{"id":"3003a853-11aa-4575-a0c6-0cafd518b5b3","arxiv_id":"1908.05467","paper_version":1,"verdict":"REJECT","confidence":"HIGH","novelty_score":5.0,"correctness_risk":"high","formal_verification":"none","parameter_count":3,"one_line_summary":"Applying the HOMFLY knot polynomial to feature maps of unfolded rib CT images yields a training-set precision of 0.60 for fracture images, but the approach does not predict on unseen images.","lead":"Researchers tested whether the HOMFLY polynomial, a knot theory invariant, can classify rib fractures in postmortem CT scans. The method partially separated fracture from non-fracture images, but it cannot yet classify new images, and its reported precision is an in-sample measure.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The reported 0.60 precision is a list-overlap statistic, not a classifier: the paper admits the pipeline cannot predict unlabelled images, so the central classification claim is unsupported.","rationale":"The reader's verdict of REJECT is sound, but the most load-bearing concern is not merely that the topological stability assumption fails; it is that the evaluation metric in Eq. 2 is not a classification metric at all. The paper's own Section 4 admission that the pipeline cannot predict unlabelled images directly contradicts the abstract and conclusion's classification claims. A list-overlap proportion computed on the full dataset cannot support 'automated detection' or 'classification' of rib fractures, regardless of whether the topological premise had held. The CNN baselines (F1 around 0.72-0.73) are standard and provide a reasonable comparison point, but they do not validate the HOMFLY classifier. The paper is transparent about several limitations, which is creditworthy, but transparency does not convert a descriptive overlap statistic into a predictive model. A leave-one-case-out test would settle the matter: if the polynomial lists from training images cannot label held-out images above chance, the central claim fails precisely as stated. I therefore leave the verdict unchanged, while noting that the reader's identified weakest assumption and my concern overlap substantially.","tokens_in":8710,"tokens_out":3225,"duration_ms":33237,"concrete_test":"Use the exact pipeline (pseudo-color, 13 kernels, 4 convolution cycles, HOMFLY computation) under a leave-one-case-out protocol: hold out all three images from one case, compute the class-specific lists of HOMFLY polynomials from the remaining cases, then assign each held-out image to the class whose list contains its polynomial (tie: undecided; absent from both lists: abstain). Measure accuracy, precision, recall, and abstention rate over all cases. To eliminate selection bias, the kernel subset and convolution-cycle choice must be re-derived inside each training fold using only training images. If held-out images are not classified at rates significantly above chance, or if most test polynomials are absent from training lists, the pipeline fails as a classifier. Also report the extent of within-class knot-type variability in the training folds to directly test the stability assumption.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim that HOMFLY polynomials classify rib fracture images is unsupported because no predictive classification was actually performed. Precision is defined in Eq. 2 as (K1-K12)/K1, the fraction of knot types in the fracture class list that are absent from the no-fracture list. This is a descriptive overlap measure over the entire dataset; it is not TP/(TP+FP), involves no train/test split, and provides no decision rule for a new image. Section 4 explicitly states: 'we could not use this pipeline for predicting unlabelled images.' The conclusion's claim 'we have shown how it can be applied to classify CT images of rib fractures' consequently overstates what was demonstrated. Additionally, the stated invariance assumption ('variations in size, color and shape ... translate into small geometrical variations ... two distinct objects will produce two distinct curves with different topological properties') is contradicted by the authors' own finding of 'more than one knot type for a given class'; thus the method reduces to list overlap rather than topology-based class identity. Hyperparameters (number of convolution cycles, 13 of 28 kernels, number of vertices) were chosen on the full dataset, so even the overlap statistic is optimistically biased. In short, the core claim requires that knot type be stable within class and predictive for unseen images, and neither condition is established.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a pipeline for automated detection of rib fractures in postmortem CT images by converting images into closed curves in RGB color space, computing HOMFLY polynomial knot invariants, and comparing the distribution of knot types between fracture and no-fracture classes. The authors report a 'precision' of 0.60 for the fracture class using this topological approach, alongside F1 scores of 0.72–0.73 for two convolutional neural network baselines (MobileNet and a custom Keras model). The paper claims in the abstract and conclusion that the method can classify CT images of rib fractures.","tokens_in":8991,"tokens_out":1923,"duration_ms":18301,"significance":"The application domain is relevant: automated rib fracture detection in forensic PMCT could reduce radiologist workload. The paper includes a legitimate comparison of two CNN baselines, with reported recall, precision, and