{"id":"775ecb3a-ebb1-48f7-aa20-c28ad5a2889a","arxiv_id":"2506.20267","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A surface vision transformer diagnoses dementia by weighing how each cortical patch resembles a corresponding prototype from real training brains, with competitive accuracy and clinically aligned explanations.","lead":"The paper introduces X-SiT, a surface vision transformer that diagnoses dementia by comparing cortical surface patches to learned prototypes taken from real training brains. It aims to make predictions understandable for clinicians while remaining competitive with black-box models.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Interpretability claim rests on unvalidated latent-space semantics: the cosine similarity in the trained encoder is not shown to track raw cortical-feature similarity, so the prototypes may be illustrative rather than causal.","rationale":"The reader's weakest assumption focused on template-registration correspondence. That is a reasonable concern, but FreeSurfer registration to fsaverage6 is standard and the paper's fixed-vertex assumption is a conventional approximation; even if registration were perfect, the central interpretability claim could still fail. The more load-bearing issue is the unvalidated semantic grounding of the latent-space comparisons. The architecture is transparent in its computation, but transparency is not the same as interpretability: if the learned latent space does not align with human-perceivable patch similarity, then the prototypes, though visually plausible, may not explain the model's decisions. The paper provides no faithfulness metric, no correlation analysis between latent and input similarity, and no intervention experiment; the only support is qualitative inspection of group-level mean activations and a few individual examples. This gap is particularly important because the model is trained end-to-end with only a classification loss, leaving the encoder free to exploit latent structure unrelated to the displayed features. The abstract's 'state-of-the-art' overclaim is also confirmed by Table 2, but this is a presentation issue rather than a fundamental flaw. Given the architectural contribution is coherent and the empirical results are competitive, a conditional verdict is appropriate: the authors should validate the faithfulness of their prototype explanations before the interpretability claim can be accepted. Hence no change to the reader's verdict.","tokens_in":8001,"tokens_out":5217,"duration_ms":57226,"concrete_test":"Compute, for each test subject and each prototype i, the rank correlation between the decision-relevant latent similarity cos(x_i, ξ_i) and the similarity of the raw input patch features (e.g., Pearson or Spearman correlation of the thickness/sulcal-depth/curvature vectors after standard normalization) across all test patches. If the average correlation is low (e.g., Spearman ρ < 0.3), the latent similarity is not reflecting the visualized patch features, and the interpretability claim fails. Alternatively, perform an intervention: replace patch i in a test input with the prototype's raw feature patch (or an average CN patch at the same location) and measure whether the change in P(c|x) tracks the change in raw-feature similarity as implied by the explanation; a mismatch indicates the prototypes are not causally faithful.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that X-SiT is inherently interpretable via case-based reasoning with prototypes. The decision rule in Eq. (2) is a weighted sum of cosine similarities between encoded patches and prototypes, but both are in a latent space produced by the SiT encoder. Section 2.3 states the SiT is trained with binary cross-entropy only, and the text says it maps input surfaces to 'an optimal latent space for comparison with the prototypes' — optimal for classification, not for semantic correspondence to the raw interpretable features (cortical thickness, sulcal depth, curvature). There is no regularization (e.g., a reconstruction or clustering loss, as in ProtoPNet) enforcing that high latent cosine similarity implies high similarity of the visualized input patches. Consequently, the encoder could map visually distinct patches to nearby latent vectors if that improves accuracy, making the prototype comparison an opaque computation whose apparent meaning in Figure 4 is not guaranteed to reflect the model's actual reasoning. The interpretability evidence in Section 3.4 is entirely qualitative: mean activation maps and a few example prototypes. The model's prediction is based on latent vectors, while the prototype visualizations show raw features; no experiment connects the two. Thus the paper's primary contribution — inherent interpretability — is not established. A secondary issue is the abstract's 'state-of-the-art performance': Table 2 shows X-SiT (80.2 AD, 79.6 FTD) is slightly below SiT (80.5, 80.6), and the text later concedes 'competitive performance.' While the registration-correspondence assumption noted by the reader is relevant, it is a standard FreeSurfer template step; the lack of faithfulness validation in the latent space is more directly load-bearing for the interpretability claim.","agreement_with_reader":"disagree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript