{"id":"14a9cd71-381c-4180-91bc-62b2ab627082","arxiv_id":"2506.16398","paper_version":3,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"HyperPath places slide, region, and patch features in hyperbolic space with text-guided losses and classifies by geodesic distance, improving most TCGA benchmarks.","lead":"This paper introduces HyperPath, a method that represents whole slide images of tissue in hyperbolic space and uses text descriptions of cancer classes to guide a semantic hierarchy for classification. It reports improved F1 scores over prior methods on four cancer benchmarks from The Cancer Genome Atlas.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"No Euclidean control: reported gains may stem from text-guided alignment, not hyperbolic geometry.","rationale":"The reader's weakest assumption focused on pseudo-label noise in LAMA. That is a valid implementation-level risk, but it is downstream: even if pseudo-labels are perfect, the paper still does not establish that hyperbolic geometry contributes to the gains. The ablation in Table 2 shows the text-guided alignment loss is responsible for the large improvement, and angular alignment is not hyperbolic-specific. Therefore the more load-bearing concern is the missing Euclidean control, because it directly tests the mechanism claimed in the title and abstract. My verdict remains CONDITIONAL: the paper is a reasonable engineering contribution, but its central mechanistic claim requires an ablation that isolates geometry from alignment. The reader's conditional verdict already anticipates the need for additional validation, so no change in verdict category is needed; the condition should explicitly include an Euclidean counterpart ablation.","tokens_in":9715,"tokens_out":5749,"duration_ms":62176,"concrete_test":"Train an Euclidean version of HyperPath using the same architecture, hierarchical aggregation, adapters, learnable prompts, and LAMA (with Euclidean cosine similarity replacing Lorentzian exterior angles); replace LSHC with a Euclidean margin/contrastive hierarchy loss and replace geodesic classification with Euclidean-distance softmax. Run on all four TCGA tasks under the identical nested split protocol and compare Table 1. If the Euclidean version matches HyperPath within one standard deviation on OOD AUC/F1, hyperbolic geometry is not the driver and the central claim should be revised; if HyperPath is clearly better across tasks, the concern is resolved.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central claim is that hyperbolic semantic hierarchy modeling yields superior WSI classification. However, all comparisons are against Euclidean baselines that do not use HyperPath's text-guided losses or adapters, so the experiments cannot separate geometric benefit from alignment benefit. Table 2 is telling: the no-loss hyperbolic model achieves BRCA OOD AUC 0.864 (below ABMIL's 0.898); adding LAMA alone jumps to 0.925, and adding both losses gives 0.933. Thus nearly all the gain over baselines comes from the angular modality alignment loss, which is a cross-modal objective that does not require hyperbolic geometry (cosine/angular distances are Euclidean concepts). The entailment-cone consistency loss is hyperbolic-specific, but in isolation it collapses performance to 0.539, so it cannot be the driver. Without an Euclidean counterpart trained with the same adapters, prompts, and alignment losses, the headline claim that hyperbolic embeddings are the source of the improvement is underdetermined.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes HyperPath, a method for whole-slide image classification that maps CONCH visual and textual features into a Lorentz hyperbolic space, hierarchically aggregates patch and region features, aligns modalities with an angular modality alignment loss (LAMA) and a semantic hierarchy consistency loss (LSHC), and classifies via geodesic distance to class prototypes. Experiments on four TCGA tasks compare HyperPath against Euclidean MIL baselines and report AUC and F1 scores; an ablation isolates the effects of LAMA and LSHC, and the source code is made available.","tokens_in":9894,"tokens_out":5472,"duration_ms":57075,"significance":"If the central claim were established, HyperPath would be a useful contribution: it is among the few hyperbolic approaches to WSI analysis, it builds on standard pathology foundation-model features, it evaluates on four clinically relevant TCGA tasks, and it releases code. The ablation and the hyperbolic-embedding visualization are informative. However, the experiments do not yet separate the benefit of hyperbolic geometry from the benefit of text-guided alignment, and the statement of \"significant gains across all tasks\" is contradicted by several cells in Table 1. The core geometric claim therefore needs additional evidence before the paper can be accepted.","major_comments":[{"comment":"The claim that HyperPath \"achieves significant gains in both AUC and F1 Score across all tasks\" is not supported by the reported numbers. On LUAD EGFR, HyperPath's OOD AUC is 0.637±0.044, slightly below HIT's 0.638±0.037, and on BRCA HER2 IND, HyperPath's AUC is 0.732±0.157 versus HIT's 0.740±0.144. The text should either restrict the claim to the settings where it holds or report paired statistical tests (e.g., bootstrap or DeLong) that justify the word \"significant.