F1 scores, which is a strength. However, the central novel claim—that HOMFLY polynomial knot invariants classify rib fracture images—is not supported by the experiments as presented. The reported 0.60 precision is not a predictive accuracy but a descriptive list-overlap statistic with no train/test split and no decision rule for new images. The paper's own Discussion (Section 4) explicitly states that the pipeline cannot be used for predicting unlabelled images. Consequently, the title, abstract, and conclusion overstate what was demonstrated. The work is at best a preliminary exploration of a topological descriptor, not a validated classifier.","major_comments":[{"comment":"The quantity called 'precision' for the HOMFLY approach is defined as (K1 - K12)/K1, where K1 is the number of knot types in one class and K12 is the number shared between classes. This is a measure of list disjointness, not classification precision in the sense of TP/(TP+FP). It is computed over the entire dataset without any train/test split, and the paper provides no decision rule that would assign a label to a new image. Section 4 explicitly states: 'we could not use this pipeline for predicting unlabelled images.' Therefore, the abstract's claim of a 'precision of 0.60' and the conclusion's statement that the method 'can be applied to classify CT images of rib fractures' are not supported.","section":"Eq. (2), Section 2.8, and Section 3"},{"comment":"The load-bearing invariance assumption stated in Section 4—that variations in size, color, and shape for the same object translate into small geometrical variations of the closed curve, so that distinct objects produce distinct curves with different topological properties—is contradicted by the authors' own finding: 'we ended up with more than one knot type for a given class.' This means the topological invariant is not stable within a class, and the method is reduced to comparing lists of knot types between two fixed collections of images. The claimed classification mechanism therefore does not hold, and the observed separation may be an artifact of the particular dataset rather than a generalizable property.","section":"Section 4"},{"comment":"The hyperparameters of the topological pipeline—the number of convolution cycles (4), the subset of 13 out of 28 kernels, and the use of 52 vertices—were selected using the full dataset. Equation (1) defines the kernel-selection criterion D as a function of the centroids of the two classes' feature clusters, and Figure 3 shows the resulting values for all 28 kernels. No held-out validation or cross-validation is performed for the topological pipeline. As a result, even the descriptive list-overlap statistic is optimistically biased, and no reliable estimate of generalization performance is provided.","section":"Sections 2.5 and 2.6"}],"minor_comments":[{"comment":"'To access the performance' should read 'To assess the performance'.","section":"Abstract"},{"comment":"There is a typographical error: 'HOMLFY polynomial' should be 'HOMFLY polynomial'.","section":"Section 2.7"},{"comment":"The word 'Euclidian' appears in Section 2.7 and elsewhere; the standard spelling is 'Euclidean'.","section":"Throughout"},{"comment":"The phrase 'we convolved each channel of the input images with a 3 × 3 kernel function ... followed by a downsampling using the softmax function' is unclear: softmax is not a downsampling operation. Please clarify the exact pooling or downsampling procedure used.","section":"Section 2.5"},{"comment":"The statement 'Using 52 vertices, we could classify all fractures' is ambiguous because the subsequent precision values indicate that knot types are shared between classes; please clarify what 'classify all fractures' means in this context.","section":"Section 3"}],"recommendation":"reject","confidential_remarks":"The paper's own Discussion section admits the pipeline cannot predict unlabelled images, which directly contradicts the abstract and conclusion's classification claims. This is not a matter of presentation alone; the core methodological contribution, as evaluated, does not constitute a classifier. The CNN baseline results are reasonable but are not the paper's claimed novelty. If the authors wish to pursue this line, they would need to reformulate the approach as a genuine predictive method with a validation set and a decision rule, or clearly reposition the work as a descriptive topological analysis."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The idea here is genuinely interesting: apply a knot polynomial (HOMFLY) to the coordinates produced by convolution kernels on rib CT images, and see if knot type separates fractures from non-fractures. I have not seen that specific pipeline before, and the paper deserves credit for attempting a topology-based alternative to black-box deep learning in medical imaging. The CNN baselines (F1 0.73) are standard and reasonable, and the authors are unusually candid about what they could not do.\n\nThe problem is that the central claim—\"we have shown how it can be applied to classify CT images of rib fractures\"—is not supported by the evidence. The precision of 0.60 reported for the HOMFLY method comes from Eq. 2: (K1 − K12)/K1, i.e., the fraction of knot types in the fracture class list that do not also appear in the no-fracture list. That is a descriptive overlap measure over the training set, not a predictive accuracy. There is no train/test split, no decision rule for a new image, and the authors explicitly state they could not use the pipeline for predicting unlabelled images. So the conclusion overstates what was demonstrated.