introduces X-SiT, an interpretable transformer for cortical surface classification. A Surface Vision Transformer (SiT) encoder produces patch-wise latent embeddings, and a prototypical surface patch (PSP) decoder computes the class probability as a weighted sum of cosine similarities between encoded patches and learned prototypes that are kept in spatial correspondence. Prototypes are periodically replaced with the most similar patches from training data, so the model is claimed to use case-based reasoning on interpretable cortical features. The method is evaluated for Alzheimer's disease and frontotemporal dementia diagnosis on ADNI and NIFD data, with balanced accuracy comparable to prior non-interpretable models, and qualitative group-level and individual-level visualizations are provided as interpretability evidence.","tokens_in":8291,"tokens_out":5550,"duration_ms":59969,"significance":"The architectural idea is timely and potentially valuable: extending prototype-based interpretability to cortical surface data with template correspondence is a natural and clinically relevant direction, and the sparse weighted decision rule is transparent in structure. The authors report competitive accuracy on two clinical tasks using public datasets, and the prototype-stability overlap statistic is a useful addition. However, the paper's core contribution, \"inherent interpretability,\" is not yet established: the visualizations and evaluations do not demonstrate that the latent-space comparisons underlying the decision rule reflect the raw interpretable features shown to the user. The performance claim in the abstract is also stronger than the test results support. Provided these points are addressed with additional analysis, the method would be a solid contribution to the interpretable medical imaging literature.","major_comments":[{"comment":"The decision rule is defined entirely on latent embeddings xi and ξi, while the interpretability visualizations (Figure 4) show raw cortical thickness, sulcal depth, and curvature. The encoder is trained with binary cross-entropy only, and there is no loss term (e.g., reconstruction, clustering, or similarity-alignment) that forces high cosine similarity in latent space to correspond to high similarity of the visualized raw features. Therefore, the prototypes may be illustrative rather than faithful: two raw patches that appear very different could be close in latent space if that helps classification. To support the central interpretability claim, the authors must connect latent-space similarity to raw-feature similarity, for example by measuring agreement between latent nearest neighbors and raw-feature nearest neighbors on held-out patches, or by adding an explicit alignment or clustering regularizer. Without such evidence, \"inherently interpretable\" is not established.","section":"Section 2.2, Eqs. (1)-(2); Section 2.3; Figure 4"},{"comment":"The interpretability evaluation is exclusively qualitative. Figure 3 shows mean activation maps and two individual examples, and Figure 4 shows selected prototypes. There is no user study, no quantitative explanation metric (e.g., fidelity, consistency, or agreement with expert-defined regions), and no statistical test that the prototypes align with known disease patterns beyond visual inspection. The claim that prototypes \"reveal classification errors\" rests on anecdotal cases. As a result, the paper does not demonstrate that X-SiT explanations are more faithful or more useful than the attention weights or saliency maps it criticizes in the introduction. The authors should add a quantitative evaluation of explanation quality, or at minimum an inter-rater or expert-agreement study.","section":"Section 3.4"},{"comment":"The abstract's claim of \"state-of-the-art performance\" is not supported by Table 2. On both test sets, the non-interpretable SiT achieves equal or higher balanced accuracy (AD: 80.5±0.5 vs 80.2±2.0; FTD: 80.6±1.8 vs 79.6±1.6), and X-SiT is second to SiT in both cases. The differences are within one standard deviation, so a fair statement is \"competitive\" or \"matches accuracy,\" which is the phrasing used in the introduction and conclusion. The abstract should be revised to avoid an unsupported claim. Additionally, in the FTD task, cognitively normal subjects are pooled from ADNI and NIFD while patients come only from NIFD, so scanner and protocol differences are a potential confound that should be discussed or mitigated.","section":"Abstract; Section 3.2; Table 2"},{"comment":"The interpretability mechanism relies on patch index i denoting the same anatomical region for every subject and every prototype, stated as an assumption in Section 2.1. The paper does not verify this correspondence after FreeSurfer registration to the FsAverage template. Residual registration error or folding variability could make \"corresponding prototypes\" and the group-level average maps in Figure 3 anatomically misaligned, which would undermine the clinical meaning of the explanations. A demonstration of registration quality (e.g., overlap of cortical labels on a validation population, or within-subject repeat-scan correspondence) would strengthen the central claim. If correspondence is not reliable, the cosine similarities and the resulting explanations are not anatomically meaningful.","section":"Section 2.1; Section 2.2"}],"minor_comments":[{"comment":"The naming is inconsistent: the related work uses \"ViT-Net\" while the adapted surface model is called \"SiT-NeT\" in Table 2 and in the text; please use a single consistent name.","section":"Related work and Table 2"},{"comment":"There is a typo in the first sentence of Section 3.2: \"T able 1\" should be \"Table 1.