether FID is computed on reconstructions, on attribute-edited counterfactual images, or on both; which real-image set serves as the reference distribution; how many samples are used; and which Inception feature layer is employed. This matters because ProjectedEx is trained to reconstruct x with L1, LPIPS, and latent-space losses, so a FID computed on near-identity reconstructions would be artificially low and would not measure the quality of counterfactual explanations. Please report the exact protocol and, ideally, separate FID values for reconstructions and counterfactual edits.","section":"Section IV.C, Table III"},{"comment":"No training configuration is given for the [Re]StylEx baseline. The comparison is meaningful only if the baseline is trained and evaluated under identical conditions: the same PI-CAI train/test split, the same 64x64 three-channel preprocessing, the same training budget (number of steps, learning rate, batch size), and the same attribute budget. Only ProjectedEx details are provided. Please supply the baseline configuration for every classifier row in Table III, or a precise pointer to the implementation and hyperparameters used.","section":"Section IV.B"},{"comment":"The reported FID values are single numbers with no error bars, no repeated-run variance, and no statistical test. This is particularly important for the ShuffleNetv2 row, where the difference is 149.29 versus 144.71 (about 3%), so measurement noise could change the qualitative conclusion. Please report means and standard deviations over multiple seeds, or bootstrap confidence intervals, for the FID comparison.","section":"Section IV.C, Table III"}],"minor_comments":[{"comment":"The text says 'three classifiers: MobileNetv2, ResNet18, ShuffleNetv2 and EfficientNet-B', but four classifiers are listed; please correct to 'four classifiers'.","section":"Section IV.C"},{"comment":"The paragraph discussing FID results refers to 'Table II', but the FID results are presented in Table III; the cross-reference should be corrected.","section":"Section IV.C"},{"comment":"The subsection title 'Evaluation Matrices' should be 'Evaluation Metrics'.","section":"Section IV.A"},{"comment":"The PI-CAI dataset is used but no reference or URL is provided; please add a citation for the dataset.","section":"Section IV.A"},{"comment":"The conclusion claims 'state-of-the-art results', but the only empirical comparison is against [Re]StylEx; no comparison is made to other medical-image counterfactual explanation methods, such as the chest X-ray StylEx adaptation cited in the related work, so the 'state-of-the-art' wording overstates the evidence.","section":"Section V"},{"comment":"The loss notation is slightly inconsistent: Lrec is said to include Lx_rec, LLPIPS, and Lw_rec, but Lx_rec and Lw_rec are not defined in the equation block; please clarify the notation.","section":"Section III.D"}],"recommendation":"major_revision","confidential_remarks":"The paper is a plausible engineering contribution, but its primary quantitative claim depends on a baseline comparison that is currently under-specified. If the authors can provide the missing experimental protocols and rerun the evaluation with appropriate statistical uncertainty, the paper could be acceptable; as written, the central FID claim is not yet reproducible."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is an incremental but sensible extension of StylEx, and the method section reads clearly. The problem is that the empirical core—the FID comparison against [Re]StylEx—is underspecified to the point of being unverifiable. I don't see a fatal internal error, but the results as reported are not yet evidence.\n\nThe genuinely new bits are the feature pyramid encoder with four discriminators and the use of random 1x1 convolutions for channel mixing. These are standard tools, but their combination in this context is a reasonable thing to try. The paper also does a decent job explaining the StyleSpace construction and the loss function. It's honest about using the [Re]StylEx reimplementation as a baseline.\n\nThe large soft spot is exactly what you flagged: there is no description of what images are scored for FID, which reference set is used, how many samples, or whether the numbers come from reconstructions or attribute-edited counterfactuals. Since ProjectedEx is trained with L1, LPIPS, and latent-space reconstruction losses, its outputs will obviously be close to the input. If FID is computed on those reconstructions, it will be artificially low compared to a protocol that scores edited counterfactuals. The paper doesn't say, so the comparison to [Re]StylEx could collapse entirely. There are also no error bars or statistical tests, and no training details for [Re]StylEx. The word 'significantly' in the claim is doing no work.\n\nI'd also push back on the 'state-of-the-art' claim in the conclusion. The comparison is against a single baseline, and the classifier numbers are just single runs. That's not enough to call anything state-of-the-art. The related work section is padded with many self-citations, some only tangentially relevant; not a fatal flaw, but it doesn't help.\n\nWho is this for? People working on counterfactual explanations in medical imaging might want to try the feature pyramid idea. But they'd need to ask the authors for the missing details.