{"id":"5063b3b6-0907-44da-a5f4-023fa0ddc4d7","arxiv_id":"2508.17916","paper_version":2,"verdict":"UNVERDICTED","confidence":"LOW","novelty_score":5.0,"correctness_risk":"unknown","formal_verification":"none","parameter_count":0,"one_line_summary":"EndoUFM claims state-of-the-art monocular depth estimation for endoscopic images using dual foundation models, RVLoRA adaptation, and mask-guided smoothness.","lead":"EndoUFM estimates depth from single endoscopic images by combining two foundation models with adaptive fine-tuning. It reports state-of-the-art results on four surgical datasets while keeping the model small.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"SOTA claim is unverifiable from the supplied evidence: the abstract gives no numbers and the full text is corrupted and even carries an unrelated arXiv header.","rationale":"The strongest claim is not a theorem; its truth depends on controlled experiments, evaluation protocols, and reported comparisons. The only way to assess it is through tables and protocol descriptions, and the supplied material blocks that check at every point. The reader's weakest assumption about foundation-model transferability is reasonable but secondary: even if the transfer assumption were wrong, the SOTA claim could still fail or hold for other architectural reasons, and conversely, transferability alone would not validate the reported superiority without fair baselines and identical splits. The load-bearing issue is therefore the unavailability of the empirical record. The proposed check is the minimal verification that would resolve it: run the public code on the public split and compare metrics. This pass neither accepts nor rejects the paper; it leaves the reader's UNVERDICTED verdict unchanged.","tokens_in":23412,"tokens_out":4227,"duration_ms":46640,"concrete_test":"Download the released repository at https://github.com/RealMindyY/EndoUFM and run its provided inference/evaluation script on the public SCARED test split used in the paper, computing AbsRel, RMSE, and delta<1.25; then compare the reproduced metrics with the values in a clean copy of the paper. If the table row or baseline protocol cannot be identified, or if the reproduced numbers differ beyond rounding tolerance, the SOTA claim is not sustained.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is empirical: EndoUFM achieves state-of-the-art monocular depth on SCARED, Hamlyn, SERV-CT, and EndoNeRF while keeping an efficient model size. For that claim to hold, the paper must provide comparison tables with the same metrics and splits as the cited baselines, a clear unsupervised training protocol, and a parameter/FLOPs comparison supporting the efficiency claim. None of that support is visible: the abstract contains no quantitative results, and the supplied full text is mojibake that cannot be parsed into tables or ablations. More concerning, the body contains an inserted and unrelated arXiv identifier, arXiv:2508.17908v2 [cond-mat.stat-mech] 3 Mar 2026, indicating that the provided text is not a trustworthy copy of the submission. I do not see an internal logical inconsistency in the method as sketched; the soft spot is that the decisive evidence is inaccessible, which justifies withholding judgment rather than a technical rejection. The public code link is a positive mitigating factor, but code alone does not establish the reported SOTA numbers.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes EndoUFM, an unsupervised monocular depth estimation framework for endoscopic images that combines two foundation models with a Random Vector Low-Rank Adaptation (RVLoRA) fine-tuning strategy, a Residual block based on Depthwise Separable Convolution (Res-DSC), and a mask-guided smoothness loss. The abstract claims state-of-the-art performance on SCARED, Hamlyn, SERV-CT, and EndoNeRF datasets while maintaining an efficient model size, and it provides a public code repository. However, the supplied full text is corrupted and unreadable, and the abstract contains no quantitative results, comparison tables, or ablations, so the central empirical claims cannot be verified from the materials provided.","tokens_in":23609,"tokens_out":3066,"duration_ms":30486,"significance":"If the claimed state-of-the-art results are reproducible, EndoUFM would be a practically relevant contribution to surgical depth estimation, offering a concrete recipe for transferring natural-image foundation models to the endoscopic domain. The release of code at https://github.com/RealMindyY/EndoUFM is a positive step for reproducibility. The problem of domain shift between pretraining and endoscopic images is well identified, and the combination of low-rank adaptation, depthwise separable residual blocks, and a mask-guided smoothness loss is a plausible design. Nevertheless, because the contribution is empirical and the supporting evidence is currently inaccessible, the significance of the paper cannot be assessed until a legible manuscript with actual numbers is provided.","major_comments":[{"comment":"The central claim that the method \"achieves state-of-the-art performance\" is unsupported in the abstract and in the supplied text: no numerical metrics (e.g., Abs Rel, RMSE, delta thresholds), no baseline comparisons, no ablations, and no error bars are reported anywhere. Since the paper's contribution is empirical, the abstract should state concrete results for at least one dataset and the body must contain full comparison tables with the same evaluation protocols as the cited baselines.","section":"Abstract"},{"comment":"The provided full text is corrupted mojibake that cannot be parsed into equations, tables, or experimental descriptions. In addition, the body embeds an unrelated arXiv identifier, \"arXiv:2508.17908v2 [cond-mat.stat-mech] 3 Mar 2026,\" which indicates that the supplied copy is not a trustworthy version of the submission. This makes it impossible to check the method details, training protocol, implementation, or the claimed experiments. The authors must resubmit a clean, completely readable manuscript before the technical content can be reviewed.","section":"Full text (unreadable copy)"},{"comment":"The phrase \"maintaining an efficient model size\" is not quantified. To support this claim, the paper must report parameter counts, FLOPs, or inference time for EndoUFM and for all compared baselines. Additionally, the individual contribution of RVLoRA, Res-DSC, and the mask-guided smoothness loss should be demonstrated through an ablation study; no such ablation is visible in the supplied materials.","section":"Abstract (efficiency claim)"}],"minor_comments":[{"comment":"The underlined letter in \"U̲tilizes\" appears to be a formatting artifact; the abstract should be plain text.","section":"Abstract"},{"comment":"The term \"Random Vector Low-Rank Adaptation (RVLoRA)\" is not defined in the abstract; a one-sentence description or a reference is needed for readers unfamiliar with this variant.","section":"Abstract"},{"comment":"The supplied document lacks a clear section structure with page numbers, making it hard to navigate even after decoding; the final version should follow the journal's formatting guidelines.","section":"General"}],"recommendation":"major_revision","confidential_remarks":"The supplied full text is corrupted and contains an unrelated arXiv header, which strongly suggests a problem in the submission or conversion pipeline. The editor should ask the authors for a clean source version (e.g., PDF or LaTeX) before sending the paper for further technical review. Because the central claim is empirical and currently unverifiable, the paper should not be considered for acceptance until the manuscript is readable and the quantitative comparisons are provided."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"One thing you should know: this is an unsupervised monocular-depth framework for endoscopic images, combining two foundation models with RVLoRA fine-tuning, a Res-DSC block, and a mask-guided smoothness loss. The abstract claims state-of-the-art performance on SCARED, Hamlyn, SERV-CT, and EndoNeRF at an efficient model size. The problem: our copy of the full text is mojibake and contains an unrelated arXiv identifier (2508.17908v2, cond-mat), so the claim is supported only by the abstract, which gives no numbers.\n\nWhat is actually new: not a lot, but the combination is reasonable. RVLoRA is a low-rank adaptation variant, Res-DSC is a residual depthwise-separable block, and the mask-guided smoothness uses segmentation to regularize depth. Each piece is known, and the paper is upfront about borrowing foundation-model priors. The motivation—natural-image priors need light adaptation to endoscopy—is sound, and releasing code helps.\n\nSoft spots, in order of importance. First, the abstract's SOTA sentence is unverifiable: no metrics, no baselines, no ablations, no parameter/FLOPs counts, no error bars. We cannot check whether they compare on the same splits as the cited methods. Second, the full text is unreadable, and the embedded cond-mat arXiv ID means we cannot trust this copy as the submission. That may be a pipeline artifact, but it blocks any assessment of the math and the tables. Third, the novelty is incremental. That is not fatal, but it means the contribution rests entirely on the empirical claims.\n\nNo internal contradiction is visible in the sketch, and the method as described is coherent. The code link is a positive sign. Yet I cannot verify soundness, and the citation pattern is invisible because the reference list is corrupted.\n\nWho this is for: people working on depth estimation in surgical video or on adapting foundation models to medical imaging. They would want the actual tables before citing it.\n\nMy recommendation: if a clean PDF shows the comparison tables and ablations, this deserves peer review at a surgical-imaging or medical-AI venue. The version we have is not reviewable as-is; I would ask the authors for a readable manuscript and the exact metrics/splits before spending referee time. Do not desk reject on principle, but do not accept the SOTA claim on faith.","headline":"EndoUFM is a plausible assembly of known components with a SOTA claim on four endoscopic benchmarks, but our copy is illegible and contains a stray arXiv header, so the claim rests entirely on an abstract with no numbers.","tokens_in":24116,"tokens_out":4481,"would_cite":false,"duration_ms":47672,"reading_group":"no","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"EndoUFM claims that adapting two pretrained foundation models with lightweight random-vector low-rank adapters produces state-of-the-art unsupervised monocular depth estimates for endoscopic scenes while keeping the model small.","keywords":["monocular depth estimation","endoscopic imaging","foundation models","low-rank adaptation","unsupervised learning","depthwise separable convolution","surgical navigation","3D reconstruction"],"falsifier":"Retrain EndoUFM from randomly initialized encoders with the same RVLoRA, Res-DSC, and mask-guided smoothness loss on the same four datasets; if the randomly initialized model matches the pretrained-backbone version on SCARED and Hamlyn metrics, the transferability premise is falsified. Alternatively, test on an endoscopic domain outside the training distribution, such as a different surgical instrument or patient cohort, and check whether the margin over baselines collapses.","tokens_in":23241,"feed_emoji":"🩺","tokens_out":4756,"duration_ms":49744,"temperature":0.7,"pith_summary":"The