\"","section":"Section 3.2"},{"comment":"The prototype robustness overlap (76.3% and 71.7%) is reported without a definition of how the overlap is computed across random initializations; please specify the metric.","section":"Section 3.4"},{"comment":"The yellow ellipses indicating hallmark regions are described in the caption but are difficult to identify in a print version of the figure; consider adding explicit region labels to the figure itself.","section":"Figure 3"},{"comment":"The statement that the temporal lobe received comparatively low weight in the FTD experiments is somewhat confusing in light of the preceding reference to frontal and temporal regions as FTD hallmarks; please clarify whether the model's focus is expected to be incomplete.","section":"Section 3.4"}],"recommendation":"major_revision","confidential_remarks":"The paper's central claim is \"inherently interpretable\" via case-based reasoning, but the current evidence does not close the gap between the latent-space computation and the raw-feature visualizations. This is fixable with additional experiments or a more modest claim, so I recommend major revision rather than rejection. Please also ensure the abstract's performance claim is aligned with the presented results, and consider whether the \"first inherently interpretable neural network\" wording is proportionate given the existing literature on interpretable surface-based methods."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The core idea is legitimate and worth a look: anchoring each patch's comparison to a prototype at the exact same patch index on a registered cortical surface. That is the right design for anatomy, unlike global prototype search in ProtoPNet or PIP-Net. The paper makes the architectural case cleanly, and the experiments are run honestly: five seeds, multiple baselines, means and standard deviations. X-SiT stays within noise of SiT on both tasks and beats the adapted ViT-NeT, so the decoder is not an accuracy liability.\n\nThe soft spots are in proportion. First, the abstract says 'state-of-the-art performance' while Table 2 shows X-SiT slightly below SiT. The text later says 'competitive'—the abstract should say that. Second, the FTD experiment mixes CN subjects from ADNI and NIFD while all FTD patients come from NIFD. That is a site confound; the model may be picking up scanner differences. It needs a site-harmonization step or at least an analysis of site effects. Third, the interpretability evidence is qualitative, and more importantly, nothing links the latent-space cosine similarity to the raw-feature visualizations. The encoder is trained only with binary cross-entropy, so the prototypes could be similar in latent space for reasons invisible in the rendered thickness maps. A clustering or reconstruction regularizer, or a faithfulness experiment, would make 'inherently interpretable' credible. The paper does none of this. The registration assumption is less of a worry; FreeSurfer registration to FsAverage is standard, so the reader's strongest concern is not the load-bearing one.\n\nNone of this sinks the idea. It is a smart adaptation, and the qualitative figures are suggestive. But the central claim is overreaching, and the missing code makes reproduction harder.\n\nWho is this for: people building interpretable models on cortical surfaces, and anyone working on prototype-based explainability. It deserves a serious referee, but the revision needs to address the latent-space faithfulness gap, temper the abstract, and deal with the FTD confound.","headline":"A genuinely novel spatial-prototype decoder for cortical surfaces, but the 'inherently interpretable' claim needs stronger faithfulness evidence and an honest abstract.","tokens_in":8885,"tokens_out":3311,"would_cite":true,"duration_ms":35610,"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":"The paper presents X-SiT, an inherently interpretable surface vision transformer that classifies Alzheimer's disease and frontotemporal dementia with accuracy matching non-interpretable models, explaining each prediction through…","keywords":["explainable AI","cortical surface analysis","vision transformer","prototypical network","case-based reasoning","Alzheimer's disease","frontotemporal dementia","surface-based morphometry"],"falsifier":"Register the same set of brains to a common template twice, for example with different template choices or surface-processing settings, and check whether the same patch indices still correspond to the same anatomy; if the cosine similarities, the sparse weights, or the prototypes change substantially under re-registration, the interpretability claim is tied to the registration pipeline rather than to cortex structure. Alternatively, replace the template with an independently built average surface and see whether the learned prototypes still land on the known disease regions, namely temporal and temporoparietal areas for