\n\nRecommendation: If this lands on my desk, I'd send it to a referee—the architecture is concrete and the medical setting is relevant—but the referee should be told to focus on the evaluation protocol. The paper needs major revisions before any FID claim can be accepted.","headline":"A reasonable StylEx variant for prostate MRI whose FID claim cannot be checked from the paper as written.","tokens_in":10189,"tokens_out":3014,"would_cite":false,"duration_ms":29221,"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":"ProjectedEx makes counterfactual prostate MRI explanations more realistic.","keywords":["ProjectedEx","explainable AI","counterfactual explanations","prostate cancer MRI","StyleSpace","feature pyramid encoder","generative adversarial networks","PI-CAI dataset"],"falsifier":"Retrain [Re]StylEx and ProjectedEx from the same PI-CAI 8:2 split with identical $64\\times64$ three-channel DWI/T2WI/ADC inputs, the same batch size, step count, learning-rate schedule, and FID reference statistics; if the FID gap shrinks to zero or reverses, the central claim fails. A second, complementary check: train ProjectedEx with the feature pyramid encoder removed and measure whether the FID returns to baseline levels.","tokens_in":9208,"feed_emoji":"🩺","tokens_out":7165,"duration_ms":63760,"temperature":0.7,"pith_summary":"ProjectedEx is a generative framework that tries to make AI prostate-cancer classification explainable by producing counterfactual MRI explanations: images that look like the patient's scan but are altered just enough to flip or shift a classifier's low-risk/high-risk decision. The paper argues that the standard approach for such explanations, which was designed for natural images, produces unrealistic and clinically murky MRI edits, and that its own modifications fix this. It reports that, on the PI-CAI dataset with DWI, T2WI, and ADC as three channels, ProjectedEx lowers FID relative to the [Re]StylEx baseline for every classifier tested, with the best FID of 108.63 versus 127.08 using EfficientNet-B. If these results hold under matched conditions, the architectural changes—multiscale feature extraction and a logit-conditioned latent space—are a working recipe for closing the realism gap in medical-image explanation.","feed_headline":"ProjectedEx lowers FID for MRI explanations with every classifier tested","feed_subtitle":"More realistic generated counterfactuals make AI prostate-cancer decisions easier for clinicians to inspect.","key_machinery":"The load-bearing mechanism is a logit-conditioned StyleSpace built from multiscale features. The feature pyramid encoder extracts features at resolutions $64^2$, $32^2$, $16^2$, and $8^2$; each layer passes through a lightweight spectral-normalized discriminator, and all discriminator outputs are brought to a fixed 128-resolution layout and concatenated into 512 dimensions. Adding the classifier's two raw logits gives a 514-dimensional latent code, so the generator is told what decision the regenerated image is supposed to provoke. Differentiable random projections via Cross-Channel Mixing—a randomly initialized $1\\times1$ convolution—mixes channels at each scale without needing an invertible rotation matrix. The training loss combines adversarial loss, path regularization, L1 and LPIPS reconstruction, latent-space reconstruction, and KL divergence between classifier outputs, and after training, individual StyleSpace coordinates are iteratively tested to find the ones that move the classifier's logits most.","core_discovery":"The central claim is that the realism of GAN-based counterfactual explanations for prostate MRI can be substantially improved by upgrading the encoder and the adversarial feedback pathway rather than by changing the explanation logic. Concretely, ProjectedEx replaces a plain encoder with a feature pyramid encoder whose four layers are each scored by a lightweight spectral-normalized discriminator; the outputs are unified to a fixed resolution, concatenated into a 512-dimensional StyleSpace, and then combined with the classifier's two logits to form a 514-dimensional latent code. The paper further replaces rotation-matrix projections with differentiable random projections via a randomly initialized $1\\times1$ convolution using a standard weight-initialization scheme. The result, as reported, is consistently lower FID than [Re]StylEx across MobileNetV2, ResNet18, ShuffleNetV2, and EfficientNet-B, with the largest relative improvement on MobileNetV2 and the best absolute FID of 108.63 paired with the best classifier, EfficientNet-B at 83.97% accuracy. On the paper's own terms, this shows that multiscale feedback plus classifier conditioning improves generated explanation quality and that attribute edits in StyleSpace shift classifier logits between low-risk and high-risk categories.","pith_inferences":["Because Cross-Channel Mixing replaces rotation matrices with a randomly initialized $1\\times1$ convolution, the same projection trick can likely be dropped into other GAN-based medical explainers without retraining the discriminator stack; that is our extrapolation, not the paper's.","The $64\\times64$ crop centered on a lesion removes surrounding anatomy, so the method may miss location- or context-dependent cues; a natural test is to apply ProjectedEx at full field of view and compare FID and attribute interpretability.","The binary ISUP split puts Gleason 3+4 and 