{"id":"e61d7b46-0fe3-435e-a31b-dce30d67a7b1","arxiv_id":"2502.06632","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"high","formal_verification":"none","parameter_count":2,"one_line_summary":"On 12 SPECT images, Prototypical Networks with a ResNet-18 backbone reached 93.33% validation accuracy for three tissue classes, while a 2D-adapted PRNet was evaluated only by its training reconstruction loss.","lead":"This paper adapts two existing deep learning methods, Prototypical Networks and PRNet, to classify and locate tissues in small SPECT heart scans. It reports 93.33% validation accuracy for classifying three tissue types, but the localization result is only a training loss with no independent evaluation.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"PRNet localization is supported only by a minimized training loss, with no validation or coordinate-error metric; the localization claim is the weakest load-bearing link.","rationale":"The reader's stated weakest_assumption focuses on possible leakage in the classification split—a real concern given the undocumented use of only 12 images. The reader's rationale also notes that PRNet's only reported result is a training loss, and that is the concern I find most load-bearing. I therefore partially agree: the classification split issue is secondary, while the localization evidence gap is decisive. The paper's own proof-of-concept framing lowers the stakes, but it does not supply the missing held-out evaluation. A minimized training loss is not evidence of generalization; PRNet could be memorizing patches or optimizing Lrec without learning useful spatial coordinates. Because anatomical localization is one of the two stated contributions, and because no validation metric for it is provided, the central claim is unsupported as written. This is not an internal inconsistency or a disagreement with consensus; it is an evidential gap that additional experiments could close. The rejection verdict stands, and no change to the reader's verdict is needed.","tokens_in":3489,"tokens_out":3761,"duration_ms":32444,"concrete_test":"Hold out one or more SPECT slices (ideally entire patients) from PRNet training. After training, sample random pairs of points on the held-out slices, predict their relative offset d'ji, and compare to the true offset dji using mean Euclidean distance. Also compute PSNR/SSIM on reconstructed held-out patches. If the coordinate error is not clearly below the distance between random point pairs, or if reconstruction quality is low, the anatomical localization claim is unsupported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's second central claim—anatomical localization—rests entirely on the PRNet training loss of 1.395 reported in Section III. The self-supervised loss in Eq. 3 is Lssl = Ldis + Lrec, and training explicitly minimizes this quantity, so a low value is expected by construction and cannot by itself demonstrate that predicted offsets d'ji correspond to true anatomical coordinates. The paper reports no separate values for Ldis and Lrec, no evaluation on held-out SPECT slices, and no quantitative localization metric such as mean Euclidean error between predicted and true landmark offsets. The qualitative reconstruction in Fig. 1 is a single example and is not accompanied by PSNR/SSIM or any held-out reconstruction error. The abstract's statement that PRNet 'accurately reconstructing patches and capturing spatial relationships' therefore overstates the available evidence. Additionally, the classification result relies on an undocumented split of only 12 images; if slices from the same patient appear in both training and validation, the 93.33% validation accuracy may reflect leakage. Both issues concern generalization, but the PRNet issue is more load-bearing because anatomical localization is presented as the second main contribution and no validation evidence for it exists at all.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript adapts two existing deep-learning methods to 2D SPECT images: a Prototypical Network with a pre-trained ResNet-18 backbone for few-shot classification of three tissue types (ventricles, myocardium, liver), and a 2D adaptation of the Propagation-Reconstruction Network (PRNet) for anatomical localization via self-supervised patch reconstruction and relative-offset prediction. Using a dataset of only 12 SPECT images cropped around the heart, the authors report 96.67% training and 93.33% validation classification accuracy, and a PRNet training loss of 1.395, which they interpret as evidence of accurate reconstruction and spatial-relationship learning. The paper concludes that these methods are promising for limited-data SPECT analysis.","tokens_in":3712,"tokens_out":3044,"duration_ms":26597,"significance":"If the claims were supported by rigorous evaluation, the paper would demonstrate a useful proof of concept for few-shot tissue classification and self-supervised landmark localization in SPECT, an area where labeled data are scarce. The choice of established methods (Prototypical