{"id":"c445e17a-7bee-478b-bffc-577f03ebeb1e","arxiv_id":"2411.10709","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"PathTree improves whole slide image classification by organizing disease categories as a binary tree and using expert-written pathological text to guide feature aggregation, achieving state-of-the-art results on three histopathology datasets.","lead":"PathTree is a method that turns cancer classification of whole slide images into a tree-shaped diagnostic hierarchy and uses written pathology descriptions to guide how the model combines image regions. It reports small but consistent accuracy gains over eight standard methods on lung, prostate, and breast datasets, and larger gains on few-sample settings.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The SOTA claim is not yet established: 5-fold CV is reported without patient-level splits or significance tests, and on PANDA/BRACS multiple WSIs per patient make slide-level folds vulnerable to patient-identity leakage, which could explain the small H-F1 margins.","rationale":"I read the paper in good faith and accept that the method is well-specified enough to reproduce in principle. The reader's weakest assumption (the hand-built tree and expert texts are the right inductive bias) is real but partially addressed by the ablations: Table 5 and Fig. 10 show performance is insensitive to text phrasing and to the graph encoder, which suggests the method does not critically depend on the exact wording; the tree structure itself is not ablated, but on three datasets the chosen trees match standard pathology groupings. I therefore did not make that the deciding issue. The more load-bearing weakness is the evaluation protocol. The paper's strongest quantitative assertion is the 'consistently competitive' SOTA claim; that assertion is only as strong as the CV setup. The manuscript explicitly gives patient counts for BRACS and slide counts for PANDA, yet never states patient-level stratification. With multiple slides per patient and a five-fold split, random slide-level splitting will place same-patient slides in different folds with high probability, creating a leakage path that favors larger models. The reported margins on the key H-F1 metric are within one standard error in several cases, so significance testing is not a formality but a necessary check. The remaining issues I noticed (Eq. 13 and Eq. 14 define H-Precision and H-Recall identically, and Algorithm 1's routing scores use diag(I T^T) while Eq. 7 uses b_alpha t_alpha^T) are reproducible-writing defects that a code release would settle; they do not change my verdict because the empirical claim is what is at stake. Verdict remains conditional: accept only if patient-level splits and significance tests are provided, and code is released to resolve the specification mismatches.","tokens_in":25305,"tokens_out":14445,"duration_ms":148823,"concrete_test":"Ask the authors to rerun the full comparison using patient-stratified 5-fold cross-validation on BRACS and PANDA (all slides of each patient in a single fold), and to report paired bootstrap or Wilcoxon signed-rank tests over the folds for fine-grained ACC, AUC, and H-F1. If PathTree no longer achieves the best or statistically tied H-F1 on all three datasets, the SOTA claim in Tables 2-3 should be softened.","verdict_should_be":"UNCHANGED","load_bearing_attack":"PathTree's central claim is that it 'consistently' outperforms eight WSI baselines on hierarchical F1 (best on all three datasets, Table 3) and is competitive on ACC/AUC. The evidence for that claim depends on the evaluation protocol. The paper reports only '5-fold cross-validation' (Section 4.4) and never states that folds are stratified by patient. This matters because BRACS comprises 547 WSIs from 189 patients (Section 4.1) and PANDA is a multi-biopsy-per-patient cohort in the public PANDA challenge; if the 5-fold split is at slide level, slides from the same patient appear in both training and test folds. Paired slides share staining, tissue, and disease state, so a model can exploit patient-specific cues. PathTree is the highest-capacity model in the comparison (2N-1 attention heads plus a learned tree-GAT and aggregation), so it is the most likely to benefit from such leakage. Moreover, the headline margins are small: e.g., H-F1 on PANDA is 73.29 +/- 0.50 vs CLAM's 73.22 +/- 0.37, and on BRACS 71.09 +/- 3.97 vs DSMIL's 69.72 +/- 3.30. No paired significance test is reported. Without patient-level splits and significance testing, the 'consistently competitive' conclusion is not supported by the data as presented.