{"id":"fdc6fcb8-8fda-48a6-bf99-e52277efb175","arxiv_id":"2506.02542","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"A heterogeneous graph neural network that includes immune cells around kidney glomeruli achieves the best cross-patient F1 score (0.73) among tested models on a six-patient kidney transplant dataset.","lead":"This paper builds graphs from kidney tissue images, with glomeruli, macrophages and T-cells as linked nodes, and trains a heterogeneous graph neural network (HIEGNet) to classify each glomerulus as healthy, sclerotic, or dead. On a six-patient dataset, HIEGNet reached the best cross-patient F1 score (0.73) compared with the tested CNN and random forest baselines.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Between-patient result lacks a no-immune control; segmentation metrics suggest immune nodes are near-noise on unseen patients, so the 0.73 F1 may be carried by glomeruli-only structure.","rationale":"The reader's verdict is already conditional, and the segmentation-quality weakness is the right place to look. The reported between-patient F1 of 0.73 for HIEGNet is the strongest evidence for the paper's central claim. For that evidence to support the specific claim that the immune environment helps cross-patient classification, the immune nodes on patients 004–006 need to carry signal. Appendix B shows they may not: macrophage segmentation AP-AUC is consistently low, with 0.003 in some between-patient columns, and the table's patient labels do not even include the actual test patients 005/006. The paper's only immune-environment ablation, Appendix K, is within-patient, so it cannot establish that immune nodes help on unseen patients. Without a glomeruli-only graph control, the 0.13 F1 gap over Random Forest could be explained entirely by rg,g message passing or by the graph's inductive bias, not by the novel immune-component representation. This is a concrete, addressable weakness rather than a rejection: adding a between-patient no-immune ablation and a homogeneous GNN baseline would settle it. The apparent typo in the HIEGNet update equation (the inner sum over N(v) uses h_v^(ℓ) instead of the neighbour states) is a secondary reproducibility issue but does not change the empirical concern. I therefore keep the reader's CONDITIONAL verdict unchanged.","tokens_in":16068,"tokens_out":7585,"duration_ms":80354,"concrete_test":"Run the between-patient setting (train 001–003, test 004–006) with a glomeruli-only variant of HIEGNet: remove all immune-cell nodes and all edges in Ri∪Rig, retaining only rg,g edges and the identical glomerulus node features, hyperparameters, and random-initialization protocol. If this variant's macro F1 remains at or near 0.73, the immune environment is not responsible for the headline cross-patient result; if it drops materially (e.g., >0.05), the concern is resolved in the authors' favor. As a secondary check, report segmentation AP-AUC for patients 005 and 006 and add a standard homogeneous GNN baseline on glomeruli only.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The load-bearing step is the implicit assumption that the immune-cell nodes used to build HIEGNet's graph are informative on the unseen test patients 004–006. The paper's own segmentation evaluation (Appendix B) undermines this: for macrophages, the selected contour-detection method achieves an AP-AUC of only 0.265 within patients and between-patient values as low as 0.003–0.202 (Table 3); fine-tuned Cellpose is 0.000 everywhere. T-cell segmentation is better but still limited (fine-tuned Cellpose AUC 0.49–0.55). At these quality levels, most macrophage nodes on new patients are likely false detections with uninformative features. The only ablation that supports the immune environment (Appendix K) removes edge groups only in the within-patient setting, where segmentation is best (Rig removal drops F1 by 0.07). No equivalent between-patient ablation is reported, and no homogeneous GNN baseline on a glomeruli-only graph is provided. Since the Random Forest baseline already uses the same hand-crafted glomerulus features, the observed 0.73 vs 0.60 between-patient gap cannot be attributed to the immune environment without such a control. The Appendix B between-patient columns are also internally inconsistent (test patients listed as 001–004 rather than 004–006), so segmentation quality on the actual test patients 005/006 is not reported.