\"","section":"§3.2, Table 1"},{"comment":"The ablation does not isolate the geometric benefit. The no-loss version of HyperPath underperforms ABMIL on BRCA OOD AUC (0.864 versus 0.898), while adding LAMA alone recovers most of the gain (0.925, and 0.933 with the full loss). LSHC alone collapses performance to 0.539 on the same metric. Because LAMA is a cross-modal angular alignment loss that could be implemented with Euclidean cosine similarity, and all baselines are Euclidean methods without textual alignment, the current experiments do not establish that hyperbolic geometry is the source of the improvement. Please add an Euclidean counterpart trained with the same adapters, prompts, and alignment losses.","section":"§3.2, Table 2"},{"comment":"The alignment loss LAMA is applied to patch- and region-level pseudo-labels selected by cosine similarity between raw CONCH visual features and class semantic features. The paper does not report the value of the top-K threshold, the accuracy of these pseudo-labels, or a sensitivity analysis. If the pseudo-labels are noisy, misalignment is propagated into the region and slide representations used for geodesic classification, so the validity of this selection step is load-bearing for the method's performance.","section":"§2.3"}],"minor_comments":[{"comment":"The sentence reporting improvement ranges \"1.9%–9.2%\" and \"2.6%–8.8%\" excludes HIT as an \"outlier,\" but HIT is a competitive baseline on BRCA HER2 and LUAD EGFR. Please justify this exclusion and reconcile the ranges with the actual table entries.","section":"§3.2"},{"comment":"The negative sampling strategy is not specified. Please state how the negative hyperbolic embeddings v^- are selected (other classes, other hierarchical levels, or batch negatives) and whether the sum in the denominator includes one or many negatives.","section":"§2.3, Eq. (3)"},{"comment":"The half-aperture constant alpha and the margin beta are set to 0.1 and 0.8 without sensitivity analysis; please add a short ablation or a reference justifying these choices.","section":"§2.4, Eq. (4)"},{"comment":"The shapes in the aggregation equation are not fully defined: the superscript T on f^I_{h',m} and the orientation of the resulting vector are unclear. Please clarify the dimensions so that the formula can be checked.","section":"§2.2, Eq. (1)"},{"comment":"The caption does not explain the color scheme or markers, and in grayscale the distinct categories are difficult to separate. Please add a legend and describe the axes.","section":"Fig. 3"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is within the scope of the journal and appears to be a genuine empirical contribution, but the central geometric claim needs the Euclidean control experiment described in Major Comment 2 before publication."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: a competent, well-engineered adaptation of hyperbolic embeddings to WSI classification, with two new losses and solid experimental effort. What is actually new is the package: hyperbolic CONCH features, text-guided semantic hierarchy, geodesic-distance classification, and the LAMA/LSHC losses. That combination is new for WSIs. The code link is a plus.\n\nExperiments are reasonably thorough—four TCGA tasks, IND/OOD splits, several MIL baselines, ablations, visualizations. On most cells HyperPath improves over the best baseline, and OOD gains look like a real robustness improvement.\n\nSoft spots, in proportion. First, the claims outrun the table. Section 3.2 says \"significant gains across all tasks,\" but Table 1 shows ties and reversals: BRCA F_OOD is 0.696 against CLAM's 0.695; LUAD EGFR IND F1 is below HIPT; BRCA HER2 IND AUC is below HIT. No significance tests, so most differences are within 1 SD. That phrasing must change.\n\nSecond, and more important, the attribution to hyperbolic geometry is underdetermined. The stress-test concern lands: the no-loss hyperbolic model gets 0.864 AUC on BRCA OOD, below ABMIL's 0.898; LAMA alone jumps to 0.925. Most of the headroom comes from the alignment loss, which is not a purely hyperbolic mechanism. There is no Euclidean baseline trained with the same adapters, prompts, and losses. The full method may still be a good system, but the paper doesn't isolate the geometric contribution.\n\nThe pseudo-label selection for LAMA (top-K by cosine similarity on raw CONCH features) is assumed clean; no sensitivity analysis. Minor. And LSHC alone collapses to near chance, which suggests the consistency loss is only effective after alignment, i.e., the design is more tuned than principled. Also minor.