{"id":"8b6d88b0-1114-4075-bafc-aa57714eadf0","arxiv_id":"2504.21562","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"high","formal_verification":"none","parameter_count":4,"one_line_summary":"Neural Cellular Automata segment bleeding and estimate depth on an ESP32-S3 microcontroller in under a second, with segmentation accuracy near nnUNet and depth quality shown only qualitatively.","lead":"A swallowed camera pill records hours of digestive tract video that doctors must review. This paper packs tiny neural networks onto a microcontroller inside the pill to detect bleeding and estimate depth, potentially cutting review time and enabling future capsule self-localization.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Depth claim rests on unvalidated pseudo-labels; no ground-truth evaluation supports 'reliable depth estimation' on the device.","rationale":"The paper's novelty is the combination of (a) NCA-based bleeding segmentation, (b) distilled depth-estimation NCA, and (c) an ESP32 port. The segmentation claim is backed by quantitative Dice scores on KID2 (Table 1), and the port is demonstrated with runtime measurements, though a subtle training/deployment stochastic-update mismatch (80% in Sec. 3.1 vs. 50:50 in Sec. 3.3) could affect on-device fidelity. The depth claim, however, has no quantitative validation at all; it is trained and judged against the same pseudo-label source, creating a circularity risk: the student NCA can only be as good as the teacher, and neither is checked against physical depth. The paper explicitly flags this limitation in Sec. 4, and it is the weakest load-bearing assumption. A simulator-based ground-truth evaluation is readily available via EndoSLAM or VR-Caps and would settle the question. This concern matches the reader's weakest_assumption, so I retain the CONDITIONAL verdict and see no change needed.","tokens_in":9103,"tokens_out":4398,"duration_ms":43811,"concrete_test":"Evaluate eNCApsulateD on EndoSLAM's simulated WCE depth test set (Ozyoruk et al., MedIA 2021) or VR-Caps rendered images with metric depth ground truth, computing AbsRel, RMSE, and delta1 accuracy. If eNCApsulateD's errors are large (e.g., AbsRel > 30%) or merely match Depth Anything V2's errors rather than improving on them, the 'reliable depth estimation' claim fails. If no simulator-based evaluation is feasible, the authors should at minimum release the curated pseudo-label set and an on-device depth output comparison against DAv2 on the same images.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim of 'reliable depth estimation on a miniaturized device' depends entirely on Depth Anything V2 pseudo-labels as the source of supervision (Sec. 3.2). The paper itself admits in Sec. 4: 'One central limitation of our approach is the absence of a proper ground truth for depth, forcing us to use the best-performing foundation model as a pseudo ground truth.' There is no external depth ground truth for training or validation; depth evaluation is qualitative on 25 selected KvasirCapsule images, with no quantitative metrics such as AbsRel or RMSE. Since Depth Anything V2 is trained predominantly on natural images, its depth estimates for capsule endoscopy may carry systematic errors on folds, bubbles, and dark luminal regions. If the pseudo-labels are biased, eNCApsulateD can only learn that bias, and the reliability claim has no empirical support. This is not a disagreement with consensus; it is a missing-evidence problem in a load-bearing assertion. The segmentation contribution is quantitatively supported by Table 1, but the depth pillar is not, and the headline claim bundles both.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes eNCApsulate, a method that trains Neural Cellular Automata (NCAs) for bleeding segmentation and monocular depth estimation in wireless capsule endoscopy, and ports the trained models to an ESP32-S3 microcontroller. Segmentation is trained on manual labels from the KID2 dataset and evaluated against lightweight U-Net baselines in Table 1, where eNCApsulateS achieves a Dice score of 0.576 with 47,152 parameters. Depth estimation is trained by distilling Depth Anything V2 pseudo-labels on 727 curated KID2 images and is evaluated qualitatively on 25 selected KvasirCapsule images. The paper also describes a temporal early-stopping scheme that reduces NCA inference steps by roughly a factor of five and reports SIMD-based runtime optimizations that reduce per-image inference time on the ESP32-S3 from 9 s to 3 s, and further to under 1 s with early stopping. The central claim is that this is the first work enabling reliable bleeding segmentation and depth estimation on a miniaturized device, as a step toward on-capsule diagnosis and visual-odometry-based localization.","tokens_in":9266,"tokens_out":4685,"duration_ms":48229,"significance":"If