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes HIEGNet, a heterogeneous graph neural network for classifying glomeruli in whole-slide kidney histopathology images. Nodes represent glomeruli, macrophages, and T-cells; edges are created from distance thresholds and k-nearest-neighbour rules; node features are hand-crafted LBP and shape descriptors. The authors evaluate HIEGNet on the EXC dataset of six kidney-transplant patients in two settings: within patients (patients 001-003, 85/15 split) and between patients (train on 001-003, test on 004-006). They report that HIEGNet reaches a between-patient macro F1 of 0.73 ± 0.01, outperforming Random Forest (0.60), ResNet-18 (0.59), EfficientNetV2 (0.55), and U-Net (0.54). The implementation is publicly available.","tokens_in":16319,"tokens_out":5386,"duration_ms":56830,"significance":"If substantiated, the between-patient generalisation result would be a useful step for graph-based histopathology, since it targets the practically important scenario of classifying glomeruli in patients not seen during training. Strengths of the paper include the public code, the detailed hyperparameter-search description, the reported computational cost of graph construction and training, and the explicit edge-importance ablation within patients. However, the evidence is currently thinner than the central claim: the between-patient comparison rests on a single train/test patient split, the segmentation of immune cells is reported to be very poor on some between-patient settings, and no homogeneous-GNN or glomeruli-only control is provided. The significance of the contribution is therefore conditional on closing these gaps.","major_comments":[{"comment":"The segmentation evaluation is internally inconsistent and does not cover the actual test patients. Table 3 reports between-patient AP-AUC values for macrophages as low as 0.003 and 0.000, and the column headers list test patients 001-004, whereas Section 5 defines the between-patient test set as patients 004-006 and Table 1 assigns only patient 004 patches to the between-patient test set. Consequently, the manuscript provides no segmentation-quality measurement on the patients used for the reported 0.73 between-patient F1 score. Since the immune-cell nodes in the graph are produced by this segmentation, the claim that the immune environment contributes to cross-patient performance is not supported without either improved segmentation metrics on patients 004-006 or a control that removes the immune nodes.","section":"Appendix B, Tables 2 and 3"},{"comment":"The only ablation that isolates the contribution of the immune environment is performed in the within-patient setting (Appendix K, Table 10), where segmentation quality is highest. No equivalent between-patient ablation is reported, and no homogeneous GNN or glomeruli-only graph control is provided. Because the Random Forest baseline already uses the same hand-crafted glomerulus features, the observed 0.73 versus 0.60 between-patient gap cannot be attributed to the immune-cell nodes or to the heterogeneous graph structure without such a control. The authors should add a between-patient ablation that removes immune-cell edge groups, and ideally a homogeneous GNN baseline on a glomeruli-only graph.","section":"Section 5 and Appendix K"},{"comment":"The between-patient result is based on a single partition of the six patients into training (001-003) and test (004-006) sets, with no significance testing across patients. The reported standard deviations come from 20 parameter initialisations and therefore quantify only initialisation variance, not the variance across possible patient splits. To support the claim that HIEGNet 'generalises best between patients among all baseline models', the authors should report patient-level cross-validation (or a permutation test) and, if possible, confidence intervals for the macro-F1 differences. Without this, the 0.13-point gap over Random Forest could be driven by the particular choice of patients.","section":"Section 5, 'between patients setting'"},{"comment":"The baseline set contains no standard GNN (e.g., GCN, GraphSAGE, or RGCN) applied to the same node features and graph structure. The comparison is therefore between HIEGNet, CNNs, and a Random Forest that ignores graph topology. This omission makes it impossible to isolate the contribution of the proposed heterogeneous message-passing architecture from the contribution of the graph representation itself. Adding a homogeneous GNN baseline on the same graph and features is necessary to support the architectural claims in Section 4.","section":"Section 5, Baselines"},{"comment":"The description of how the contour-detection thresholds were selected is contradictory and raises a possible leakage issue. The text states that 'the thresholds were optimised on the training set', but then says 'the final thresholds were determined on the test set within patients 001 - 003 with an intensity threshold value of 60 and an area threshold of 160 pixels'. If the threshold values were chosen using the within-patient test split, then the within-patient evaluation is not a clean evaluation, and the model selection performed on that setting may be compromised. The authors must clarify the protocol and, if thresholds were indeed selected on the test set, re-run the experiments with a proper training/validation split.","section":"Appendix B, threshold selection"}],"minor_comments":[{"comment":"The SAGE aggregation term writes 'sum_{u in N(v)} h_v / |N(v)|', which appears to be a typo: the summed quantity should be the neighbour embedding h_u, not h_v. The neighbouring set in the first term also reads 'N(v) ∪ {u}', which is unclear and should be written as N(v) or N(v) ∪ {v} depending on whether self-loops are intended.","section":"Section 4, display equation after Eq. (1)"},{"comment":"Precision and recall for ResNet-18 and EfficientNetV2 in the between-patients columns are reported as '0.