{"id":"126974ee-d28e-4422-8030-5830a75f94e5","arxiv_id":"2411.14833","paper_version":4,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"A point-based transformer framework tracks cells directly from raw microscopy frames, without segmentation, using event-guided sampling and rolling windows.","lead":"CAP is a one-stage cell tracking method that treats each cell as a point, removing the need for separate segmentation or detection steps, and runs 8 to 32 times faster than existing methods while matching their accuracy. It could make automated cell tracking in time-lapse microscopy much faster and cheaper for biology labs.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The one-stage claim is undermined by KIT-GE first-frame initialization, and the reported 8–32x speedups likely exclude this segmentation preprocess, so the central efficiency and one-stage claims are not yet supported.","rationale":"The reader's weakest assumption identified the external KIT-GE initialization as undermining the one-stage claim. My stress-test pass reaches the same conclusion and sharpens it: the issue is not only conceptual but also quantitative. The paper's own implementation details reveal that CAP's testing procedure begins with KIT-GE, a segmentation-based method, contradicting the central claim that detection/segmentation is eliminated. Moreover, the reported inference times for CAP are almost certainly tracking-only times, since they are an order of magnitude smaller than any full-pipeline baseline time in the same tables. Adding KIT-GE's runtime changes the efficiency comparison, which is one of the two pillars of the paper's contribution (the other being elimination of segmentation labels). I also considered the AEG sampling formula, which is confusing and potentially dimensionally inconsistent, and the contradictory code-availability statements, but these are secondary: they can be clarified without destroying the method's utility. The KIT-GE dependency directly affects the validity of the headline claims. This does not warrant rejection, because the underlying tracker may still work well when initialized appropriately; however, the claims must be restated and the timing methodology must be disclosed. Hence the reader's CONDITIONAL verdict is appropriate, and my read does not change it.","tokens_in":21430,"tokens_out":4249,"duration_ms":45037,"concrete_test":"Instrument the released CAP code to measure end-to-end wall-clock time from raw input frames to final tracking output, explicitly including the KIT-GE first-frame preprocessing used to select query points Q. Recompute the speedup ratios in Tables 2 and 3 using this end-to-end time rather than CAP's tracking-only time. If the speedup drops below the claimed 8–32x, or CAP becomes slower than KIT-GE, the efficiency claim must be revised. As a secondary check, rerun CAP with first-frame query points replaced by centroids derived directly from the TRA ground-truth masks (where available) and report the change in TRA; if accuracy changes materially, the dependency on the external segmentation step is confirmed.","verdict_should_be":"UNCHANGED","load_bearing_attack":"In Section 4.1.1 (Implementation Details), the authors state: 'As for the testing (evaluation) process, we preprocess the first frame for each sequence using [70] to select query points Q and start the process of window rolling.' Reference [70] is KIT-GE, a CNN-based distance-prediction segmentation and graph-matching method. This is precisely the kind of explicit segmentation/detection stage that the Abstract, Figure 1, and Section 3 claim CAP eliminates. At inference time, CAP has no internal mechanism to obtain the initial cell locations; it relies on an external segmentation-based detector for the first frame. Thus the pipeline is not a true one-stage tracker: it is KIT-GE-based initialization followed by CAP trajectory refinement. This is not merely a semantic issue. The reported inference times for CAP (e.g., 1.3 s on HeLa, 2.6 s on U373) appear to exclude the KIT-GE preprocessing pass, while the baseline inference times (e.g., KIT-GE itself: 14.4 s on HeLa, 22.0 s on U373) are full-pipeline times. If KIT-GE's runtime is added to CAP's, the claimed 8–32x speedup shrinks dramatically and in several datasets CAP would become slower than KIT-GE's full pipeline. The paper's Limitations section acknowledges annotation-style sensitivity but does not disclose this external dependency, and Section 5 still claims CAP 'eliminates the need for a separate segmentation or detection stage.' This makes the strongest claim as stated inaccurate and the efficiency comparison potentially misleading.