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes PathTree, a weakly supervised WSI classification method that organizes fine-grained classes into a binary tree and uses expert-written pathological text descriptions for each node. Patch features are extracted with PLIP, multiple slide-level representations are generated either by multi-gated attention or by multi-head Nyström attention, and a text-guided recursive aggregator fuses child embeddings along the tree path. Training combines cross-entropy on text-slide cosine similarity (Eq. 11) with a path-alignment loss (Eq. 8) and a tree-aware triplet loss (Eq. 10). The method is evaluated on an internal lung cryosection dataset, the PANDA prostate grading dataset, and the BRACS breast subtyping dataset against eight WSI baselines, with additional few-shot comparisons against Linear-Probe, CoOp, and TOP.","tokens_in":25633,"tokens_out":5994,"duration_ms":57269,"significance":"The core idea—encoding pathologist-derived binary diagnostic trees and using text embeddings both to guide attention and to serve as the classification weight matrix—is a reasonable and fairly original combination for hierarchical WSI classification. The paper includes a broad baseline comparison, three datasets, pathologist-verified prompt construction, and systematic ablations of text prompts, graph encoders, patch/text encoders, and loss coefficients. The hierarchical F1 improvements over the best planar baselines are directionally consistent across datasets, and the few-shot experiments are a useful addition. However, the evidence as presented is not yet sufficient to support the headline claim of consistent superiority: the hierarchical precision and recall formulas appear to be misprinted, and the evaluation protocol does not state whether folds are patient-stratified, allowing patient-level leakage to inflate the reported margins. If the metric definition is corrected and the evaluation is redone with patient-exclusive, significance-tested folds, PathTree would make a valuable contribution to weakly supervised hierarchical WSI classification.","major_comments":[{"comment":"The printed formulas for H-Precision and H-Recall are identical: both have denominator Σ_j |Ct(j)|. The standard definitions use the predicted label set Σ_j |Cp(j)| in the denominator of H-Precision and the true label set in the denominator of H-Recall. As written, H-Precision and H-Recall would always be equal, which contradicts the different values reported in Table 3. Because H-F1 is the paper's headline metric, the formula must be corrected and the computations in Table 3 must be verified against the intended definitions.","section":"Section 4.3, Eqs. (13) and (14)"},{"comment":"The paper reports only \"5-fold cross-validation\" and does not state whether folds are stratified at the patient level. BRACS consists of 547 WSIs from 189 patients, and PANDA is a multi-biopsy-per-patient cohort in the public PANDA challenge. If the split is performed at the slide level, WSIs from the same patient can appear in both training and test folds. Because PathTree adds trainable components beyond the baselines (the tree-structured GAT prompt encoder and multiple attention heads), it may benefit disproportionately from such leakage. The authors should specify the split granularity and, if patient-level stratification was not used, re-run the evaluation with patient-exclusive folds.","section":"Sections 4.1 and 4.4 (evaluation protocol)"},{"comment":"The reported margins over the best baselines are often small relative to their standard errors: for PANDA H-F1, PathTree obtains 73.29±0.50 versus CLAM's 73.22±0.37; for BRACS H-F1, 71.09±3.97 versus DSMIL's 69.72±3.30. Many fine-grained ACC differences in Table 2 also overlap within one standard error. No paired significance test is reported. Because the paper claims PathTree is \"consistently competitive\" and \"outperforms\" state-of-the-art methods, the authors should provide paired per-fold comparisons (for example, a paired t-test or corrected repeated k-fold test) and temper the conclusion accordingly if the differences are not statistically reliable.","section":"Tables 2 and 3, Section 4.5.1"}],"minor_comments":[{"comment":"The text refers to \"IUSP\" when the intended abbreviation is \"ISUP\" (International Society of Urological Pathology).","section":"Section 4.5.2"},{"comment":"The figure axes are labeled 0.0 to 1.0 while the caption and surrounding text report results in percent; please clarify whether the plotted values are normalized scores or percentages.","section":"Figure 9 and its caption"},{"comment":"The BRACS dataset is described as containing 547 WSIs in Section 4.1 and Table 1, but the Figure 9 panel header states \"549 slides\"; the count should be made consistent.","section":"Section 4.1 and Figure 9"},{"comment":"Minor language issues: \"messages information\" should be \"exchanges information\", and \"slice-level labels\" should be \"slide-level labels\".","section":"Abstract and Section 4.1"},{"comment":"The same text embeddings appear both as targets in the path-alignment and tree-aware losses (Eqs. 8 and 10) and as classifier weights in Eq. (11). This is a legitimate prompt-tuning formulation, but the paper should state explicitly that