4+3 in different risk classes, so the reported attribute edits reflect coarse risk rather than fine grade; a three- or five-class version could reveal whether the same StyleSpace coordinates remain semantically coherent."],"forward_implications":["ProjectedEx reports lower FID than [Re]StylEx under every classifier tested (178.21 to 117.78 for MobileNetV2, 134.20 to 113.42 for ResNet18, 149.29 to 144.71 for ShuffleNetV2, 127.08 to 108.63 for EfficientNet-B), so the paper's recipe improves generative realism without changing the explanation logic.","Because StyleSpace coordinates are found by scanning which latent directions move classifier logits, the framework yields per-attribute counterfactuals: editing the top four attributes flips or shifts images between low-risk and high-risk classifications, which is the ingredient clinicians would inspect.","The best FID (108.63) coincides with the strongest classifier (EfficientNet-B, 83.97% accuracy, 91.26% F1), suggesting that explanation quality benefits from pairing with a capable classifier.","The loss already includes LPIPS and latent-space reconstruction, so the framework is engineered to preserve patient-specific appearance while editing only decision-relevant attributes, a direct corollary of the reconstruction terms in the objective."],"supporting_citations":[{"why":"Defines the base framework: a GAN trained in StyleSpace to explain a classifier; this is the method ProjectedEx modifies.","marker":"[8]"},{"why":"Is the reimplemented PyTorch baseline ([Re]StylEx) whose FID scores ProjectedEx claims to beat.","marker":"[29]"},{"why":"Provides the generator backbone and the path-regularization term used in the training loss.","marker":"[31]"},{"why":"Supplies the weight-initialization scheme used for the $1\\times1$ convolution in Cross-Channel Mixing.","marker":"[33]"},{"why":"Defines the adversarial loss used in the overall objective.","marker":"[34]"},{"why":"Provides the LPIPS perceptual-loss term used in reconstruction.","marker":"[35]"},{"why":"Defines the ISUP grade-group mapping that determines the low-risk/high-risk labels.","marker":"[36]"}],"fun_headline_variants":["ProjectedEx: Sharper MRI counterfactuals for prostate cancer AI","Multiscale feedback lowers FID for prostate MRI explanations","Feature pyramid encoder improves explainable prostate cancer AI","Improving counterfactual realism in prostate MRI with ProjectedEx"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The gap in FID scores assumes the baseline [Re]StylEx was trained and measured under exactly the same protocol as ProjectedEx; because the paper gives training details only for ProjectedEx, differences in data split, preprocessing, training budget, or FID computation could explain the observed improvement.","fun_headline_variants_meta":{"raw":{"variants":["ProjectedEx: Sharper MRI counterfactuals for prostate cancer AI","Multiscale feedback lowers FID for prostate MRI explanations","Feature pyramid encoder improves explainable prostate cancer AI","Improving counterfactual realism in prostate MRI with ProjectedEx"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000713,"raw_usage":{"total_tokens":3237,"prompt_tokens":1002,"completion_tokens":2235,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":618,"completion_tokens_details":{"reasoning_tokens":2165}},"tokens_in":618,"tokens_out":2235,"duration_ms":15073,"temperature":1.0,"reasoning_tokens":2165,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T22:27:59.470657+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Retrain [Re]StylEx and ProjectedEx from the same PI-CAI 8:2 split with identical $64\\times64$ three-channel DWI/T2WI/ADC inputs, the same batch size, step count, learning-rate schedule, and FID reference statistics; if the FID gap shrinks to zero or reverses, the central claim fails. A second, complementary check: train ProjectedEx with the feature pyramid encoder removed and measure whether the FID returns to baseline levels.","supporting_citations":[{"cited_title":"Explaining in style: Training a gan to explain a classifier in stylespace,","cited_arxiv_id":null,"evidence_quote":"Defines the base framework: a GAN trained in StyleSpace to explain a classifier; this is the method ProjectedEx modifies."},{"cited_title":"[re] explaining in style: Training a gan to explain a classifier in stylespace,","cited_arxiv_id":null,"evidence_quote":"Is the reimplemented PyTorch baseline ([Re]StylEx) whose FID scores ProjectedEx claims to beat."},{"cited_title":"Analyzing and improving the image quality of stylegan,","cited_arxiv_id":null,"evidence_quote":"Provides the generator backbone and the path-regularization term used in the training loss."},{"cited_title":"Delving deep into rectifiers: Surpassing human-level performance on imagenet classification,","cited_arxiv_id":null,"evidence_quote":"Supplies the weight-initialization scheme used for the $1\\times1$ convolution in Cross-Channel Mixing."},{"cited_title":"Gen- erative adversarial nets,","cited_arxiv_id":null,"evidence_quote":"Defines the adversarial loss used in the overall objective."},{"cited_title":"International society of urological pathology (isup) grading of prostate cancer–an isup consensus on contemporary grading,","cited_arxiv_id":null,"evidence_quote":"Defines the ISUP grade-group mapping that determines the low-risk/high-risk labels."}],"review_version":1}