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper proposes CAP (Cell as Point), a one-stage cell tracking framework that treats each cell as a point and uses a transformer with cross-trajectory/time attention to jointly predict cell trajectories and visibilities. The method introduces adaptive event-guided (AEG) sampling to address division-event imbalance and a rolling-as-window (RAW) inference strategy for long sequences. The authors evaluate on DeepCell and ISBI CTC benchmarks, reporting competitive TRA scores (e.g., 0.985 on U373) and 8-32x inference speedups over existing methods, with claimed elimination of detection/segmentation stages and reduced annotation requirements.","tokens_in":21748,"tokens_out":7388,"duration_ms":65697,"significance":"Strengths: the paper evaluates on multiple public benchmarks with standard TRA metrics, provides ablations for the main components, and makes code/checkpoints available. If the one-stage claim and efficiency comparison were substantiated, CAP would be a practically useful contribution to cell tracking. However, the central claims are not yet supported: the testing procedure relies on an external segmentation-based method (KIT-GE) for first-frame initialization, and the reported inference times appear to exclude this preprocessing, undermining the 'one-stage' and speedup claims. The AEG probability formula is also internally inconsistent with its stated purpose. These issues are addressable but require substantive revision.","major_comments":[{"comment":"The testing process states: 'we preprocess the first frame for each sequence using [70] to select query points Q and start the process of window rolling.' Reference [70] is KIT-GE, a CNN-based distance-prediction segmentation and graph-matching method. This external dependency directly contradicts the central claim that CAP 'eliminates the need for explicit detection or segmentation' and the conclusion's statement that CAP 'eliminates the need for a separate segmentation or detection stage.' Moreover, the inference times reported for CAP in Tables 2 and 3 (e.g., 1.3 s on HeLa) appear to exclude the KIT-GE preprocessing pass, while the baseline times are full-pipeline times. The claimed 8-32x speedup is therefore not an end-to-end comparison. Please report full end-to-end inference time including initialization, or provide an internal initialization mechanism, and revise the claims accordingly.","section":"Section 4.1.1 (Implementation Details); Section 5"},{"comment":"The probability of applying AEG is defined as PAEG = Ndiv×Tdiv/T, where Ndiv is the total number of divisions in the sequence and Tdiv is the duration of a division event. This quantity is not guaranteed to be in [0,1]: on HeLa, for instance, 189 divisions with Tdiv on the order of a few frames and T=42 would give PAEG>1. Conversely, on a dataset with very few divisions (PC-3 has 5 divisions in 50 frames), PAEG becomes small, so AEG sampling would rarely be triggered precisely when division events are rarest. This is the opposite of the stated goal of 'prioritizes cell division events.' Please clarify the intended formula, add normalization or an alternative definition, and demonstrate that the sampling actually increases the proportion of division-containing windows.","section":"Section 3.2.1, Eq. for P_AEG"},{"comment":"The inner loop 'for i←0 to lwin do tcur ← tcur + i; ...' increments tcur by i for each i, so after the loop tcur advances by lwin(lwin+1)/2 rather than by lwin. This makes the algorithm as written non-reproducible and inconsistent with the text, which says the window slides frame-by-frame. Please correct the pseudocode to tcur ← tcur + 1 (or an equivalent step) and ensure the 'find new cell(s)' condition is defined.","section":"Algorithm 1 (RAW inference)"},{"comment":"All TRA values are reported as single numbers without error bars, confidence intervals, or multiple runs. Several comparisons are close (e.g., 3T3: CAP 0.854 vs GNN 0.857; GOWT1: CAP 0.960 vs KIT-GE 0.966), so the claims of 'consistent improvements' and 'promising performance' cannot be assessed statistically. Please provide variance estimates or multiple-seed results, at least for the main comparisons.","section":"Tables 2-6"}],"minor_comments":[{"comment":"The abstract states that 'The code and model checkpoints will be available soon,' while the paper header lists a GitHub URL and says the code is available. Please make these statements consistent.","section":"Abstract