\n\nCitations are appropriate; no invented entities. The paper is for people doing hyperbolic representation learning in medical imaging and computational pathology researchers chasing OOD robustness. It deserves peer review. I would send it out, but with explicit requests: add a Euclidean counterpart with the same losses, report confidence intervals or paired significance, and soften the all-task superiority claim. Revised, this is a solid paper.","headline":"Useful engineering adaptation of hyperbolic embeddings to WSI classification, but the headline claim that hyperbolic geometry drives the gains is undercut by a missing Euclidean control.","tokens_in":10406,"tokens_out":4201,"would_cite":false,"duration_ms":40881,"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":"HyperPath models whole-slide images as patch-region-slide hierarchies in hyperbolic space, guided by textual class knowledge, to improve cancer classification.","keywords":["hyperbolic space","whole slide image analysis","multiple instance learning","vision-language model","semantic hierarchy","entailment cones","computational pathology","cancer subtyping"],"falsifier":"Run HyperPath on the four TCGA tasks with the Angular Modality Alignment Loss replaced by an equivalent Euclidean cosine alignment loss, keeping the same hierarchical aggregation, text guidance, and entailment loss; if the Euclidean variant matches or exceeds HyperPath's AUC and F1, hyperbolic geometry per se is not what causes the gains. Alternatively, compare the CONCH cosine-similarity pseudo-labels with pathologist- or supervised-model-derived patch and region labels; low agreement would make the alignment signal suspect.","tokens_in":9533,"feed_emoji":"🔬","tokens_out":8040,"duration_ms":75336,"temperature":0.7,"pith_summary":"HyperPath is a method for whole-slide image (WSI) classification that treats a slide as a semantic hierarchy — patches, regions, slide — and embeds that hierarchy in hyperbolic space, a geometry whose exponential growth gives hierarchical levels natural room to separate. The paper claims that text-guided hyperbolic modeling produces more coherent representations and better classification than standard multiple instance learning (MIL) baselines, which mostly work in Euclidean space. HyperPath aligns visual features from the CONCH pathology model with textual class semantics using an angular alignment loss, and it enforces entailment and contradiction relations with a semantic hierarchy consistency loss. Classification is read directly from geodesic distances between slide and class embeddings, so no linear classifier is needed. On four TCGA tasks, the paper reports consistent AUC and F1 gains over non-hierarchical and hierarchical baselines, with the largest gains in out-of-domain settings.","feed_headline":"Hyperbolic hierarchy lifts cancer-slide classification accuracy","feed_subtitle":"HyperPath aligns patch, region, and slide levels with class text in hyperbolic space, then classifies with geodesic distance.","key_machinery":"The load-bearing object is the Lorentz model of hyperbolic space $H^k_\\rho$, a manifold whose exponential volume growth gives hierarchy an intrinsic notion of distance. Visual and textual features from CONCH are mapped into this space; an attention aggregator builds region and slide features from patches; the Angular Modality Alignment Loss $\\mathcal{L}_{\\mathrm{AMA}}$ measures similarity through exterior angles $\\theta(u,v)$ and angular distance $\\varphi(u,v)$; the Semantic Hierarchy Consistency Loss $\\mathcal{L}_{\\mathrm{SHC}}$ uses hyperbolic entailment cones, cones of points that a concept semantically entails, with half-aperture $\\phi(u)=\\sin^{-1}(2\\alpha/(\\sqrt{\\rho}\\|u_s\\|_E))$, to enforce entailment and contradiction structure; and slide classification uses geodesic distance $d_G(u,v)=\\sqrt{1/\\rho}\\,\\cosh^{-1}(-\\rho\\langle u,v\\rangle_H)$ between slide and class embeddings, replacing the linear classifier.","core_discovery":"The central discovery claimed is that the natural patch-region-slide organization of a whole slide image is better represented as a hyperbolic semantic hierarchy than as a Euclidean bag of patches. In the Lorentz model, broad class concepts sit near the origin with wide entailment cones, while slide, region, and patch features radiate outward at increasing specificity; HyperPath argues this geometry is the right inductive bias for tissue structure. To make it work, the paper contributes an Angular Modality Alignment Loss based on exterior angles, which avoids the scale mismatch between general textual embeddings and specific visual embeddings, and a Semantic Hierarchy Consistency Loss that pushes entailed features inside cones and contradictory features outside. Classification then reduces to comparing the slide embedding with each class embedding by geodesic distance. The paper supports the claim with experiments on four TCGA tasks and ablations showing that both losses are needed together.","pith_inferences":["The top-$K$ pseudo-label selection by raw CONCH cosine similarity is the fragile point; an obvious untested extension is iterative self-training with uncertainty-aware