{"id":"e32ac1e7-67f9-4a30-b305-28e8cf6e5f66","arxiv_id":"2504.21544","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"SAM4EM, a prompt-free SAM2 adapter with LoRA, multi-scale feature fusion, and memory attention, reports improved Dice scores over H-SAM and SAMed on four EM benchmarks, including a new glia/synapse dataset.","lead":"This paper adapts Meta's Segment Anything Model to automatically segment 3D electron microscopy images of brain tissue, focusing on mitochondria, glia, and synapses. It combines LoRA fine-tuning, a multi-scale feature enhancer, and a lightweight memory attention mechanism to segment without manual prompts.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The claimed gains on glia and synapses rest entirely on an unreleased, self-annotated benchmark, so the central SOTA claim is not yet independently verifiable.","rationale":"The reader identified the same load-bearing weakness: the headline performance on glia and synapses depends entirely on a novel, unreleased, self-annotated dataset. My reading of the full text confirms this. Section 5 describes the dataset in one paragraph and defers details to absent supplementary materials; Section 6 reports the large synapse gain without error bars or significance testing; and Table 1 has no ViT-based SOTA baselines despite the related-work section naming FragViT and ATFormer as state-of-the-art mitochondria methods. The Lucchi metric swap is a concrete internal inconsistency that further reduces confidence in the numeric reporting, though it is not by itself fatal to the method. The architecture description is plausible and the ablation study on Lucchi is internally coherent, so the paper merits conditional acceptance rather than rejection. The reader's conditional verdict is appropriate, and my concern does not move it.","tokens_in":12291,"tokens_out":2093,"duration_ms":25448,"concrete_test":"Release the Mice dataset with explicit train/validation/test splits and the full annotation protocol, then independently re-run H-SAM and SAM4EM on the identical split using at least three random seeds and report Dice/mIoU with standard deviations. If the 11.5-point synapse gap and 1.8-point glia gap persist under these controlled conditions, the central claim is supported; if the gaps shrink to statistical noise, the claim fails. As a secondary check, recompute the Lucchi Dice and mIoU from the released predictions to resolve the swapped values in Section 6 versus Table 1.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The abstract and Section 1 claim that SAM4EM 'outperforms existing solutions in the segmentation of complex processes like glia and post-synaptic densities,' and Table 1 is summarized as 'consistently outperforms baseline approaches across all datasets.' The decisive evidence for that claim is the 1.8-point Dice gain on Mice-Glia and the 11.5-point gain on Mice-Synapses over H-SAM. Both numbers come only from the new 'Mice-' dataset, which the authors curated from the stack in [2] and annotated themselves. Section 5 says 'Additional details about the novel released dataset can be found in the supplementary materials,' but the supplementary material is not present, no dataset release link appears in the main text (despite the abstract saying the dataset will be released), and no annotation protocol, split definition, or inter-annotator agreement is given. Since the benchmark is self-created, self-annotated, and unreleased, the reported advantage on the very structures central to the paper could be an artifact of annotation bias, label leakage, or the specific train/test split. The issue is compounded by the Lucchi metric swap in Section 6 (text reports Dice 86.1 and mIoU 92.4, while Table 1 reports Dice 92.4 and mIoU 86.1), which undermines confidence in the reported numbers, and by the absence of the stronger ViT baselines (e.g., FragViT, ATFormer) that the paper itself discusses as SOTA for mitochondria. These are addressable, but until the dataset and code are released and independently re-run, the central claim is not robustly supported.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes SAM4EM, a prompt-free adapter built on SAM2 for 3D electron microscopy segmentation of mitochondria, glia, and synapses. The architecture combines LoRA-based fine-tuning of the image encoder, a multi-scale feature enhancer, a lightweight memory encoder with attention and exponential moving average updates (Eqs. 1–2), and a bi-directional self-prompting two-stage decoder. Experiments compare against H-SAM, SAMed, and UN-SAM on the Lucchi dataset and a newly curated 'Mice-' dataset, reporting consistent improvements in Table 1, together with a component ablation in Table 2.","tokens_in":12604,"tokens_out":5271,"duration_ms":51672,"significance":"If the reported results are reproducible and the dataset is made available, the paper would be a useful contribution: it presents one of the first SAM adapters specifically for EM neuroscience data, with parameter-efficient fine-tuning and a new benchmark for complex structures. The architecture is described in enough detail to be implemented, and the main equations are clearly specified, with no equation-level circularity. However, the