vs. Section 4.1.1"},{"comment":"The notation '2s' in the correlation volume formula is ambiguous; the text describes a kernel size of 2^s × 2^s, so the equation should use superscripts consistently.","section":"Equation (2)"},{"comment":"The columns and numbers in Table 4 are not aligned correctly; for example, the GOWT1 row appears to contain '0.8830.9210.878' with no separators, and the sequence labels do not line up with the values. Please reformat.","section":"Table 4"},{"comment":"The caption states '2.9s' as the inference time, but no dataset in Tables 2 or 3 has exactly 2.9 s; please clarify whether this is an average and over which datasets.","section":"Figure 1(b)"},{"comment":"The term 'T anchors' is introduced without definition; please define the anchor set and explain how it is used in sampling.","section":"Section 3.2.1"},{"comment":"The condition 'find new cell(s)' is not defined; the model's visibility output presumably yields new cells, but the pseudocode should specify how this is determined from the predicted visibilities.","section":"Algorithm 1"}],"recommendation":"major_revision","confidential_remarks":"This manuscript is within the scope of the journal, but the main claims overstate the contribution. The external initialization dependency and the timing comparison should be addressed in revision. I have no concerns about citation integrity."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: CAP adapts CoTracker to cell tracking with a genuinely nice representation—cell point trajectory plus three visibility slots for mother and two daughters—and the ablations show each component earns its keep. The TRA numbers on U373 (0.985) and HeLa (0.926) are strong. But the paper's headline claim that it 'eliminates the need for explicit detection or segmentation' is not true as stated. Section 4.1.1 says the first frame is preprocessed with KIT-GE [70] to select query points Q. KIT-GE is a CNN-based distance-prediction segmentation plus graph-matching method. So at inference, CAP is effectively KIT-GE initialization followed by trajectory refinement—a two-stage pipeline, however you frame it. Worse, the reported inference times (1.3 s on HeLa, 2.6 s on U373) appear to exclude the KIT-GE pass, while the baseline times (KIT-GE itself: 14.4 s on HeLa, 22.0 s on U373) are full-pipeline. Adding KIT-GE's runtime to CAP's would shrink the claimed 8–32x speedup dramatically; on some datasets CAP would be slower than KIT-GE. The Limitations section acknowledges annotation-style sensitivity but never discloses this external dependency, and the conclusion repeats the 'eliminates separate segmentation/detection' claim. That is load-bearing, not cosmetic.\n\nSecond issue: the AEG sampling probability formula P_AEG = N_div * T_div / T. If divisions are rare, N_div is small, so P_AEG is small, meaning AEG is rarely used. That seems backwards for a method whose purpose is to prioritize rare division events. The ablation shows AEG helps a lot, so either the formula is misstated or the intent is different. Needs clarification. Also: no error bars; the abstract says code is available at a URL while the same paragraph ends 'will be available soon'; and the training uses only TRA masks (good) but the first-frame initialization requires a segmentation-based detector (bad for the label-light claim).\n\nWhat is genuinely good: the representation embeds division into a trajectory predictor cleanly, the RAW inference strategy for long sequences is sensible, and the experiments cover multiple datasets with standard TRA metrics and informative ablations. This is a real contribution to cell tracking, just not the one-stage revolution the text claims.\n\nWho is this for? Anyone working on cell tracking or point-tracking adaptations. It deserves a serious referee, but the authors need to fix the initialization disclosure, rerun the efficiency comparison with the full pipeline, and correct or clarify the AEG formula. As is, the central claims are not supported as written.","headline":"CAP is a sensible CoTracker adaptation for cell tracking with a clean division representation, but the 'one-stage / no segmentation' claim falls apart on the KIT-GE first-frame initialization and the speedups are likely overstated.","tokens_in":22236,"tokens_out":3239,"would_cite":false,"duration_ms":31386,"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":"A one-stage cell tracker that treats each cell as a point matches multi-stage pipelines at 8–32x lower inference cost.","keywords":["cell tracking","one-stage tracking","cell