Alzheimer's disease and frontal and temporal areas for frontotemporal dementia.","tokens_in":7814,"feed_emoji":"🧠","tokens_out":7762,"duration_ms":79188,"temperature":0.7,"pith_summary":"This paper introduces X-SiT, a transformer network that classifies cortical surface data into dementia diagnoses while explaining each prediction by pointing to specific brain regions of real training cases. The central claim is that X-SiT is the first inherently interpretable neural network for cortical surfaces that keeps pace with non-interpretable state-of-the-art models: on Alzheimer's disease detection it reaches a balanced accuracy of 80.2% versus 80.5% for the plain surface transformer, and on frontotemporal dementia 79.6% versus 80.6%. The explanations are not post-hoc saliency maps but the network's actual decision rule: a weighted cosine similarity between encoded surface patches and learned prototypes at anatomically corresponding locations. If this holds, clinicians could review why a model made a diagnosis and spot unreliable predictions without giving up accuracy.","feed_headline":"Interpretable dementia AI matches black-box accuracy","feed_subtitle":"Novel prototype decoder points to the exact cortical regions behind each dementia diagnosis.","key_machinery":"The central mechanism is the prototypical surface patch (PSP) decoder attached to a Surface Vision Transformer encoder. The input cortical surface is assumed pre-registered to a common template and partitioned into a fixed set of non-overlapping triangular patches, so patch index $i$ denotes the same anatomical region across subjects and prototypes. The encoder maps each patch to a latent embedding; a prototypical sequence $\\xi=(\\xi_1,\\dots,\\xi_N)$ is obtained by encoding learned prototype patches. For each index the decoder computes the cosine similarity $\\cos(x_i, \\xi_i)$ between the input embedding $x_i$ and the prototype embedding $\\xi_i$, rectifies all embeddings with ReLU so that absent features cannot increase class probability, scales each similarity by a sparse weight $w_i$ (with $\\sum_i w_i=1$ and weights below $1/N$ set to zero), and sums to form $P(c|x)=\\sum_i w_i\\cos(x_i,\\xi_i)$. Every five epochs each prototype's latent representation is replaced by the most similar patch from the training set at the same location, keeping prototypes interpretable as real anatomy. This pipeline is what makes the explanation identical to the decision rule.","core_discovery":"On the paper's own terms, the central discovery is that case-based reasoning can replace the opaque classification head of a surface vision transformer without a performance penalty. X-SiT's decoder compares each input patch embedding to a learned prototype embedding at the same spatial index, applies ReLU so that only positive evidence counts, and forms the class probability as a sparse weighted sum of cosine similarities. The prototypes are real training patches refreshed during training, so predictions are literally \"this region resembles that region of a patient in the training set.\" In experiments on two dementia tasks, X-SiT's test balanced accuracy (80.2% for AD, 79.6% for FTD) is competitive with the non-interpretable Surface Vision Transformer (80.5% and 80.6%), and its learned prototypes activate in the temporal and temporoparietal regions for Alzheimer's disease and in frontal and temporal regions for frontotemporal dementia, matching published cortical-thinning patterns. The paper further reports that prototype sets are stable across random initializations (76.3% overlap for AD, 71.7% for FTD) and that individual explanations of misclassified cases deviate from the group-level pattern, suggesting the explanations can flag errors.","pith_inferences":["If the anatomical-correspondence assumption is satisfied across larger, multi-site datasets, the same prototype decoder could be applied to differential diagnosis among several dementia subtypes, where the spatial pattern of atrophy is the key discriminator.","The sparse weighting scheme implies the model deliberately ignores below-average regions; a testable extension would be to compare sparse against dense weighting to see whether diffuse but real disease patterns are being discarded for interpretability's sake.","A natural stress test is to feed X-SiT surfaces registered with different templates or from different preprocessing pipelines; if the prototypes and explanations do not remain stable, the interpretability claim would be tied to a specific registration pipeline rather than to the cortex itself."],"forward_implications":["X-SiT delivers per-prediction explanations that are the model's actual decision rule, not post-hoc attributions, so a clinician can inspect the exact cortical regions that drove a dementia diagnosis.","The learned prototypes double as global disease maps: the same weights and similarities that classify a patient can be averaged across a cohort to show which regions the model finds characteristic of Alzheimer's disease or frontotemporal dementia.","Because prototypes are replaced with real training patches, every explanation can be traced to concrete patient anatomy, making the model's reasoning auditable against clinical knowledge.","The method matches the balanced accuracy of the non-interpretable surface transformer while outperforming