these losses regularize the representation toward the prediction head rather than provide fully independent external text supervision.","section":"Sections 3.5 and 3.6"}],"recommendation":"major_revision","confidential_remarks":"The duplicate H-Precision/H-Recall formula makes me suspect that Table 3 may have been generated with a different metric implementation than the printed equations; the authors should be asked to provide the computation script or a clear correction. The patient-level split question is the most consequential issue: PANDA and BRACS are both public datasets with known multi-biopsy-per-patient structure, so slide-level folding is a realistic risk of leakage. If the authors can show patient-stratified results and the corrected hierarchical metrics, the paper could become acceptable."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"PathTree is a genuinely assembled recipe rather than a single new idea, and that is fine. It combines a binary tree over fine-grained classes, expert-written pathological text prompts, a bidirectional graph text encoder, multiple attention-based slide features, a text-guided aggregator, and two tree-specific losses. I have not seen this exact assembly in the WSI literature, and it is clearly a step beyond the authors' earlier ROI-level DeepTree. The paper is honest enough to evaluate on three datasets (one in-house, two public) with 5-fold CV against eight baselines, and it includes ablations on prompt wording, graph encoder, patch/text encoders, and loss coefficients. Hierarchical F1 is best on all three datasets, and the few-shot results are suggestive. Those are real positives: the method is specified in enough detail to reimplement, and the ablations show that the tree losses and text prompts matter.\n\nThe soft spots are mostly about evidence, not about the idea. First, no code is released, and the in-house SYSFL dataset is private; reproducibility is limited. Second, the margins over the strongest baselines are often within one standard error, and no significance testing is reported. Third, the paper says only '5-fold cross-validation' and never states that folds are stratified by patient. Since PANDA and BRACS have multiple WSIs per patient, slide-level splits could leak patient identity. That is a legitimate concern, though I do not think it is damning: the baselines share the same split, so the comparison is fair unless PathTree's higher capacity exploits the leakage more than the others, which is plausible but unproven. Fourth, the Ny variant underperforms on BRACS, so the claim of being 'consistently competitive' overstates things; the gated-attention variant is the consistent one.\n\nThe circularity worry—using the same text embeddings as both training targets and classifier weights—is standard CLIP-style metric learning and not a real flaw. The text descriptions are external and expert-written, so the loop is not vacuous.\n\nWho is this for? Computational pathology researchers working on WSI subtyping and anyone interested in injecting structured expert knowledge into weakly supervised slide classifiers. It deserves a serious referee: the idea is plausible, the experiments are extensive, and the weaknesses are fixable by releasing code, adding patient-stratified folds, and reporting significance tests. I would send it to peer review with those requests on the table.","headline":"PathTree is a well-specified text-guided hierarchical WSI classifier with consistent, if modest, gains; the SOTA claim needs stronger evaluation before it lands.","tokens_in":26192,"tokens_out":1973,"would_cite":false,"duration_ms":24249,"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":"PathTree turns whole-slide diagnosis into a text-guided binary tree and reports the best hierarchical F1 on three pathology datasets.","keywords":["whole slide image","hierarchical classification","text-guided representation learning","binary decision tree","pathological diagnosis","slide-level labels","multiple instance learning","vision-language alignment"],"falsifier":"A controlled run on any of the three datasets with the same tree structure and encoders but with the professional text prompts replaced by pathology phrases randomly assigned to nodes would settle it: if hierarchical F1 and AUC stay flat, then text semantics are not doing the work, and the reported gains come from the tree pooling itself.","tokens_in":1801,"feed_emoji":"🔬","tokens_out":2404,"duration_ms":102751,"temperature":0.7,"pith_summary":"PathTree claims that hard-to-separate pathology classes—such as the progressive lung adenocarcinoma subtypes AAH, AIS, MIA, and IAC—are recognized more accurately when a model is forced to reason along the same coarse-to-fine decision tree a pathologist uses, rather than treating every class as an independent flat label. The paper converts each classification task into a binary tree, writes a professional pathological description for every