{"id":"e46642fd-b31f-4162-b5ff-7da88a46b9a7","arxiv_id":"2505.17921","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"Prototypical Networks with ImageNet-pretrained ResNet backbones classify kidney stone subtypes from endoscopic patches as accurately with 25% of the training data as a conventional ResNet trained on the full dataset.","lead":"This paper evaluates Prototypical Networks, a few-shot learning method, for classifying kidney stone types from endoscopic images, and reports that it matches or beats a conventional deep learning model even when trained on a quarter of the data. A generalist might read it as a case study in building medical image classifiers when labeled images are scarce.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The ProtoNet-vs-traditional comparison is confounded because the traditional ResNet-34 baseline is unspecified, especially whether it uses ImageNet initialization; the headline gap may be a transfer-learning artifact.","rationale":"The reader's weakest assumption is the same as the most load-bearing concern: the traditional DL baseline is under-specified, so the headline comparison could be an artifact of undertraining or missing transfer learning. The abstract's claim that ProtoNet with 25% of data beats a full-data traditional model is supported only by Table III, and the baseline column has no implementation details in Section III.B. The paper does give ProtoNet implementation details in Section II.D.4, but no analogous paragraph appears for the non-FSL ResNet-34. Since the ProtoNet method explicitly uses ImageNet-pretrained backbones, failing to state whether the baseline also uses pretraining introduces a direct confound: any competitive advantage could be due to transfer learning rather than few-shot learning. The proposed test would settle this by rerunning the baseline under standard fine-tuning and under the exact ProtoNet recipe. The patch-level split by source image is a strength and mitigates leakage concerns, so the issue is not with dataset construction but with the fairness of the method comparison. The citation mismatch between [9] and [12] for ProtoNet is minor and does not change the verdict. Overall, the conditional acceptance recommendation remains appropriate pending the baseline reimplementation and details.","tokens_in":8298,"tokens_out":4112,"duration_ms":37442,"concrete_test":"Re-run Table III's Traditional DL row with a standard fine-tuning recipe on the same 25/50/75/100 splits: ImageNet-pretrained ResNet-34, SGD lr 1e-3 with cosine decay, batch 32, 50 epochs, random crop/flip augmentation, 5 seeds, reporting mean and standard deviation. Compare the 100%-data baseline to ProtoNet's best 25%-data entries (SUR 88.77, SEC 95.22, MIX 90.17). If the tuned baseline meets or exceeds these numbers, the headline claim fails; if it remains below, the few-shot advantage is supported. As a minimal control, also run the baseline with the exact ProtoNet training recipe (Adam lr 1e-4, 1000 iterations) to isolate the effect of pretraining from the choice of method.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Central claim depends on Table III: ProtoNet at 25% data (SUR 88.77, SEC 95.22, MIX 90.17) beats traditional ResNet-34 at 100% data (85.17, 91.75, 88.42). Section III.B gives no implementation details for the traditional model: optimizer, epochs/iterations, batch size, augmentation, and crucially whether it is initialized from ImageNet weights. Section II.C states ProtoNet backbones are ImageNet-pretrained; if the non-FSL baseline is trained from scratch, the comparison conflates transfer learning with few-shot learning. On 500-2000 patches, a randomly initialized ResNet-34 under-trained for a fixed budget can plausibly sit near 77-85%, making the reported ProtoNet advantage an artifact. The absence of variance or standard deviations for the baseline rows in Table III further means the 3-5 point gaps may be within noise. A secondary issue is that ProtoNet is cited to [9], a survey, rather than the actual source [12]. The load-bearing fix is to specify and re-run the baseline fairly.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This manuscript evaluates few-shot learning methods, specifically Prototypical Networks, for kidney stone type recognition in ex vivo ureteroscopy images. The authors use a dataset of 409 endoscopic images from which 6,000 patches per view (surface, section, and mixed) are extracted, split by image into 80% training/validation and 20% test. Prototypical Networks are tested with ResNet-18/34/50 backbones, multiple shots (5, 10, 15, 20), and training-data fractions (100%, 75%, 50%, 25%). The central claim, stated in the abstract, is that Prototypical Networks trained on up to 25% of the data can match or exceed a traditional deep learning model trained on the full dataset. The experimental setup is mostly clear and all results are tabulated, but the key comparisons in Table III rely on an under-specified baseline and lack variance information.","tokens_in":8478,"tokens_out":3730,"duration_ms":37285,"significance":"If the comparison were valid, the result would be practically useful: for a domain where annotated kidney stone images are scarce and class frequencies are imbalanced, demonstrating that a few-shot method can match or beat a traditional model with a quarter of the data would support