Networks, PRNet) is reasonable, and the adaptation to 2D imaging is a sensible first step. However, the significance is currently undermined by the absence of any validation evidence for the localization claim, by an incompletely described and very small classification experiment, and by the lack of quantitative or comparative evaluation. The paper does not provide machine-checked proofs, reproducible code, or a public dataset; in its present form it is a brief technical report rather than a substantiated journal contribution.","major_comments":[{"comment":"The anatomical localization claim rests entirely on the reported training loss of 1.395 for the self-supervised loss Lssl of Eq. (3). This quantity is exactly what the training procedure minimizes, so a low value is expected by construction and cannot by itself demonstrate that the predicted relative offsets d'ji correspond to true anatomical coordinates. The paper reports no held-out evaluation, no separate values for Ldis and Lrec, and no quantitative localization metric such as mean Euclidean error between predicted and true landmark offsets. The qualitative example in Fig. 1 is a single case and is not accompanied by PSNR/SSIM or any held-out reconstruction error. Consequently, the abstract's statement that PRNet 'accurately reconstructing patches and capturing spatial relationships' is unsupported.","section":"Section III (PRNet results)"},{"comment":"The description of the train/validation split for the 12 SPECT images is missing. It is not stated whether the split is at the level of patients or individual slices, how many slices from each patient appear in each set, or how the tissue masks were generated. If slices from the same patient appear in both training and validation, the reported 93.33% validation accuracy could reflect data leakage rather than true generalization. The experimental design also lacks multiple seeds, error bars, and any comparison with a baseline classifier, so the reported accuracy is not a reliable estimate of few-shot performance.","section":"Section III (Prototypical Network)"},{"comment":"The claim that PRNet 'accurately reconstructs input patches' is supported only by the total training loss (1.395) and a single visual example. Since Lrec is a component of the minimized objective, the total loss does not separate reconstruction quality from distance-error quality, and no quantitative reconstruction metric (e.g., PSNR, SSIM) is given for any held-out patches. Without such evidence, the reconstruction aspect of the localization contribution cannot be assessed.","section":"Section III (PRNet reconstruction)"},{"comment":"The few-shot classification experiment uses only 10 episodes with 3 support and 6 query images per class. Standard few-shot evaluation protocols typically use hundreds or thousands of episodes to obtain stable accuracy estimates; with 10 episodes, the reported 96.67% training and 93.33% validation accuracies have very large variance and cannot be interpreted as robust performance. The absence of confidence intervals or seed variation makes the result essentially anecdotal.","section":"Section III (episodic evaluation)"}],"minor_comments":[{"comment":"The denominator of the softmax in Eq. (1) is typeset as a bare 'P' with an index; it should be an explicit summation over all classes k' to be mathematically readable.","section":"Equation (1)"},{"comment":"The text states that a 'pretrained ResNet-18 [8]' is used, but reference [8] is a few-shot classification paper (Chen et al.) and not the original ResNet reference; the authors should cite He et al. for the ResNet architecture.","section":"Section II.A, reference [8]"},{"comment":"The sentence reporting '96.67% accuracy and 0.486 loss on the training set' does not specify over which examples or episodes the training accuracy was computed; this should be clarified (e.g., training episodes, support set, or the entire training split).","section":"Section III (classification results)"},{"comment":"The phrase 'high training and validation accuracies' in the Conclusion is an overstatement given the very small evaluation; the wording should be tempered to 'reported accuracies' unless additional experiments are provided.","section":"Abstract and Conclusion"}],"recommendation":"reject","confidential_remarks":"The manuscript is extremely thin for a journal submission: the central localization claim has no validation at all, and the classification experiment lacks basic experimental hygiene (split description, error bars, baselines). These are not local presentation issues; the evidence presented cannot support the stated conclusions. Even with a major revision, the authors would need to add substantial new experiments (e.g., a proper cross-validation with patient-level splits, multiple seeds, a coordinate-error metric for PRNet on held-out slices) that go beyond the current proof-of-concept scope. I do not see a path to acceptance without such new evidence, hence my recommendation to reject rather than request major revision. I would encourage the authors to revisit the evaluation design and consider making the dataset and code available to strengthen any resubmission."