node, and lets those texts steer how patch features extracted from a whole slide image are pooled into a global slide representation. Prediction is made by comparing the slide representation with the fine-grained text prompts, and two tree-specific losses keep the slide aligned with every node on its diagnostic path while pushing it away from sibling and unrelated classes. On three datasets—an in-house lung cryosection set, the public PANDA prostate grading set, and the public BRACS breast subtyping set—PathTree reports the best hierarchical F1 scores and competitive or best fine-grained accuracy and AUC against eight whole-slide baselines. If this is right, slide-level diagnosis can be improved without pixel-level annotation by supplying the model with the same diagnostic structure and vocabulary that pathologists use.","feed_headline":"Text-guided tree improves three whole-slide pathology benchmarks","feed_subtitle":"PathTree routes slide features along a pathologist's decision tree, raising hierarchical F1 on lung, prostate, and breast slides.","key_machinery":"The load-bearing mechanism is the text-guided tree-like aggregator, a recursive bottom-up fusion defined as $b_\\gamma = s_\\alpha b_\\alpha + s_\\beta b_\\beta + i_\\gamma$, where $i_\\gamma$ is the slide embedding of parent node $\\gamma$, $b_\\alpha$ and $b_\\beta$ are the fused embeddings of its two children, and the fusion weights are softmax scores of slide-text similarities: $s_\\alpha = \\frac{\\exp(b_\\alpha t_\\alpha^\\top)}{\\exp(b_\\alpha t_\\alpha^\\top)+\\exp(b_\\beta t_\\beta^\\top)}$, $s_\\beta=1-s_\\alpha$. This makes the text tree control which visual evidence accumulates into the root. Two losses reinforce the hierarchy: path alignment, $L_{\\text{path}} = \\frac{1}{|P|}\\sum_{k\\in P}\\|g - t_k\\|_2^2$, pulls the global slide feature $g$ toward every text embedding $t_k$ on its root-to-leaf path, and tree-aware matching combines triplet losses with margins ordered as $\\lambda_{\\text{leaf}} > \\lambda_{\\text{sibling}} > \\lambda_{\\text{parent}}$ so that sibling nodes are kept closer than unrelated leaves but farther than the parent.","core_discovery":"At the center of the paper is a new way to use text in weakly supervised slide classification: the diagnostic description is not just a label or a prompt that selects a feature, it is the routing rule for hierarchical pooling. PathTree begins with a hand-designed binary tree whose nodes are coarse and fine categories, each described in pathologist-verified language. Text embeddings for all nodes are passed through a bidirectional tree graph so sibling and parent semantics inform one another. The image side produces one slide embedding per tree node using two attention modules, and a text-guided aggregator fuses child embeddings into their parent with weights computed from slide-text similarity; the root embedding is the global slide feature. Classification scores are cosine similarities between that global feature and the fine-grained text prompts, and training adds a path-alignment loss and a tree-aware triplet loss. The claimed net effect is that, compared with flat multiple-instance baselines, PathTree improves hierarchical F1 on all three datasets and improves or matches fine-grained accuracy and AUC, with the detailed text prompts outperforming plain class names in ablation.","pith_inferences":["Beyond the paper: the same design could learn the tree from data rather than from experts, which would test whether the hierarchy itself, rather than the text encoder, drives the reported gains.","Beyond the paper: because prediction is computed by slide-text similarity, PathTree classes are inherently nameable; a natural extension is zero-shot evaluation on a new hospital's categories using only written descriptions and no retraining.","Beyond the paper: the triplet margins encode a claim about semantic distance, with parent closer than sibling; these margins could be calibrated per dataset from pathologist agreement rates, an option the paper does not explore.","Beyond the paper: the Nyström variant improves on SYSFL and PANDA but not BRACS, hinting that global-context aggregation helps some tasks while local high-malignancy regions matter more in breast subtyping; a testable extension is to choose the aggregation module per node type."],"forward_implications":["If PathTree is right, hierarchical labels are not just extra supervision but a better description of the task, so future whole-slide models should be evaluated with hierarchical precision, recall, and F1, not only flat accuracy.","If the gains hold, pathologist-written text acts as a cheap source of expert knowledge that reduces the number of labeled slides needed, since the few-shot experiments show PathTree improving most at 4, 8, and 16 shots per class.","If the ablation is representative, prompt design directly changes clinical performance, meaning that replacing generic templates