pseudo-labels, which could reduce propagated misalignment.","The large F1 improvements on HER2 and EGFR tasks relative to AUC suggest the geometry mainly counters majority-class bias; a per-class calibration and confusion-matrix analysis would test that directly.","The same hyperbolic hierarchy machinery could be applied to survival prediction or tumor microenvironment characterization, where scale and tissue organization also matter, but the paper does not evaluate those tasks.","Since hierarchy consistency alone degrades performance, one testable hypothesis is that angular alignment normalizes feature distributions before entailment constraints become useful; scheduling the two losses in sequence may be simpler and stronger than weighting them jointly."],"forward_implications":["The standard MIL pipeline can drop the learned linear classification head: prediction becomes a nearest-class lookup by geodesic distance in hyperbolic space.","Patch, region, and slide features become semantically layered by specificity, so the model offers an interpretable account of which tissue scale supports a slide-level diagnosis.","Because angular alignment does not depend on matching geodesic scales, the approach should transfer to other gigapixel or hierarchical biomedical images where class concepts can be written as text prompts.","Text-guided hyperbolic embeddings could improve out-of-domain robustness in computational pathology, since the reported out-of-domain gains exceed the in-domain gains on most tasks.","The two losses are complementary rather than individually sufficient: ablation results show alignment alone helps, hierarchy consistency alone hurts, and the combination gives the best result."],"supporting_citations":[{"why":"Supplies the pathology vision-language feature extractor whose visual and textual embeddings HyperPath adapts to hyperbolic space.","marker":"[18]"},{"why":"Provides the hyperbolic image-text representation learning setup, including the Lorentz model, that HyperPath extends with angular alignment.","marker":"[5]"},{"why":"Defines hyperbolic entailment cones and the cone-loss formulation that the Semantic Hierarchy Consistency Loss builds on.","marker":"[6]"},{"why":"Serves as a MIL baseline that HyperPath compares against on all four tasks.","marker":"[24]"},{"why":"Serves as the hierarchical interaction baseline closest in spirit to HyperPath, using asynchronous hierarchical transformers.","marker":"[10]"},{"why":"Serves as the hierarchical self-supervised vision transformer baseline for gigapixel WSIs.","marker":"[2]"},{"why":"Serves as a widely used attention-based MIL baseline in the comparison.","marker":"[19]"},{"why":"Serves as a double-tier feature distillation MIL baseline in the comparison.","marker":"[32]"}],"fun_headline_variants":["HyperPath: hyperbolic space maps slide hierarchy for cancer diagnosis","Hyperbolic embeddings beat Euclidean for whole-slide cancer classification","Text-guided hyperbolic hierarchy boosts WSI classification accuracy","Geometry-aware MIL: hyperbolic hierarchy for cancer slides"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the top-$K$ patch and region pseudo-labels chosen by cosine similarity between raw CONCH visual features and class text features are accurate enough to train the cross-modal alignment loss; if those pseudo-labels are noisy, the misalignment propagates into the slide representation used for classification.","fun_headline_variants_meta":{"raw":{"variants":["HyperPath: hyperbolic space maps slide hierarchy for cancer diagnosis","Hyperbolic embeddings beat Euclidean for whole-slide cancer classification","Text-guided hyperbolic hierarchy boosts WSI classification accuracy","Geometry-aware MIL: hyperbolic hierarchy for cancer slides"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000184,"raw_usage":{"total_tokens":1306,"prompt_tokens":924,"completion_tokens":382,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":540,"completion_tokens_details":{"reasoning_tokens":319}},"tokens_in":540,"tokens_out":382,"duration_ms":4358,"temperature":1.0,"reasoning_tokens":319,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T19:26:41.515203+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run HyperPath on the four TCGA tasks with the Angular Modality Alignment Loss replaced by an equivalent Euclidean cosine alignment loss, keeping the same hierarchical aggregation, text guidance, and entailment loss; if the Euclidean variant matches or exceeds HyperPath's AUC and F1, hyperbolic geometry per se is not what causes the gains. Alternatively, compare the CONCH cosine-similarity pseudo-labels with pathologist- or supervised-model-derived patch and region labels; low agreement would make the alignment signal suspect.","supporting_citations":[{"cited_title":"In: Proceedings of the IEEE/CVF International Conference on Computer Vision","cited_arxiv_id":null,"evidence_quote":"Serves as the hierarchical interaction baseline closest in spirit to HyperPath, using asynchronous hierarchical transformers."}],"review_version":2}