deployment in low-resource settings. The paper's strengths are its multi-view evaluation (SUR, SEC, MIX), the range of backbones and data fractions, the image-level split that avoids patch leakage, and the transparent reporting of accuracy/precision/recall/F1 in Table II. However, the headline claim rests on an unspecified traditional baseline and on selected ProtoNet configurations, so the current evidence is suggestive rather than conclusive.","major_comments":[{"comment":"The 'Traditional DL model' baseline is severely under-specified. The text only states that a ResNet-34 was implemented without FSL, but gives no optimizer, learning rate, number of epochs/iterations, batch size, data augmentation, or initialization scheme. Section II.C explicitly says the ProtoNet backbones are initialized with ImageNet-pretrained weights; if the traditional ResNet-34 is trained from scratch on 500-2000 patches, the comparison conflates transfer learning with few-shot learning and the reported ProtoNet advantage could be an artifact of the baseline's undertraining. The authors must specify the baseline training details and, ideally, retrain the baseline with ImageNet initialization and a comparable tuned budget.","section":"Section III.B, Table III"},{"comment":"Table III reports single accuracy values with no standard deviations, confidence intervals, or number of seeds. The mean +/- std in Table II is computed across 16 heterogeneous configurations (combinations of shots and data percentages), so it cannot serve as an uncertainty estimate for the specific rows used in the headline comparison. The gaps used to support the central claim (e.g., SUR 88.77 vs. 85.17, SEC 95.22 vs. 90.00) may be within run-to-run noise. Repeated runs with different seeds and variance reporting for the key ProtoNet and baseline rows are needed.","section":"Table III and Section III.A"},{"comment":"The claim that 'the 6-ways-10-shots configuration consistently achieves the best performance across all views and various data percentages' is contradicted by Table III itself. For example, at 25% data on SEC, 6-15 achieves 95.22% while 6-10 achieves 89.92%; at 50% data on SUR, 6-15 achieves 88.37% while 6-10 achieves 83.77%; at 50% MIX, 6-15 achieves 90.52% while 6-10 achieves 87.42%. The text should either be revised to state that the best configuration varies, or the selection procedure for choosing the 'best' ProtoNet configuration per view and data fraction should be described explicitly, since this selection affects the strength of the central claim.","section":"Section III.A, paragraph after Table II"}],"minor_comments":[{"comment":"The sentence 'the dataset size was gradually reduced by randomly selecting subsets in 25' is incomplete and should be finished, presumably as 'in 25% increments' or similar.","section":"Section II.D.1"},{"comment":"Prototypical Networks are attributed to reference [9], which is a survey, while Section II.C correctly cites the original source [12]. The citation in Section II.B should be [12] or should include [12] alongside the survey.","section":"Sections II.B and II.C"},{"comment":"The caption states that each metric is the average over experiments with varying shots and data proportions, but the text says 'sixteen models were generated.' It should be clarified whether the standard deviation is across independent training runs or across the heterogeneous configurations; if the latter, the current notation is misleading.","section":"Table II caption and text"},{"comment":"The patch extraction procedure is not described in enough detail for reproducibility: it is unclear how many patches are extracted per image, whether patches overlap, and whether the standardization mean/standard deviation is computed only on the training split.","section":"Section II.A"},{"comment":"Figure 3 is presented as a qualitative comparison, but the subfigures show only example outputs, not a systematic visualization of the quantitative difference. The caption should make clear that the figure is illustrative, and the quantitative evidence should come from the tables.","section":"Figure 3"}],"recommendation":"major_revision","confidential_remarks":"The paper fits the scope of a medical-imaging evaluation venue and the empirical question is relevant. However, the headline comparison is not yet convincing because the baseline implementation is underspecified and the key table lacks variance. These issues are fixable within the manuscript's scope by adding baseline details, rerunning with a comparable pretrained/tuned baseline, and reporting repeated-run statistics. The citation inconsistency between [9] and [12] also suggests a careful reference check before resubmission."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is a useful empirical study, but the core comparison is not yet trustworthy because the traditional baseline is described in one sentence. The paper applies Prototypical Networks with ImageNet-pretrained ResNet backbones to ex vivo kidney stone patch classification, with ablations over data fraction (100-25%), shots (5-20), and backbone depth (ResNet-18/34/50). That specific systematic comparison is new for this niche medical imaging domain, and the data split by image (not patch) is a sound choice. The reported numbers do support the abstract's claim in several configurations: e.g., ProtoNet at 25% data reaches 88.77% (SUR) and 95.22% (SEC) versus 85.17% and 91.75% for the traditional model at 100% data.