central claim of consistent state-of-the-art performance currently rests on an unreleased, self-annotated dataset, missing ViT baselines, and an internal metric contradiction, so the significance cannot yet be fully assessed.","major_comments":[{"comment":"The text states that SAM4EM achieves 'a Dice score of 86.1% and an exceptional mIoU of 92.4%' on Lucchi, while Table 1 reports Dice=92.4 and mIoU=86.1 for the same entry. The surrounding sentence claims a 'meaningful improvement over H-SAM' whose Lucchi Dice is 90.5; that claim is coherent only if SAM4EM's Dice is 92.4. Please correct the text to match Table 1 and re-verify all reported numbers for consistency.","section":"Section 6, Lucchi paragraph"},{"comment":"The novel 'Mices-' dataset is curated by the authors from the EM stack in [2] and annotated by the authors themselves. The main text provides no dataset release link, annotation protocol, split definition, or inter-annotator agreement, and the supplementary material is not present. Because the claimed gains on glia (1.8 Dice points) and synapses (11.5 Dice points) over H-SAM come exclusively from this dataset, the paper's central claim of consistent state-of-the-art performance is not independently verifiable. The authors should either release the dataset and code (as the abstract promises) or provide full annotation and evaluation details in the main text.","section":"Section 5, Datasets"},{"comment":"The manuscript says 'the model's performance was optimized through random grid search over hyperparameters' and that the resulting alpha=0.3 and M=8 memory slots were used for the results in Table 1. If this grid search was performed on the same datasets whose test metrics are reported, then the reported numbers are optimistically biased by hyperparameter selection on the test data. Please specify the validation procedure used for hyperparameter selection, or report results obtained with hyperparameters fixed before any test evaluation.","section":"Section 5, hyperparameters"},{"comment":"The Related Work section discusses ATFormer [24] and FragViT [20] as state-of-the-art vision-transformer methods for mitochondria segmentation, and the abstract claims improvement over 'other vision transformer-based approaches.' However, Table 1 compares only with H-SAM, SAMed, and UN-SAM. Please add the ViT-based mitochondria baselines to Table 1 (or explain why they cannot be compared) and adjust the claims accordingly, since the current comparison set does not support the abstract's broad claim.","section":"Section 2 and Table 1"}],"minor_comments":[{"comment":"The dataset is referred to as 'Mices-' in Section 5 but as 'Mice-' in Section 6 and Table 1; please use one consistent name throughout.","section":"Section 5, dataset naming"},{"comment":"The text says 'While UN-SAM results were unavailable for comparison,' but Table 1 reports UN-SAM results for all datasets. Please clarify the intended meaning or remove the contradiction.","section":"Section 6, Overall performance"},{"comment":"The ablation study runs only 20 epochs on Lucchi and reports single mIoU values without variance, while Table 1 reports the full model at 200 epochs. Please report mean and standard deviation over multiple runs and state the training schedule explicitly, so the ablation can be compared with the main results.","section":"Section 6, Ablation experiments"},{"comment":"There are several typos: 'synapsis' should be 'synapses,' 'we plane to extend' should be 'we plan to extend,' and 'focussed ion beam' should be 'focused ion beam.'","section":"Throughout"}],"recommendation":"major_revision","confidential_remarks":"The paper's central claim hinges on the new Mice- dataset. I strongly encourage the editor to require dataset and code release, or at least detailed annotation and split descriptions, before accepting. The hyperparameter search description should also be clarified to rule out test-set selection. If the authors cannot provide these, the claims should be substantially softened. The comparison set is also too narrow to support 'outperforms existing solutions' as stated."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague, here's my take on arXiv:2504.21544.\n\nThe genuinely new thing is the package: LoRA-adapted SAM2 image encoder, multi-scale feature enhancer, lightweight memory attention for z-consistency, and two-stage self-prompting decoding. Individually these are borrowed from LoRA, H-SAM, SAM2, but the combination for prompt-free EM segmentation is not something I've seen. The second contribution is the Mice- dataset for glia and synapses, which fills a real gap in the field—if it actually materializes.\n\nWhat the paper does well: the architecture is clearly motivated, the medical-image rationale for multi-scale features and memory propagation is sensible, and the ablation on Lucchi (though short) shows each module adds something. The writing is straightforward and the related work is adequate.\n\nThe soft spots are real and mostly concern evidence rather than mechanism.