– ± 0.–', which is not a usable result. The authors should report the actual values or explicitly state why they are unavailable (for example, due to undefined metrics from a missing class).","section":"Appendix G, Table 6"},{"comment":"The text says 'all four patients are affected by glomerulosclerosis', but the dataset contains six patients and the following sentences refer to six patients. This is presumably a typo and should be corrected.","section":"Appendix E, first paragraph"},{"comment":"The column headers 'Between patients: 001 002 003 004' conflict with the main text, which uses patients 004-006 as the between-patient test set. The table headers should be relabelled to match the actual experimental setting.","section":"Appendix B, Tables 2 and 3"},{"comment":"The manuscript contains several typographical or formatting issues, including 'Heterogenous' in the title, 'Gird Search Space' in Appendix H, and unusual spacing in 'F uture W ork'. These should be corrected during revision.","section":"Throughout"}],"recommendation":"major_revision","confidential_remarks":"The central claim is plausible and the paper makes a clear practical contribution, but the evidence is not yet sufficient for acceptance. In particular, the missing between-patient ablation and the inconsistent segmentation evaluation directly affect the interpretation of the headline F1 score. These issues are addressable within the scope of the manuscript, so I would prefer major revision over rejection. If the authors cannot add patient-level cross-validation, they should substantially temper the 'generalises best' claim."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nQuick take: this is a credible, mostly honest computational pathology paper that does something genuinely new – it builds a heterogeneous graph with glomeruli, macrophages, and T-cells as separate node types, and proposes a GNN (HIEGNet) with edge-type-specific message passing. The between-patient evaluation (train on patients 001-003, test on 004-006) is the right protocol, and the authors report segmentation quality, ablations, and public code, which is more than many papers in this area do.\n\nThe headline result is HIEGNet reaching 0.73 macro-F1 between patients versus 0.60 for Random Forest, 0.59 for ResNet-18, 0.54 for U-Net. That gap looks good. But the claim that the immune environment is what carries the cross-patient performance is not actually supported. The macrophage segmentation on unseen patients is near-noise: AP-AUC of 0.003-0.202 between patients, and fine-tuned Cellpose gets 0.000. T-cell segmentation is better (0.49-0.55 for fine-tuned Cellpose). The only ablation that removes immune-glomerulus edges (Appendix K) is within-patient; there is no between-patient ablation. And there is no homogeneous GNN baseline on a glomeruli-only graph, so the 0.73 vs 0.60 gap could equally come from modeling glomerulus-glomerulus relationships or from hand-crafted features working better in a graph model.\n\nOther soft spots: only one between-patient split, no significance testing or patient-level variance; the 20 random seeds only cover the same split. The equation in Section 4 has an index typo (the GraphSAGE sum should be over h_u, not h_v, and the formal message passing definition mixes layer indices ell and ell-1). And Appendix B's between-patient columns list test patients as 001-004 instead of 004-006, so we never see segmentation quality on the actual test patients 005/006. These are all fixable.\n\nWhat the paper does well: the graph construction pipeline is detailed, the feature engineering is thoughtful, the hyperparameter search is described, and the authors don't hide the segmentation failures. The between-patient protocol is standard model selection, not circularity.\n\nBottom line: this deserves a serious referee – it is a solid method paper for the computational pathology subfield. I would not desk-reject it. But the immune-environment narrative needs a homogeneous GNN baseline and a between-patient edge ablation before I'd trust the central claim. The equation and appendix inconsistency should also be fixed.