\n\nThe soft spot is exactly where the stress test points. Section III.B gives no implementation details for the 'Traditional DL model': no optimizer, no epochs, no batch size, no augmentation, and no statement about whether the ResNet-34 was initialized from ImageNet weights or trained from scratch. The ProtoNet backbones are explicitly ImageNet-pretrained (Section II.C). If the baseline trains from scratch or under a weak budget, the comparison conflates transfer learning with few-shot learning, and a randomly initialized ResNet-34 on 500-2000 patches could plausibly land at 77-85%, right where the reported baseline numbers sit. So the headline gap may be an artifact. This is not a manufactured flaw; the missing baseline specification is load-bearing.\n\nTwo smaller issues: Table III reports point estimates without variance or confidence intervals, so the 3-5 point gaps could be noise; the paper does give mean±std for the aggregate Table II, but not for the key Table III rows. Also, ProtoNet is cited to a survey [9] in Section II.B while the original Snell et al. [12] appears later, which looks like a citation slip. And there is a broken sentence in Section II.D.1 ('randomly selecting subsets in 25') that should be completed.\n\nWho this is for: researchers working on medical image classification with limited data, especially kidney stone recognition. If the authors add a fair, fully specified baseline and report variance, the result would be worth taking seriously. As submitted, I would not trust the headline until the baseline is re-run and reported.\n\nRecommendation: send to peer review; this is an empirical evaluation with a testable claim, and the flaws are addressable with revision rather than desk-reject material.","headline":"A useful empirical evaluation of ProtoNet for kidney stone classification, but the headline comparison is weakened by an under-specified traditional baseline and missing variance.","tokens_in":8980,"tokens_out":2779,"would_cite":true,"duration_ms":24394,"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":"This paper claims that Prototypical Networks, trained on only 25% of available ureteroscopic kidney-stone patches, match or exceed a traditional ResNet-34 trained on the full dataset.","keywords":["kidney stone classification","few-shot learning","prototypical networks","ureteroscopy","endoscopic imaging","ResNet","data scarcity","medical image classification"],"falsifier":"Train a traditional ResNet-34 from the same codebase on the same SUR, SEC, and MIX splits with the same optimizer (Adam, learning rate 0.0001), iteration count, and patch preprocessing used for ProtoNet, sweeping epochs and augmentation. If the tuned baseline reaches or exceeds ProtoNet's accuracy at 25%, 50%, 75%, and 100% data fractions, the central claim is refuted.","tokens_in":8135,"feed_emoji":"🩺","tokens_out":5859,"duration_ms":55440,"temperature":0.7,"pith_summary":"This paper asks whether few-shot learning can classify six kidney stone types in ureteroscopic images when labeled examples are scarce. The central claim is that Prototypical Networks, given only 25% of the training patches, perform as well as or better than a traditional ResNet-34 deep learning model trained on the complete dataset. The authors report 88.77% versus 85.17% accuracy on surface-view images and 95.22% versus 90.00% on section-view images at the 25% data fraction. The motivation is medical: standard stone typing takes weeks, expert visual recognition is rare, and hospital datasets are small and imbalanced.","feed_headline":"Few-shot model beats full-data network for kidney stones","feed_subtitle":"Prototypical Networks using 25% of training patches match or beat a standard ResNet-34 trained on all data across all three views.","key_machinery":"The load-bearing mechanism is Prototypical Networks (ProtoNet). For each class, a prototype is the mean embedding of that class's support examples after a shared ResNet encoder; every query image is classified by its Euclidean distance to the nearest prototype. Training is episodic: the model repeatedly sees simulated few-shot tasks, which prepares it for genuine data scarcity. A one-step transfer step initializes the ResNet backbone with ImageNet weights, which the authors argue is crucial when starting from limited target-domain data. This machinery is what lets a 6-way, 10- or 15-shot configuration trained on 25% of the patches stay competitive with a full-data classifier.","core_discovery":"On the paper's own terms, the discovery is that Prototypical Networks with an ImageNet-pretrained ResNet-34 backbone maintain high discriminative performance even when the training set is cut to a quarter. Across surface (SUR), section (SEC), and mixed (MIX) views, the best Prototypical Network configuration (6-way, 10 or 15 shots) produces accuracies comparable to or higher than those of a traditional ResNet-34 trained on the same data fractions, and in several cells the few-shot model trained on 25% of the data matches or exceeds the traditional model trained on 100%. For example, on the SEC view, ProtoNet at 25% data reaches 95.22% accuracy, versus 90.00% for the traditional model at the same fraction. The authors interpret this as evidence that the episodic, prototype-based