the claims hold, the contribution is practically significant: NCAs on a capsule-sized microcontroller could enable on-device bleeding screening and reduce the amount of video data that must be transmitted, and the depth-estimation work is a step toward sensor-less capsule localization. The segmentation contribution is quantitatively supported by a five-fold ensemble evaluation with standard deviations, and the parameter advantage is stark (47,152 parameters versus the smallest baseline's 14,342,596). The deployment engineering, including the ANSI C implementation, SIMD usage, and buffer reuse, is a concrete and reproducible advance. The paper also provides a public code repository, which strengthens reproducibility. However, the depth-estimation claim is not established at the same evidentiary level: it relies entirely on Depth Anything V2 pseudo-labels, has no external ground truth, and is evaluated only qualitatively. The temporal-regularization claim likewise lacks quantitative segmentation metrics at the early cutoff.","major_comments":[{"comment":"The depth-estimation claim is supported only by qualitative comparison on 25 selected KvasirCapsule images (Figs. 5 and 6), with no quantitative metric such as AbsRel, RMSE, or delta accuracy, and with no ground-truth depth for validation or test. The text itself acknowledges 'the absence of a proper ground truth for depth' in Section 4. Because eNCApsulateD is trained and assessed inside the loop of Depth Anything V2 pseudo-labels, any systematic error in the teacher is propagated and cannot be detected by the presented evaluation. I ask the authors to either evaluate on a dataset with real or simulated depth ground truth (e.g., EndoSLAM or VR-Caps) or to remove the word 'reliable' from the depth claim and present the depth model as a qualitative pseudo-label distillation.","section":"Section 4, Depth Estimation with eNCApsulateD"},{"comment":"The stochastic update probability is inconsistent between training and device deployment. Section 3.1 states that 'only 80% of cells are stochastically updated in each timestep,' while Figure 2's caption says the update is applied 'by a chance of 50%,' and Section 3.3 says the microcontroller code uses a '50:50 condition.' If the deployed model updates only 50% of cells per step but the evaluated PyTorch model was trained with 80%, the on-chip behavior likely differs from the reported quantitative results. Please align the training and deployment probabilities and report device-side segmentation metrics under the actual deployed probability.","section":"Sections 3.1, 3.3, and Figure 2"},{"comment":"The early-stopping claim 'at the same segmentation quality' is not quantitatively supported. The paper reports only the total number of NCA steps with and without early stopping (1,222,998 vs. 6,988,560), not the Dice or IoU at the early cutoff. Please report segmentation metrics, with variance, for full inference versus early-stopped inference on a per-frame basis, along with the distribution of stopping steps, so that the accuracy-runtime trade-off is actually verified.","section":"Section 4, Temporal Regularization"},{"comment":"The depth training data are curated with a hand-set normalized gradient magnitude threshold of 1.1, and only 727 samples remain after curation. The paper neither justifies this threshold nor analyzes its sensitivity. Please provide a histogram of the curation criterion, an ablation over thresholds, and a breakdown of the remaining samples by patient or anatomy, so that reviewers can assess whether the small curated set supports the generalization claimed from the qualitative results.","section":"Section 3.2, Depth curation"},{"comment":"The size advantage is stated inconsistently: the Abstract claims 'more than 100x fewer parameters,' while Section 4 says the model is '10x smaller than the smallest baseline' (MobileNetV3s). In addition, the column header 'Size [B]' is ambiguous: it should state whether the values are parameter counts or stored bytes, and if bytes, the numeric format and quantization should be specified. Because the miniaturization claim is central to the paper, this inconsistency and ambiguity should be resolved.","section":"Table 1 and Abstract"}],"minor_comments":[{"comment":"The stochastic update probability is described as 80% in the text and 50% in the Figure 2 caption; the figure and text should agree (this is also noted in the major comments, but the figure itself needs a correction).","section":"Section 3.1"},{"comment":"The sentence 'We also used similar settings and hyperparameters for the baseline training' is vague; please report the exact optimizer, learning rate, loss, and number of epochs used for each baseline model.","section":"Section 4, Segmentation"},{"comment":"The