\n\nFirst, the central SOTA claim—outperforming in glia and synapses—rests entirely on the new Mice- dataset, which is not available: no release link, no supplementary, no annotation protocol, no split definition. The abstract says the dataset will be released, but the main text only points to missing supplementary material. The 1.8-point Dice gain on glia and 11.5-point gain on synapses are the only evidence for the flagship claim, and they could easily be artifacts of annotation choices, leakage, or split luck. Until the dataset ships, I'd treat those numbers as unverified.\n\nSecond, there are internal contradictions. Section 6 reports Lucchi Dice 86.1 / mIoU 92.4, while Table 1 says Dice 92.4 / mIoU 86.1. One is simply wrong. Also, the text says UN-SAM results were unavailable, yet UN-SAM appears in Table 1 with numbers. That's sloppy reporting, and it undercuts confidence in the table.\n\nThird, the paper claims to outperform \"vision transformer-based approaches\" but Table 1 only compares against SAM-family methods. ATFormer and FragViT, which the paper itself calls SOTA for mitochondria, are missing. The claim is overreaching until those are included.\n\nFourth, the ablation runs 20 epochs on one dataset with no error bars, and the memory hyperparameters (alpha, M) were tuned on the same data. Minor, but worth flagging.\n\nNone of these are load-bearing architectural flaws. The method is plausible and the work is honest in intent. But the evidence as presented is not enough. I'd send it to peer review with a request for the dataset, supplementary, corrected numbers, and ViT baselines. If the authors deliver those, it becomes a decent incremental contribution for the EM segmentation community.\n\nRecommendation: engage with it, but demand the data before believing the synapse gains.","headline":"SAM4EM is a plausible engineering package, but its headline SOTA numbers sit on an unreleased self-annotated dataset plus clear reporting contradictions, so treat the claims as provisional.","tokens_in":13176,"tokens_out":3285,"would_cite":false,"duration_ms":30719,"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":"SAM4EM claims that a prompt-free, memory-aware adapter for the Segment Anything Model outperforms existing methods on 3D electron-microscopy segmentation of glia and synapses.","keywords":["electron microscopy segmentation","segment anything model","prompt-free segmentation","low-rank adaptation","LoRA","3D memory attention","glia segmentation","synapse segmentation"],"falsifier":"Once the Mice- dataset is made public, rerun the head-to-head comparison with independently produced expert annotations for glia and synapses; if SAM4EM's reported 11.5-point Dice advantage over H-SAM on synapses (53.8 vs 42.3) does not reproduce, the paper's central claim collapses.","tokens_in":12099,"feed_emoji":"🧠","tokens_out":8401,"duration_ms":83240,"temperature":0.7,"pith_summary":"This paper aims to establish that a prompt-free adapter built on the Segment Anything Model (SAM) can segment complex 3D electron-microscopy brain structures—especially glia and synapses—more accurately than existing SAM-based and transformer baselines. The method replaces manual prompts with a two-stage self-prompting loop that uses the previous slice's mask, adds low-rank (LoRA) fine-tuning for efficiency, and a memory attention mechanism for slice-to-slice consistency. The authors back the claim with experiments on mitochondria, glia, and synapses, plus a newly curated annotated dataset for the latter two structures. If the claim holds, neuroscience labs could segment difficult neural structures automatically with modest GPU budgets.","feed_headline":"No-prompt SAM4EM outperforms baselines on brain EM stacks","feed_subtitle":"LoRA fine-tuning plus two-stage decoding and 3D memory delivers the largest gains on synapses.","key_machinery":"The central object is the bi-directional self-prompting loop combined with an efficient 3D memory encoder. The first prompt encoder takes the previous slice's predicted mask as a noisy prompt; the first decoder generates a coarse prior mask; a second prompt encoder and decoder refine it, creating an iterative refinement loop. The efficient memory encoder propagates context along the z-axis by combining current slice features with previous mask features and updating memory slots through an exponential moving average, $M_t = (1-\\alpha)M_{t-1} + \\alpha F_t$, with $\\alpha = 0.3$ and eight memory slots. A multi-scale feature enhancer fuses 1/4, 1/8, and 1/16 resolution features to preserve fine boundary details, and LoRA bypass paths keep the image encoder mostly frozen, reducing trainable parameters by about 85%.","core_discovery":"SAM4EM's core claim is that a two-stage prompt-free adapter, built on SAM2 and fine-tuned with LoRA, can match or beat specialized baselines on well-bounded structures such as mitochondria and clearly beat them on irregular, boundary-poor structures such as astrocytic processes and post-synaptic densities. On the curated Mice- dataset the paper reports Dice scores of 70.5 for glia, 80.7 for mitochondria, and 53.8 for synapses, and 92.4 Dice on the public Lucchi mitochondria benchmark; in Table 1 these figures place SAM4EM ahead of H-SAM, SAMed, and UN-SAM across all datasets. The largest reported advantage is 11.5 Dice points over H-SAM on synapses.","pith_inferences":["If the memory attention generalizes