\n\nRecommendation: send to peer review, asking for those additions. For my own reading group, maybe; I'd cite it if I worked on GNNs in histopathology.","headline":"Useful, honest computational pathology paper whose cross-patient immune-environment claim is weaker than it looks once you check the segmentation quality and the missing homogeneous GNN baseline.","tokens_in":16923,"tokens_out":4667,"would_cite":true,"duration_ms":45159,"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":"HIEGNet, a heterogeneous graph network that includes immune cells as nodes, classifies glomeruli and generalises between patients better than the CNN and random-forest baselines tested on the EXC dataset.","keywords":["graph neural networks","histopathology","glomeruli classification","immune environment","heterogeneous graphs","whole slide images","kidney transplant","glomerulosclerosis"],"falsifier":"Run the between-patient protocol twice: once with the current pipeline's predicted immune-cell nodes and once with all macrophage and T-cell nodes removed entirely. If the macro F1 does not fall below 0.73 ± 0.01 when immune nodes are removed, the claim that the immune environment drives cross-patient generalisation is falsified.","tokens_in":15846,"feed_emoji":"🔬","tokens_out":6356,"duration_ms":54596,"temperature":0.7,"pith_summary":"The paper tries to establish that a heterogeneous graph representation of kidney tissue, where each glomerulus and the immune cells around it are nodes connected by distance-based edges, lets a graph neural network classify glomeruli as healthy, sclerotic, or dead, and that this representation transfers to unseen patients better than the tested convolutional baselines and a random forest. On the EXC dataset of six kidney-transplant patients with 2052 glomeruli, the proposed HIEGNet reaches a macro F1 of 0.73 between patients, against 0.60 for random forest, 0.59 for ResNet-18, 0.55 for EfficientNetV2, and 0.54 for U-Net. The paper further argues that the immune environment matters: deleting edges between glomeruli and immune cells drops the within-patient F1 by 0.07, the largest drop among the three edge groups. A sympathetic reader should care because, if true, it shows that explicitly modelling cell-level interactions with GNNs is a viable route to whole-slide histopathology classification that generalises across patients.","feed_headline":"Immune-aware graph net wins on unseen kidney patients","feed_subtitle":"HIEGNet hits 0.73 across-patient F1, beating CNNs and random forest on glomeruli.","key_machinery":"The central object is the heterogeneous graph $G = (V, E, T, R)$ with node types $T = \\{$glomeruli, macrophages, T-cells$\\}$, built from whole-slide images by segmentation (manual glomeruli masks, Cellpose or contour detection for immune cells) and hand-crafted features (Local Binary Patterns plus shape descriptors). Edges are created by Euclidean distance, with edge-type-specific construction: k-NN plus $\\epsilon$-neighbourhood ($\\epsilon = 100\\,\\mu$m) among immune cells, $\\epsilon = 277\\,\\mu$m between glomeruli and immune cells, and $\\epsilon = 138.6\\,\\mu$m between glomeruli. The architecture uses RGCN-style aggregation with a different message passing function per edge-type group, namely GraphSAGE for glomerulus–glomerulus edges, GATv2 for glomerulus–immune edges, and SchNet's CFconv for immune–immune edges, sharing parameters only within the same edge type. That design is what lets the model exchange information between a glomerulus and its immune environment and across neighbouring glomeruli.","core_discovery":"On its own terms, the paper reports that HIEGNet generalises best between patients among all models tested: macro F1 0.73 ± 0.01 on patients 004–006 after training on patients 001–003, versus 0.60 for Random Forest, 0.59 for ResNet-18, 0.55 for EfficientNetV2, and 0.54 for the U-Net. Within the same patients, the U-Net reaches 0.77 and HIEGNet 0.73; HIEGNet's between-patient drop is much smaller than the CNNs'. The paper also reports that removing glomerulus–immune-cell edges lowers within-patient F1 by 0.07, more than removing the other edge groups, which it reads as evidence that the immune environment carries classification signal.","pith_inferences":["The paper does not show that the immune-cell nodes themselves drive the between-patient gain; given the reported macrophage segmentation AUC as low as 0.003 on unseen patients, the gain may be carried by the hand-crafted glomeruli features and glomerulus–glomerulus edges, and a direct ablation removing immune nodes would separate these.","The same graph construction and edge-type-group architecture could be applied to other histopathology tasks with different object types, provided the cells of interest can be stained and segmented, though the distance thresholds and feature sets would need re-derivation.","Because the paper's stain-augmentation experiments degrade HIEGNet much more than a pre-trained U-Net, the cross-patient advantage may depend on the training and test images sharing the same staining protocol; training on multiple sites is a testable extension the paper itself flags as future work.","The between-patient test set contains only three patients with heterogeneous class distributions, so the reported superiority should be read with the small sample size in mind."],"forward_implications":["If HIEGNet's cross-patient F1 of 0.73 holds, GNNs that explicitly model immune cells become a viable alternative to CNNs for nephropathology, with the benefit of whole-slide context.","The linear-in-edges complexity and measured 1.21 s per epoch make the approach feasible on full WSIs, including graphs with up to roughly 4.1 million