\n\nThe paper also undermines its own invariance assumption. The logic is that variations in size, color, and shape should produce small geometric variations in the curve, preserving knot type within a class and distinguishing classes by knot type. But the authors report more than one knot type per class, which means the core stability assumption fails. The method degrades into comparing lists of polynomial types, which is weak evidence for classification.\n\nThe hyperparameter choices (number of convolution cycles, subset of 13 of 28 kernels, resolution) were made on the full dataset, so even the overlap statistic is optimistically biased. The small sample and the exclusion criteria also limit generalizability, but that is minor compared to the missing validation.\n\nWho is this for? Researchers curious about topological descriptors for image classification, and forensic radiologists interested in automating fracture detection. The idea is worth exploring, but the current paper is a proof of concept that does not validate a deployable classifier. It deserves a serious referee because the novelty is real and the authors are honest about limitations; I would send it to peer review, but expect major revisions or a rejection in current form. The authors should either reframe the work as an exploratory descriptive analysis or build a proper classifier with external validation.\n\nSerious thinker: yes—the paper is coherent on its own terms, and the limitations are stated plainly, even if the conclusions overreach.","headline":"Novel idea, honest write-up, but the reported HOMFLY 'classifier' is a descriptive list-overlap statistic, not a predictor.","tokens_in":9509,"tokens_out":1776,"would_cite":false,"duration_ms":17654,"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 claims the HOMFLY knot invariant can separate fractured from intact rib CT images at 0.60 precision, but cannot yet label new images.","keywords":["HOMFLY polynomial","knot theory","rib fracture detection","postmortem computed tomography","topological invariant","convolutional neural network","rib unfolding","image classification"],"falsifier":"Take a fresh cohort of rib CT images, run the same 13-kernel, 52-coordinate pipeline, and check whether each image's HOMFLY polynomial falls exclusively into the fracture or no-fracture list; if a substantial share of new images has a knot type present in both lists or in neither, the claimed classification breaks, and the paper's own statement that unlabelled images cannot yet be predicted already points toward this outcome.","tokens_in":8531,"feed_emoji":"🦴","tokens_out":8926,"duration_ms":80293,"temperature":0.7,"pith_summary":"This paper tries to establish that a topological invariant, the HOMFLY polynomial knot invariant, can serve as the classifier in a machine-learning pipeline for detecting rib fractures in postmortem computed tomography images. The authors unfold the rib cage into a 2D projection, convert the grayscale image to pseudo-color, apply hand-picked convolution kernels, and pool the feature maps until they obtain 52 coordinates that are joined into a closed curve in RGB color space; each image then receives a HOMFLY polynomial as its class fingerprint. On 195 cases, the method separates fracture images from no-fracture images with class-exclusivity precision of 0.60 for the fracture class and 0.52 for the no-fracture class, whereas two convolutional neural networks reach F1 scores of 0.72 and 0.73. The paper states that the pipeline cannot yet be used to predict unlabelled images, so the contribution is a demonstration of a topology-based class separation rather than a deployable diagnostic tool. A sympathetic reader should care because the approach offers an alternative to statistically trained networks, one with potentially exact classification if the knot-stability assumption can be made to hold.","feed_headline":"Knot polynomials split rib CTs at 0.60 precision","feed_subtitle":"A topology-only pipeline maps unfolded rib scans to HOMFLY knots and separates fracture cases, while CNN baselines reach F1 0.73.","key_machinery":"The load-bearing object is the HOMFLY polynomial, a two-variable knot invariant computed from the closed piecewise-linear curve formed by joining 52 coordinates in RGB color space. The coordinates are produced by taking a pseudo-colored, unfolded rib CT image, convolving each RGB channel separately with 13 hand-picked $3\\times3$ kernels, and softmax-pooling repeatedly until a $2\\times2$ feature map per channel supplies four coordinate values; 13 kernels times four coordinates gives the 52 vertices of the curve. The HOMFLY polynomial is then computed from this curve with a program supplied by reference [20], and each image is assigned the corresponding knot type. The mechanism works if geometrically similar images yield topologically equivalent curves, and the paper measures class separation by counting which knot types appear in only one of the two class lists.","core_discovery":"The central discovery, stated by the authors as a demonstration rather than a finished tool, is that rib CT images can be mapped to knot types whose HOMFLY polynomials separate, with measurable overlap, into a fracture list and a no-fracture list: precision is 0.60 for images with rib fractures and 0.52 for images without fractures, computed as the proportion of knot