division","lineage reconstruction","point tracking","adaptive event-guided sampling","rolling-as-window inference","time-lapse microscopy"],"falsifier":"Run CAP on the same ISBI sequences with first-frame query points initialized from (a) the segmentation-based output, (b) centroids of the tracking ground-truth masks, and (c) the same points shifted by a few pixels of Gaussian noise; if TRA collapses under (b) or (c), the framework's independence from a detection or segmentation stage is not established.","tokens_in":21251,"feed_emoji":"🔬","tokens_out":7812,"duration_ms":68201,"temperature":0.7,"pith_summary":"CAP is proposed as an end-to-end one-stage alternative to the standard detect-then-link or segment-then-link cell tracking pipelines. The paper's central claim is that representing every cell as a point trajectory, together with a visibility flag and three output slots for a mother and two potential daughter cells, is enough to capture migration, division, and lineage in a single trainable stage. If this claim holds, the practical payoff is large: laboratories can train and run cell trackers using only coarse tracking ground truth, without producing high-quality segmentation masks, and can expect inference times of seconds rather than tens or hundreds of seconds on standard benchmarks. The paper supports the claim with TRA scores up to 0.985 on U373 and an 8–32x inference speedup relative to multi-stage baselines.","feed_headline":"Cell tracking in one stage, 8–32x faster than current pipelines","feed_subtitle":"Treating each cell as a tracked point skips segmentation and reaches a tracking accuracy of 0.985 on U373.","key_machinery":"The load-bearing representation is the cell point trajectory with visibility: each cell is a point $(x,y)$, a binary visibility flag marks existence, and each predicted point owns three location slots for the mother cell and two possible daughter cells, so division is represented natively. Association is carried by 4D correlation volumes between tracking features and multi-scale image features around estimated locations, refined iteratively by a transformer with cross-trajectory/time attention. Around this core, AEG sampling selects training clips that contain complete division events, and RAW inference processes long sequences by rolling a window and appending newly born cells to the query set.","core_discovery":"The paper's central discovery is that a point-based sequence model can jointly track all cells in a microscopy video by iteratively refining their trajectories and visibilities. The model predicts, for each tracked cell, its own location and the locations of two potential daughter cells, updated through a transformer with cross-trajectory/time attention and RAFT-style correlation volumes. Two mechanisms make this practical: AEG sampling forces the training sequence to include complete division events, countering the rarity of mitosis, and RAW inference rolls a fixed-size window forward, inserting newly appearing cells as new query points in long sequences. On the DeepCell and ISBI CTC benchmarks, the framework reaches the best or competitive tracking accuracy among the tested methods, including zero structural errors and a TRA of 0.985 on U373, while requiring only 1.1–7.1 seconds per sequence.","pith_inferences":["Editorial inference: the framework's one-stage status is conditional, because first-frame cell locations come from an external segmentation-based method; replacing that initialization with cheap centroid extraction or with tracking ground-truth masks is a direct test of how much of the staging claim actually carries.","Editorial inference: the same point-trajectory machinery may transfer to other biological imaging tasks with division and merging events, such as bacterial colony or organelle tracking, where mask-level annotation is the bottleneck.","Editorial inference: AEG sampling could be generalized from division events to other rare but decisive cell behaviors, such as apoptosis, and its probability schedule could be made adaptive to per-sequence event rates rather than a dataset-global statistic."],"forward_implications":["Training no longer requires segmentation masks; tracking ground truth (coarse masks and lineage graphs) suffices.","Inference time drops to a few seconds per sequence, which is an order of magnitude faster than the compared multi-stage systems.","Rare cell divisions can be learned reliably by biasing the sampled clips toward complete division events.","Long sequences can be tracked with a fixed-size window, so memory and compute