with professional pathological descriptions is a concrete way to steer model behavior.","If the method generalizes, any multi-class pathology task with an expert decision tree can be converted to this scheme without pixel-level annotation, including new cancer grading or subtyping tasks."],"supporting_citations":[{"why":"Supplies the pre-trained pathology image and text encoders used to encode patches and text prompts.","marker":"Huang et al. 2023"},{"why":"Provides the OncoTree cancer classification system that motivates representing diagnoses as hierarchical trees.","marker":"Kundra et al. 2021"},{"why":"The earlier DeepTree work that introduces binary-tree strategies for pathology regions of interest and is extended here to slide-level learning.","marker":"Li et al. 2023a"},{"why":"Defines ABMIL attention-based multiple instance learning, the baseline and attention mechanism PathTree builds on.","marker":"Ilse et al. 2018"},{"why":"CLAM, a weakly supervised clustering-constrained attention baseline used for comparison and as a design reference.","marker":"Lu et al. 2021b"},{"why":"Provides the PANDA prostate cancer grading dataset used as one of the three evaluation benchmarks.","marker":"Bulten et al. 2022"},{"why":"Provides the BRACS breast carcinoma subtyping dataset used as another evaluation benchmark.","marker":"Brancati et al. 2022"},{"why":"Inspires the path alignment loss through hierarchy-aware global models for hierarchical text classification.","marker":"Zhou et al. 2020"},{"why":"Supplies the Nyström approximation of self-attention used in the PathTree-Ny variant.","marker":"Xiong et al. 2021"}],"fun_headline_variants":["Text-guided tree improves whole-slide pathology benchmarks","PathTree: text-guided hierarchy boosts WSI classification","Tree of text descriptions sharpens slide-level cancer diagnosis","Diagnostic text as routing rules for hierarchical slide analysis","Text-driven tree structure lifts pathology slide F1 scores"],"cache_read_input_tokens":28288,"weakest_assumption_plain":"Everything rests on the premise that the hand-built binary trees and the expert-written text descriptions actually match how pathologists distinguish the classes; if a tree groups the wrong categories or a description does not line up with the visual features, the text-guided pooling and both alignment losses will systematically misdirect the slide representation.","fun_headline_variants_meta":{"raw":{"variants":["Text-guided tree improves whole-slide pathology benchmarks","PathTree: text-guided hierarchy boosts WSI classification","Tree of text descriptions sharpens slide-level cancer diagnosis","Diagnostic text as routing rules for hierarchical slide analysis","Text-driven tree structure lifts pathology slide F1 scores"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000429,"raw_usage":{"total_tokens":2215,"prompt_tokens":988,"completion_tokens":1227,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":604,"completion_tokens_details":{"reasoning_tokens":1152}},"tokens_in":604,"tokens_out":1227,"duration_ms":9664,"temperature":1.0,"reasoning_tokens":1152,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T19:24:22.409714+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A controlled run on any of the three datasets with the same tree structure and encoders but with the professional text prompts replaced by pathology phrases randomly assigned to nodes would settle it: if hierarchical F1 and AUC stay flat, then text semantics are not doing the work, and the reported gains come from the tree pooling itself.","supporting_citations":[{"cited_title":", author Bianchi, F","cited_arxiv_id":null,"evidence_quote":"Supplies the pre-trained pathology image and text encoders used to encode patches and text prompts."},{"cited_title":", author Zhang, H","cited_arxiv_id":null,"evidence_quote":"Provides the OncoTree cancer classification system that motivates representing diagnoses as hierarchical trees."},{"cited_title":", author Tomczak, J","cited_arxiv_id":null,"evidence_quote":"Defines ABMIL attention-based multiple instance learning, the baseline and attention mechanism PathTree builds on."},{"cited_title":"o nberg, H. , author Samaratunga, H. , author Delahunt, B. , author Tsuzuki, T. , author H \\","cited_arxiv_id":null,"evidence_quote":"Provides the PANDA prostate cancer grading dataset used as one of the three evaluation benchmarks."},{"cited_title":", author Anniciello, A.M","cited_arxiv_id":null,"evidence_quote":"Provides the BRACS breast carcinoma subtyping dataset used as another evaluation benchmark."},{"cited_title":", author Ma, C","cited_arxiv_id":null,"evidence_quote":"Inspires the path alignment loss through hierarchy-aware global models for hierarchical text classification."},{"cited_title":"o mformer: A nystr \\","cited_arxiv_id":null,"evidence_quote":"Supplies the Nyström approximation of self-attention used in the PathTree-Ny variant."}],"review_version":1}