training procedure makes more efficient use of scarce endoscopic images.","pith_inferences":["A fairer comparison would tune the traditional ResNet-34 with the same optimizer schedule, epochs, and augmentation as ProtoNet; the paper does not report these details for the baseline, so part of the gap may be baseline under-training rather than few-shot learning itself.","The evaluation uses balanced 6-way episodes, but real kidney stone classes are highly imbalanced; testing ProtoNet under class-imbalanced episode sampling would show whether the advantage survives natural prevalence.","Since training and testing use 256x256 patches, a full-image variant could change the ranking; the authors themselves list full-image training as future work.","An immediate extension is to measure calibration and error patterns per stone type, since equal accuracy across classes can hide confusion between visually similar subtypes such as WW and WD."],"forward_implications":["A hospital with only a small archive of labeled endoscopic images could still train a useful six-class stone classifier, because 25% of the current patch set suffices.","Models built this way stay accurate across all three views (surface, section, mixed), so the approach does not depend on image engineers having full cross-section access.","The configuration search gives practitioners a concrete default: ResNet-34 with 6-way, 10- or 15-shot episodic training.","Because the traditional baseline at 25% data drops sharply (e.g., 77% on SUR) while ProtoNet stays near 88%, the method is particularly valuable at exactly the data-scarce operating point."],"supporting_citations":[{"why":"Supplies the ex vivo endoscopic kidney stone dataset (409 images, 6 types, SUR/SEC views) and the phantom acquisition setup used in all experiments.","marker":"[5]"},{"why":"Provides the prior deep learning kidney stone classification work from which the patch extraction and standardization protocol are taken, and the traditional DL baseline context.","marker":"[11]"},{"why":"Introduces Prototypical Networks, the few-shot method that is the paper's central object.","marker":"[12]"},{"why":"Establishes endoscopic stone recognition as a clinically relevant task and the visual features the DL models aim to automate.","marker":"[6]"},{"why":"Supplies the general few-shot image classification context that motivates the episodic evaluation.","marker":"[9]"}],"fun_headline_variants":["Few-shot kidney stone AI matches full-data model with quarter data","Prototypical nets excel at kidney stone ID with scarce images","Kidney stone recognition via few-shot learning needs only 25% data","Quarter data, full performance: few-shot kidney stone classifier","Few-shot method cuts kidney stone training data to 25%"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The headline comparison assumes the traditional ResNet-34 is a strong, fairly tuned baseline; its training details are not reported, so a weaker baseline would inflate the apparent few-shot advantage.","fun_headline_variants_meta":{"raw":{"variants":["Few-shot kidney stone AI matches full-data model with quarter data","Prototypical nets excel at kidney stone ID with scarce images","Kidney stone recognition via few-shot learning needs only 25% data","Quarter data, full performance: few-shot kidney stone classifier","Few-shot method cuts kidney stone training data to 25%"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000655,"raw_usage":{"total_tokens":2983,"prompt_tokens":912,"completion_tokens":2071,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":528,"completion_tokens_details":{"reasoning_tokens":1984}},"tokens_in":528,"tokens_out":2071,"duration_ms":15014,"temperature":1.0,"reasoning_tokens":1984,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T14:36:47.915693+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train a traditional ResNet-34 from the same codebase on the same SUR, SEC, and MIX splits with the same optimizer (Adam, learning rate 0.0001), iteration count, and patch preprocessing used for ProtoNet, sweeping epochs and augmentation. If the tuned baseline reaches or exceeds ProtoNet's accuracy at 25%, 50%, 75%, and 100% data fractions, the central claim is refuted.","supporting_citations":[{"cited_title":"Evaluation and understanding of automated urinary stone recognition methods","cited_arxiv_id":null,"evidence_quote":"Supplies the ex vivo endoscopic kidney stone dataset (409 images, 6 types, SUR/SEC views) and the phantom acquisition setup used in all experiments."},{"cited_title":"On the in vivo recognition of kidney stones using machine learning","cited_arxiv_id":null,"evidence_quote":"Provides the prior deep learning kidney stone classification work from which the patch extraction and standardization protocol are taken, and the traditional DL baseline context."},{"cited_title":"Towards automatic recognition of pure and mixed stones using intra-operative endoscopic digital images","cited_arxiv_id":null,"evidence_quote":"Establishes endoscopic stone recognition as a clinically relevant task and the visual features the DL models aim to automate."},{"cited_title":"Few-shot image classification: Current status and research trends","cited_arxiv_id":null,"evidence_quote":"Supplies the general few-shot image classification context that motivates the episodic evaluation."}],"review_version":1}