edges.","Hand-crafted, rotation- and stain-robust features (LBP and shape) appear to transfer across patients, as evidenced by Random Forest's small gap between the within- and between-patient settings.","The edge-removal ablation indicates that glomerulus–immune-cell edges carry more classification-relevant information than immune–immune or glomerulus–glomerulus edges.","The architecture's per-edge-type parameter sharing means additional cell types, such as B-cells, can be added without redesigning the network."],"supporting_citations":[{"why":"Supplies the EXC dataset and the 277 µm immune-environment radius used for glomerulus–immune-cell edges.","marker":"Merveille et al., 2021"},{"why":"Provides Cellpose, the segmentation model fine-tuned for T-cells and compared for macrophages.","marker":"Stringer et al., 2020"},{"why":"Provides the GraphSAGE message passing used for glomerulus–glomerulus edges.","marker":"Hamilton et al., 2017"},{"why":"Provides GATv2 attention used for glomerulus–immune-cell edges.","marker":"Brody et al., 2022"},{"why":"Provides SchNet's CFconv message passing used for immune-cell–immune-cell edges.","marker":"Schütt et al., 2017"},{"why":"Provides the RGCN edge-type aggregation mechanism that HIEGNet extends to different message passing functions.","marker":"Schlichtkrull et al., 2018"},{"why":"Establishes the CNN baselines (ResNet, EfficientNetV2) and the Oxford-classification context for glomerulosclerosis.","marker":"Altini et al., 2023"},{"why":"Earlier deep-learning glomeruli classification work that motivates the task and comparison.","marker":"Ayyar et al., 2018"},{"why":"Supplies the pre-trained U-Net used both as a baseline and in the Hybrid-HIEGNet variation.","marker":"Nisar and Lampert, 2024"}],"fun_headline_variants":["Immune-aware graph net beats CNNs on unseen kidney patients","Heterogeneous GNN uses immune cells to classify kidney glomeruli","Graph network that reads immune environment excels in kidney glomeruli","Immune-cell edges boost cross-patient performance of kidney GNN","HIEGNet: glomeruli classifier that generalizes best across patients"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The claim depends on the automatically segmented immune cells around a glomerulus being accurate enough on new patients to provide the cross-patient signal attributed to the immune environment; the paper's own Appendix B reports macrophage segmentation AUC as low as 0.003 in between-patient settings, so those nodes may be mostly noise outside the training patients.","fun_headline_variants_meta":{"raw":{"variants":["Immune-aware graph net beats CNNs on unseen kidney patients","Heterogeneous GNN uses immune cells to classify kidney glomeruli","Graph network that reads immune environment excels in kidney glomeruli","Immune-cell edges boost cross-patient performance of kidney GNN","HIEGNet: glomeruli classifier that generalizes best across patients"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000246,"raw_usage":{"total_tokens":1526,"prompt_tokens":918,"completion_tokens":608,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":534,"completion_tokens_details":{"reasoning_tokens":521}},"tokens_in":534,"tokens_out":608,"duration_ms":6066,"temperature":1.0,"reasoning_tokens":521,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T11:21:25.645416+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the between-patient protocol twice: once with the current pipeline's predicted immune-cell nodes and once with all macrophage and T-cell nodes removed entirely. If the macro F1 does not fall below 0.73 ± 0.01 when immune nodes are removed, the claim that the immune environment drives cross-patient generalisation is falsified.","supporting_citations":[{"cited_title":"An automatic framework for fusing information from differently stained consecutive digital whole slide images: A case study in renal histology","cited_arxiv_id":null,"evidence_quote":"Supplies the EXC dataset and the 277 µm immune-environment radius used for glomerulus–immune-cell edges."},{"cited_title":"Kipf, Peter Bloem, Rianne van den Berg, Ivan Titov, and Max Welling","cited_arxiv_id":null,"evidence_quote":"Provides the RGCN edge-type aggregation mechanism that HIEGNet extends to different message passing functions."},{"cited_title":"Performance and limitations of a supervised deep learning approach for the histopathological O xford classification of glomeruli with I g A nephropathy","cited_arxiv_id":null,"evidence_quote":"Establishes the CNN baselines (ResNet, EfficientNetV2) and the Oxford-classification context for glomerulosclerosis."},{"cited_title":"Harnessing AI for kidney glomeruli classification","cited_arxiv_id":null,"evidence_quote":"Earlier deep-learning glomeruli classification work that motivates the task and comparison."},{"cited_title":"Maximising histopathology segmentation using minimal labels via self-supervision","cited_arxiv_id":null,"evidence_quote":"Supplies the pre-trained U-Net used both as a baseline and in the Hybrid-HIEGNet variation."}],"review_version":1}