no longer grow with full sequence length.","The ablations indicate that cross-trajectory attention and a feature stride of 4 are both necessary for the reported accuracies."],"supporting_citations":[{"why":"Supplies the pretrained joint point-tracking transformer and weights that CAP adapts to grayscale cell microscopy.","marker":"[43]"},{"why":"Contributes the 4D correlation volume and iterative update machinery used to associate points across frames.","marker":"[76]"},{"why":"Provides the segmentation-based first-frame cell localization used to initialize query points, so the method's independence from segmentation depends on this step.","marker":"[70]"},{"why":"Supplies the DeepCell HeLa, PC-3, 3T3, and RAW264 datasets and their tracking ground truth.","marker":"[59]"},{"why":"Defines the ISBI Cell Tracking Challenge benchmark datasets and evaluation protocol used for U373, GOWT1, and Huh7.","marker":"[57]"},{"why":"Defines the AOGM-based TRA measure used to score the reported tracking accuracy.","marker":"[58]"},{"why":"Provides the earlier ISBI challenge datasets and official evaluation methodology that ground the comparison.","marker":"[80]"}],"fun_headline_variants":["Cell tracking: point-based one-stage, up to 32x faster","Skip segmentation: one-stage cell tracking via points","One-stage cell tracking: 8-32x faster, no segmentation","Cell as point: one-stage tracking with 32x speedup"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The claim that the pipeline is one-stage and bypasses detection and segmentation depends on treating the first-frame cell locations, which are produced by an external segmentation-based method, as harmless initialization; if those points are inaccurate or if that step counts as a detection or segmentation stage, the central claim weakens.","fun_headline_variants_meta":{"raw":{"variants":["Cell tracking: point-based one-stage, up to 32x faster","Skip segmentation: one-stage cell tracking via points","One-stage cell tracking: 8-32x faster, no segmentation","Cell as point: one-stage tracking with 32x speedup"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000229,"raw_usage":{"total_tokens":1481,"prompt_tokens":952,"completion_tokens":529,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":568,"completion_tokens_details":{"reasoning_tokens":456}},"tokens_in":568,"tokens_out":529,"duration_ms":4981,"temperature":1.0,"reasoning_tokens":456,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T14:49:04.767925+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run CAP on the same ISBI sequences with first-frame query points initialized from (a) the segmentation-based output, (b) centroids of the tracking ground-truth masks, and (c) the same points shifted by a few pixels of Gaussian noise; if TRA collapses under (b) or (c), the framework's independence from a detection or segmentation stage is not established.","supporting_citations":[{"cited_title":"Co- Tracker: It is better to track together","cited_arxiv_id":null,"evidence_quote":"Supplies the pretrained joint point-tracking transformer and weights that CAP adapts to grayscale cell microscopy."},{"cited_title":"Cell segmentation and tracking using cnn-based dis- tance predictions and a graph-based matching strategy.PLoS One, page e0243219, 2020","cited_arxiv_id":null,"evidence_quote":"Provides the segmentation-based first-frame cell localization used to initialize query points, so the method's independence from segmentation depends on this step."},{"cited_title":"Accurate cell tracking and lineage construction in live-cell imaging ex- periments with deep learning.Biorxiv, page 803205, 2019","cited_arxiv_id":null,"evidence_quote":"Supplies the DeepCell HeLa, PC-3, 3T3, and RAW264 datasets and their tracking ground truth."},{"cited_title":"The cell tracking challenge: 10 years of objective benchmarking.Nature Methods, pages 1010–1020, 2023","cited_arxiv_id":null,"evidence_quote":"Defines the ISBI Cell Tracking Challenge benchmark datasets and evaluation protocol used for U373, GOWT1, and Huh7."},{"cited_title":"Cell tracking accuracy measurement based on comparison of acyclic oriented graphs.PloS one, page e0144959, 2015","cited_arxiv_id":null,"evidence_quote":"Defines the AOGM-based TRA measure used to score the reported tracking accuracy."},{"cited_title":"An objective comparison of cell-tracking algorithms.Nature methods, pages 1141–1152, 2017","cited_arxiv_id":null,"evidence_quote":"Provides the earlier ISBI challenge datasets and official evaluation methodology that ground the comparison."}],"review_version":1}