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: a four-page proof of concept that applies two established few-shot methods to 2D SPECT images. The genuinely new thing is the application domain. The classification half is a plausible small demonstration; the localization half is not supported by the reported evidence.\n\nWhat the paper does well: it is clearly written, the adaptation of PRNet from 3D to 2D is described plainly, and the authors correctly credit the source methods. They also frame the work as a proof of concept, which is the right level of ambition for a 12-image dataset.\n\nThe soft spots are real. Classification: 12 images total, no description of how the train/validation split was made, no error bars, no seed variation, no baseline comparison. If slices from the same patient end up in both sets, the 93.33% validation accuracy is compromised. The PRNet problem is more load-bearing. The only quantitative result is a training loss of 1.395, which is the exact objective minimized during training. That number cannot establish that the network learned anatomical coordinates. The paper reports no separate distance and reconstruction losses, no held-out reconstruction error, and no coordinate-error metric. The qualitative figure is one example. So the abstract's claim that PRNet 'accurately reconstructing patches and capturing spatial relationships' goes beyond what is shown.\n\nThe stress-test note is right on target. This is a training loss presented as evidence of generalization, and the localization claim is the paper's second main contribution. I don't read the paper as deceptive—it is transparent about the setup—but the evaluation gap is fundamental.\n\nWho is this for? A reader curious whether Prototypical Networks and PRNet can be bolted onto SPECT data at all. It could be a useful reading-group example of what counts as evidence for localization. It is not ready for journal peer review.\n\nIf I were the editor, I would desk reject it, with an invitation to resubmit after adding a proper split description, baselines, error bars, and a held-out localization metric. The kernel of a proof of concept is here, but the evidence as written is not enough.","headline":"A clearly written proof of concept for few-shot classification in SPECT, but the PRNet localization claim rests entirely on a minimized training loss and needs a real held-out evaluation.","tokens_in":4254,"tokens_out":3771,"would_cite":false,"duration_ms":30405,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":false},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A Prototypical Network with a pretrained ResNet-18 backbone classifies cardiac tissue masks in 2D SPECT slices with 93.33% validation accuracy from only 12 images, while a 2D-adapted PRNet learns anatomical landmark positions by…","keywords":["few-shot learning","SPECT imaging","Prototypical Networks","PRNet","anatomical localization","tissue classification","self-supervised reconstruction","medical image analysis"],"falsifier":"Run the same Prototypical Network with leave-one-patient-out cross-validation on the 12 images: if validation accuracy falls to near the 33% chance level, the original 93.33% figure was inflated by train/validation leakage; if it remains near 90%, the few-shot result generalizes across patients.","tokens_in":3287,"feed_emoji":"🩺","tokens_out":9048,"duration_ms":69350,"temperature":0.7,"pith_summary":"Single Photon Emission Computed Tomography (SPECT) images are usually scarce, and this paper asks whether two few-shot deep learning methods can still classify tissues and locate anatomy with only 12 images cropped around the heart. It adapts a Prototypical Network to classify myocardium, ventricle, and liver tissue masks, reporting 96.67% training accuracy and 93.33% validation accuracy. It also converts the Propagation-Reconstruction Network (PRNet) from 3D to 2D and uses self-supervised patch reconstruction to learn spatial relationships between anatomical landmarks, reaching a training loss of 1.395. The paper presents these numbers as a proof of concept that tissue classification and anatomical localization are feasible under severe data scarcity.","feed_headline":"SPECT tissue classification hits 93.33% from 12 images","feed_subtitle":"Prototypical Networks classify heart tissues; a 2D PRNet reconstructs patches to localize landmarks.","key_machinery":"Two existing architectures carry the argument. The Prototypical Network represents each class by the mean embedding of its support examples and classifies a query by the softmax of negative Euclidean distances to those prototypes, which lets a three-class tissue classifier be trained from a handful of labeled slices. The Propagation-Reconstruction Network (PRNet), adapted from 3D to 2D layers, samples two random points in an image, crops fixed-size patches around them, and learns to predict their relative anatomical offset while reconstructing the patches; its