the adapted neural-tree interpretable decoder, indicating that interpretability need not be bought with accuracy in this setting."],"supporting_citations":[{"why":"Supplies the Surface Vision Transformer encoder that produces patch-wise latent embeddings for both inputs and prototypes.","marker":"[6]"},{"why":"Introduces the prototypical part-network and case-based reasoning paradigm that the PSP decoder adapts to cortical surfaces.","marker":"[3]"},{"why":"Provides the cortical surface reconstruction and template registration used to produce thickness, curvature, and sulcal depth maps.","marker":"[10]"},{"why":"Defines the neural-tree decoder baseline that the paper adapts to surfaces and consistently outperforms with the PSP decoder.","marker":"[17]"},{"why":"Documents Alzheimer's disease cortical-thinning patterns in temporal and temporoparietal regions used to validate that AD prototypes match known disease topology.","marker":"[30]"},{"why":"Documents distinct regional cortical thinning in Alzheimer's disease and frontotemporal dementia, the reference patterns for validating both sets of prototypes.","marker":"[9]"},{"why":"Serves as a non-interpretable graph-convolution baseline whose test accuracy X-SiT must beat to support the competitive-performance claim.","marker":"[13]"},{"why":"Serves as a second non-interpretable surface baseline included in the performance comparison.","marker":"[35]"}],"fun_headline_variants":["Inherently interpretable transformer matches black-box dementia accuracy","Dementia AI explains itself with cortical prototypes","Case-based reasoning: interpretable dementia diagnosis without accuracy loss","Surface transformer reveals disease patterns in brain cortex"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole explanation pipeline depends on the assumption that after registration to a common template, patch index $i$ names the same anatomical location in every subject's cortex and every prototype, so a cosine similarity at that index is a comparison of matching brain regions; residual registration error or folding variability would make the similarities and the resulting explanations anatomically meaningless.","fun_headline_variants_meta":{"raw":{"variants":["Inherently interpretable transformer matches black-box dementia accuracy","Dementia AI explains itself with cortical prototypes","Case-based reasoning: interpretable dementia diagnosis without accuracy loss","Surface transformer reveals disease patterns in brain cortex"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000244,"raw_usage":{"total_tokens":1535,"prompt_tokens":949,"completion_tokens":586,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":565,"completion_tokens_details":{"reasoning_tokens":525}},"tokens_in":565,"tokens_out":586,"duration_ms":7465,"temperature":1.0,"reasoning_tokens":525,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T22:53:25.039967+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Register the same set of brains to a common template twice, for example with different template choices or surface-processing settings, and check whether the same patch indices still correspond to the same anatomy; if the cosine similarities, the sparse weights, or the prototypes change substantially under re-registration, the interpretability claim is tied to the registration pipeline rather than to cortex structure. Alternatively, replace the template with an independently built average surface and see whether the learned prototypes still land on the known disease regions, namely temporal and temporoparietal areas for Alzheimer's disease and frontal and temporal areas for frontotemporal dementia.","supporting_citations":[{"cited_title":"In: Medical Imaging with Deep Learn- ing","cited_arxiv_id":null,"evidence_quote":"Supplies the Surface Vision Transformer encoder that produces patch-wise latent embeddings for both inputs and prototypes."},{"cited_title":"Advances in neural information processing systems 32 (2019)","cited_arxiv_id":null,"evidence_quote":"Introduces the prototypical part-network and case-based reasoning paradigm that the PSP decoder adapts to cortical surfaces."},{"cited_title":"In: International Conference on Machine Learning","cited_arxiv_id":null,"evidence_quote":"Defines the neural-tree decoder baseline that the paper adapts to surfaces and consistently outperforms with the PSP decoder."},{"cited_title":"Brain 129(11), 2885–2893 (2006)","cited_arxiv_id":null,"evidence_quote":"Documents Alzheimer's disease cortical-thinning patterns in temporal and temporoparietal regions used to validate that AD prototypes match known disease topology."},{"cited_title":"Brain130(4), 1159–1166 (2006)","cited_arxiv_id":null,"evidence_quote":"Documents distinct regional cortical thinning in Alzheimer's disease and frontotemporal dementia, the reference patterns for validating both sets of prototypes."},{"cited_title":"In: 2019 IEEE/CVF International Conference on Computer Vision Workshop (ICCVW)","cited_arxiv_id":null,"evidence_quote":"Serves as a non-interpretable graph-convolution baseline whose test accuracy X-SiT must beat to support the competitive-performance claim."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Serves as a second non-interpretable surface baseline included in the performance comparison."}],"review_version":1}