types not shared between classes. This separation is achieved without training any weights, using only the 52-coordinate curves derived from 13 convolution kernels on the pseudo-colored unfolded rib cage. The authors compare against two convolutional neural network baselines, which reach F1 scores of 0.72 and 0.73, and they report that the topology method cannot predict labels for unseen images because the current dataset does not cover the full catalogue of polynomials. On the paper's own terms, the result is an existence proof: a knot invariant can be inserted into an image-classification pipeline and yield a real, if modest, class signal.","pith_inferences":["A fair reading is that what is established is a class-separation statistic, not a classifier: the list-overlap rule cannot label a new image whose knot type appears in both lists or in neither, which the paper concedes.","A natural extension is to treat the HOMFLY polynomial type as a categorical feature and fit a small probabilistic model over knot-type distributions, instead of requiring strict exclusivity between the two class lists.","Using a coarser invariant (for instance, only the unknot versus a small set of low-crossing types) could mitigate the instability from multiple knot types per class; this direction is untested but follows directly from the paper's own observation."],"forward_implications":["If the topology-based separation holds on larger cohorts, rib-fracture screening could run without neural network training, using only hand-crafted kernels and a knot invariant.","The reported precision values set a measurable baseline: a larger dataset is expected to improve them, because the paper attributes the current ceiling to an incomplete catalogue of knot types.","Because both CNN baselines plateau around F1 0.72-0.73, the topology pipeline's 0.60 precision gives a non-statistical reference point for future automated fracture detection.","The same image-to-knot recipe should transfer to other grayscale medical images where a pathology manifests as stable shape or texture changes, since nothing in the pipeline is rib-specific."],"supporting_citations":[{"why":"Defines the HOMFLY polynomial invariant used throughout the pipeline.","marker":"[5]"},{"why":"Supplies the program that computes the polynomial from the coordinate curve.","marker":"[20]"},{"why":"Establishes the rib-unfolding CT visualization used to prepare the images.","marker":"[17]"},{"why":"Shows that single-plane rib reformations improve fracture reading, supporting the unfolding step.","marker":"[19]"},{"why":"Provides the neural network architecture used as one of the two comparison baselines.","marker":"[22]"},{"why":"Supplies the notation used to identify simple knot types from the polynomials.","marker":"[23]"},{"why":"Supplies the notation used to identify more complex knots.","marker":"[24]"}],"fun_headline_variants":["Knot theory separates rib fractures in CT scans","Topology-only rib CT screening hits 0.60 precision","Rib fracture detection via HOMFLY knots without training","Topological invariant shows class signal in rib CTs"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole method rests on the assumption that images from the same class produce closed curves with the same knot type, so that a single HOMFLY polynomial identifies the class; the paper's data already undermine this, since more than one knot type appears within a single class.","fun_headline_variants_meta":{"raw":{"variants":["Knot theory separates rib fractures in CT scans","Topology-only rib CT screening hits 0.60 precision","Rib fracture detection via HOMFLY knots without training","Topological invariant shows class signal in rib CTs"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00026,"raw_usage":{"total_tokens":1587,"prompt_tokens":937,"completion_tokens":650,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":553,"completion_tokens_details":{"reasoning_tokens":584}},"tokens_in":553,"tokens_out":650,"duration_ms":6668,"temperature":1.0,"reasoning_tokens":584,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T13:12:17.602038+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a fresh cohort of rib CT images, run the same 13-kernel, 52-coordinate pipeline, and check whether each image's HOMFLY polynomial falls exclusively into the fracture or no-fracture list; if a substantial share of new images has a knot type present in both lists or in neither, the claimed classification breaks, and the paper's own statement that unlabelled images cannot yet be predicted already points toward this outcome.","supporting_citations":[{"cited_title":"Freyd and D","cited_arxiv_id":null,"evidence_quote":"Defines the HOMFLY polynomial invariant used throughout the pipeline."},{"cited_title":"Ewing and K.C","cited_arxiv_id":null,"evidence_quote":"Supplies the program that computes the polynomial from the coordinate curve."},{"cited_title":"Ringl, M","cited_arxiv_id":null,"evidence_quote":"Establishes the rib-unfolding CT visualization used to prepare the images."},{"cited_title":"Dankerl, H","cited_arxiv_id":null,"evidence_quote":"Shows that single-plane rib reformations improve fracture reading, supporting the unfolding step."},{"cited_title":"Alexander and G.B","cited_arxiv_id":null,"evidence_quote":"Supplies the notation used to identify simple knot types from the polynomials."},{"cited_title":"Hoste, M","cited_arxiv_id":null,"evidence_quote":"Supplies the notation used to identify more complex knots."}],"review_version":1}