paper sets out to prove that monocular depth estimation inside endoscopic surgery can be made substantially more accurate by exploiting two pretrained foundation models rather than training a depth network from scratch or from a single generic backbone. It introduces EndoUFM, a fully unsupervised framework that adapts the foundation models with Random Vector Low-Rank Adaptation (RVLoRA), adds a residual depthwise-separable convolution block for local detail, and regularizes with a mask-guided smoothness loss. If the reported results hold, surgeons would get denser and more reliable depth cues from ordinary endoscope video, which matters for AR overlays, navigation, and safer minimally invasive procedures.","feed_headline":"EndoUFM tops endoscopic depth benchmarks with a slim model","feed_subtitle":"Adapting two pretrained models with RVLoRA yields state-of-the-art unsupervised depth on four surgical datasets.","key_machinery":"The framework pairs two pretrained foundation models as feature extractors, adapts them with Random Vector Low-Rank Adaptation (RVLoRA), a low-rank adaptation scheme that keeps the number of trainable parameters small while letting the pretrained features shift toward endoscopic statistics, and then feeds the adapted features through a Residual block built on Depthwise Separable Convolutions (Res-DSC) to capture fine local detail. A mask-guided smoothness loss regularizes depth within anatomical regions. RVLoRA preserves pretrained priors while limiting trainable parameters, Res-DSC sharpens local structure, and the mask-guided loss discourages depth jumps inside coherent tissue regions.","core_discovery":"EndoUFM is an unsupervised monocular depth estimation framework that takes endoscopic frames as input and produces dense depth maps without ground-truth depth supervision. Its central claim is that two pretrained foundation models, adapted with the lightweight RVLoRA fine-tuning strategy, supply semantic and geometric priors that bridge the gap between natural-image pretraining and endoscopic scenes; the Res-DSC residual block then recovers fine-grained local features, and a mask-guided smoothness loss enforces depth consistency inside anatomical structures. The authors report that this combination achieves state-of-the-art results on the SCARED, Hamlyn, SERV-CT, and EndoNeRF datasets while keeping the model size efficient.","pith_inferences":["Editorial inference: The same dual-foundation-model plus RVLoRA recipe could plausibly transfer to other narrow-domain medical imagery such as capsule endoscopy, arthroscopy, or fetoscopy, where illumination and texture also differ sharply from natural images; the paper does not test this.","Editorial inference: Because the loss is mask-guided, EndoUFM's benefit depends on the availability or quality of anatomical masks, so a natural extension would measure how performance degrades as mask noise increases.","Editorial inference: The benchmark claim is limited to the four datasets tested; a harder, untested question is whether the margin survives cross-patient, cross-instrument, or unseen-disease generalization."],"forward_implications":["If EndoUFM's reported accuracy holds, monocular endoscopic video alone can support dense 3D reconstruction, removing the need for stereo endoscopes or structured-light hardware in AR-assisted surgery.","Because RVLoRA leaves most foundation-model weights frozen, the framework's efficient footprint makes it plausible to run depth estimation in near-real-time on clinical hardware.","The mask-guided smoothness loss ties depth estimates to anatomical regions, which should reduce spurious depth jumps at tissue boundaries and improve the stability of 3D overlays.","On the four benchmark datasets, EndoUFM is claimed to outperform existing monocular depth estimators while using fewer parameters, giving a practical baseline for future endoscopic depth methods."],"supporting_citations":[],"fun_headline_variants":["EndoUFM: dual foundation models for endoscopic depth","Unsupervised depth from endoscopy via dual FMs","EndoUFM: RVLoRA-tuned FMs top endoscopic depth","Slim model, SOTA depth: EndoUFM on four datasets","Endoscopies go deep: EndoUFM adapts FMs with RVLoRA"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole method leans on the assumption that representations learned from natural images remain useful for endoscopic scenes after only lightweight RVLoRA adaptation; if that transfer does not hold, the reported gains would disappear.","fun_headline_variants_meta":{"raw":{"variants":["EndoUFM: dual foundation models for endoscopic depth","Unsupervised depth from endoscopy via dual FMs","EndoUFM: RVLoRA-tuned FMs top endoscopic depth","Slim model, SOTA depth: EndoUFM on four datasets","Endoscopies go deep: EndoUFM adapts FMs with RVLoRA"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000179,"raw_usage":{"total_tokens":1298,"prompt_tokens":941,"completion_tokens":357,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":557,"completion_tokens_details":{"reasoning_tokens":261}},"tokens_in":557,"tokens_out":357,"duration_ms":3805,"temperature":1.0,"reasoning_tokens":261,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T16:59:41.627281+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Retrain EndoUFM from randomly initialized encoders with the same RVLoRA, Res-DSC, and mask-guided smoothness loss on the same four datasets; if the randomly initialized model matches the pretrained-backbone version on SCARED and Hamlyn metrics, the transferability premise is falsified. Alternatively, test on an endoscopic domain outside the training distribution, such as a different surgical instrument or patient cohort, and check whether the margin over baselines collapses.","supporting_citations":[],"review_version":1}