self-supervised loss is $L_{\\text{ssl}} = L_{\\text{dis}} + L_{\\text{rec}}$, where the first term penalizes offset error and the second penalizes reconstruction error. Skip connections in the encoder-decoder help the reconstruction preserve spatial detail, and the pretrained ResNet-18 supplies the embedding backbone for the classifier.","core_discovery":"On a dataset of 12 SPECT images cropped around the heart, the paper claims that a Prototypical Network with a pretrained ResNet-18 backbone separates tissue masks into ventricles, myocardium, and liver with 96.67% training accuracy and 93.33% validation accuracy after 10 episodes of 3 support and 6 query images per class. For localization, the paper converts the 3D Propagation-Reconstruction Network (PRNet) to 2D convolutions and trains it with self-supervised random double-cropping; the model predicts relative offsets between pairs of patches and reconstructs the patches through an encoder-decoder with skip connections, reaching a training loss of 1.395. The authors take this reconstruction behavior as evidence that the network captures spatial relationships among anatomical landmarks. The intended contribution is a proof of concept that both classification and anatomical localization are attainable in SPECT without large labeled datasets.","pith_inferences":["A stronger test of the classification claim would be a patient-disjoint validation split, since the paper does not report whether slices from the same patient appear in both training and validation.","A shuffled-offset control for PRNet, where patch positions are randomized, would reveal how much of the 1.395 loss reflects genuine anatomy rather than low-level image statistics.","The same 2D adaptation recipe could transfer to other low-count nuclear medicine studies, such as brain or bone SPECT, where labeled data are equally scarce.","Combining the classifier's tissue labels with the localizer's landmark coordinates could generate pseudo-masks that reduce manual annotation for future segmentation models."],"forward_implications":["The episodic Prototypical Network setup can separate ventricle, myocardium, and liver tissue masks using only three support images per class.","The 2D-adapted PRNet can predict relative positions of anatomical landmarks without dense annotations, using only self-supervised patch pairs.","The combination of a few-shot classifier and a self-supervised localizer provides a building block for SPECT segmentation pipelines that do not depend on large labeled atlases.","The reported 1.395 training loss shows that the reconstruction objective is learnable on the available SPECT data, supporting further training on larger cohorts."],"supporting_citations":[{"why":"Introduces the Prototypical Network objective and episodic training scheme that the paper adapts for tissue classification.","marker":"[3]"},{"why":"Supplies the PRNet encoder-decoder architecture and the self-supervised propagation-reconstruction loss that the paper converts from 3D to 2D.","marker":"[4]"},{"why":"Cited as the source of the pretrained ResNet-18 backbone used as the embedding function in the classifier.","marker":"[8]"}],"fun_headline_variants":["Few-shot SPECT: 93% tissue classification, plus localization","SPECT few-shot: 93% tissue accuracy with 12 images","12 SPECT images: 93% classification, landmark localization","Prototypical nets: SPECT tissues at 93% from few shots"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The reported validation accuracy depends on the undocumented assumptions that the 12 images split into training and validation sets are independent and that their tissue masks are correct, with no slice from the same patient appearing in both sets.","fun_headline_variants_meta":{"raw":{"variants":["Few-shot SPECT: 93% tissue classification, plus localization","SPECT few-shot: 93% tissue accuracy with 12 images","12 SPECT images: 93% classification, landmark localization","Prototypical nets: SPECT tissues at 93% from few shots"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000614,"raw_usage":{"total_tokens":2842,"prompt_tokens":922,"completion_tokens":1920,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":538,"completion_tokens_details":{"reasoning_tokens":1843}},"tokens_in":538,"tokens_out":1920,"duration_ms":13794,"temperature":1.0,"reasoning_tokens":1843,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-08T14:50:37.549757+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the same Prototypical Network with leave-one-patient-out cross-validation on the 12 images: if validation accuracy falls to near the 33% chance level, the original 93.33% figure was inflated by train/validation leakage; if it remains near 90%, the few-shot result generalizes across patients.","supporting_citations":[{"cited_title":"One-shot weakly-supervised segmentation in 3d medical images,","cited_arxiv_id":null,"evidence_quote":"Supplies the PRNet encoder-decoder architecture and the self-supervised propagation-reconstruction loss that the paper converts from 3D to 2D."}],"review_version":1}