the way it does in video foundation models, the same adapter should transfer to 2D+t microscopy or video object segmentation where structures deform between frames rather than between slices.","The claimed 11.5-point Dice improvement on synapses is concentrated on a dataset the authors annotated themselves, so an independent expert re-annotation and public release of the splits would turn that number into a reproducible benchmark result.","A natural next experiment is to reverse the slice order or initialize memory from a different slice; if segmentation quality stays the same, the memory mechanism is storing stable object context rather than just a moving average of nearby masks."],"forward_implications":["Neuroscience labs could segment entire EM volumes, including the 448-slice Mice- stack, without clicking prompt points, because the model generates its own prompts from the previous slice's mask.","The LoRA-based adapter cuts trainable parameters by about 85% and fits in roughly 4 GB of GPU memory during training, making foundation-model fine-tuning practical on a single consumer GPU.","The two-stage mask decoding plus memory attention is a general volumetric recipe that can be carried to other biomedical volumes where structures change gradually between slices.","If the glia and synapse benchmark is released as stated, it would give the community a first standard dataset for comparing SAM-based adapters on these difficult structures."],"supporting_citations":[{"why":"Supplies the Segment Anything Model itself, the promptable foundation that SAM4EM adapts.","marker":"[15]"},{"why":"Introduces the memory encoder and memory attention mechanism that SAM4EM repurposes for slice-to-slice 3D consistency.","marker":"[28]"},{"why":"Provides the Low-Rank Adaptation technique used to fine-tune the frozen image encoder with few trainable parameters.","marker":"[11]"},{"why":"The prompt-free hierarchical-decoding baseline on the same benchmarks and the source of the second-stage hierarchical pixel decoder design.","marker":"[6]"},{"why":"The prior EM study whose stacks were curated into the new Mice- benchmark for glia, mitochondria, and synapses.","marker":"[2]"},{"why":"The public Lucchi FIB-SEM mitochondria dataset used as a standard benchmark and for the ablation study.","marker":"[19]"},{"why":"SAMed is a SAM-based medical-imaging baseline compared in Table 1.","marker":"[36]"},{"why":"UN-SAM is a prompt-free baseline for nuclei included in Table 1.","marker":"[5]"}],"fun_headline_variants":["SAM4EM: prompt-free SAM2 adapter wins on synapses in EM","LoRA-tuned SAM4EM beats SOTA on glia and synapses","Prompt-free SAM4EM tops benchmarks for complex 3D EM stacks","SAM4EM's two-stage decoder gives 11.5 Dice gain on synapses"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The entire superiority claim for glia and synapse segmentation is measured on a newly curated dataset whose annotations were made by the authors and whose release details, annotation protocol, and splits are not given in the main text, so the results depend on those annotations being correct and representative.","fun_headline_variants_meta":{"raw":{"variants":["SAM4EM: prompt-free SAM2 adapter wins on synapses in EM","LoRA-tuned SAM4EM beats SOTA on glia and synapses","Prompt-free SAM4EM tops benchmarks for complex 3D EM stacks","SAM4EM's two-stage decoder gives 11.5 Dice gain on synapses"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000904,"raw_usage":{"total_tokens":3869,"prompt_tokens":902,"completion_tokens":2967,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":518,"completion_tokens_details":{"reasoning_tokens":2885}},"tokens_in":518,"tokens_out":2967,"duration_ms":21266,"temperature":1.0,"reasoning_tokens":2885,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T05:00:51.604937+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Once the Mice- dataset is made public, rerun the head-to-head comparison with independently produced expert annotations for glia and synapses; if SAM4EM's reported 11.5-point Dice advantage over H-SAM on synapses (53.8 vs 42.3) does not reproduce, the paper's central claim collapses.","supporting_citations":[{"cited_title":"Berg, Wan-Yen Lo, Piotr Dollar, and Ross Girshick","cited_arxiv_id":null,"evidence_quote":"Supplies the Segment Anything Model itself, the promptable foundation that SAM4EM adapts."},{"cited_title":"Unleashing the potential of sam for medical adaptation via hierarchi- cal decoding","cited_arxiv_id":null,"evidence_quote":"The prompt-free hierarchical-decoding baseline on the same benchmarks and the source of the second-stage hierarchical pixel decoder design."},{"cited_title":"The effects of aging on neuropil structure in mouse somatosensory cor- tex—a 3d electron microscopy analysis of layer 1","cited_arxiv_id":null,"evidence_quote":"The prior EM study whose stacks were curated into the new Mice- benchmark for glia, mitochondria, and synapses."},{"cited_title":"Supervoxel-based segmentation of mitochondria in em image stacks with learned shape fea- tures","cited_arxiv_id":null,"evidence_quote":"The public Lucchi FIB-SEM mitochondria dataset used as a standard benchmark and for the ablation study."}],"review_version":1}