axis labels in Figure 3 appear corrupted in the manuscript as sequences of '/uni0000...' characters; please replace them with readable text.","section":"Figure 3"},{"comment":"The model names eNCApsulateS and eNCApsulateD are introduced without an explicit naming convention; please define the suffix convention at first use.","section":"Section 3.2"},{"comment":"The qualitative depth evaluation is based on 'a subset of the KvasirCapsule dataset with interesting benchmark images'; please specify the selection criteria to avoid the appearance of cherry-picking.","section":"Section 4, Depth Evaluation"}],"recommendation":"major_revision","confidential_remarks":"The segmentation pillar is solid enough to warrant revision rather than rejection, but the depth pillar is currently a qualitative pseudo-label exercise whose headline claim exceeds the evidence. I would encourage the editor to require either external depth validation or a substantially tempered claim before publication. The inconsistency in the stochastic update probability between training and device code is also a correctness risk that should be addressed during revision."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\neNCApsulate is worth a look if you care about putting neural networks on capsule endoscopes. The real news is the segmentation: the authors show an NCA at 47k parameters matching nnUNet (Dice 0.576 vs 0.582) on KID2 bleeding segmentation, while beating every lightweight U-Net backbone they compared against. That is a reproducible result, with five-fold splits and standard deviations, and the hardware port to an ESP32-S3 is a genuine engineering effort: from-scratch ANSI C, SIMD, three-fold speedup. The temporal early-stopping trick that cuts inference steps by a factor of ~5 is also useful and clearly explained.\n\nThe depth estimation is the soft spot. It is trained entirely on Depth Anything V2 pseudo-labels, with no external ground truth at any point; the evaluation is qualitative on a hand-picked subset of KvasirCapsule. The paper admits this limitation in Section 4, but then the abstract and conclusion still claim 'reliable depth estimation on a miniaturized device.' That is an overclaim. If DA2's depth maps are biased for endoscopic images, the student NCA can only inherit that bias. The hand-set threshold for curating 'flat' depth maps (1.1) also needs scrutiny.\n\nTwo technical issues jump out. First, the stochastic update probability is 80% during training (Section 3.1) but 50:50 on the microcontroller (Section 3.3). That is a train/test mismatch for a stochastic process, and the authors do not address it. Second, the temporal regularization is not evaluated in terms of segmentation Dice at the early cutoff; they report step counts but not whether the output quality actually holds. Both are fixable with a few experiments.\n\nThe paper is an incremental but solid application of existing NCA machinery from Kalkhof et al. to a new domain, plus a first distillation of a large depth model into an NCA. That is enough novelty for a workshop or a solid conference paper after revision. It is not enough to support the 'first reliable on-capsule depth' claim. I would send it to peer review, but I would ask the authors to add a quantitative depth evaluation against available synthetic endoscopic depth data (e.g., EndoSLAM or VR-Caps simulator), report Dice under early stopping, reconcile the 80% vs 50:50 discrepancy, and clarify exactly what code is released and when. The segmentation half deserves to be published; the depth half needs more evidence.","headline":"Solid NCA segmentation port with a useful early-stopping trick, but the depth-estimation pillar is under-evidenced and the stochastic update mismatch needs fixing.","tokens_in":9812,"tokens_out":2543,"would_cite":false,"duration_ms":26388,"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 demonstrates that Neural Cellular Automata can run bleeding segmentation and depth estimation on a microcontroller small enough for a capsule endoscope, opening the way to on-capsule diagnosis and visual-odometry localization.","keywords":["Wireless Capsule Endoscopy","Neural Cellular Automata","Bleeding Segmentation","Monocular Depth Estimation","Model Distillation","Edge AI","Microcontroller Inference","Visual Odometry"],"falsifier":"Measure eNCApsulateD's output against known geometry from a simulated capsule environment or a calibrated phantom; if absolute depth error is large, or if the flatness filter removes most informative frames, the depth claim fails. A patient-disjoint bleeding segmentation test would also settle whether the Dice advantage over lightweight U-Nets holds beyond the data used.","tokens_in":8855,"feed_emoji":"💊","tokens_out":9185,"duration_ms":92795,"temperature":0.7,"pith_summary":"Wireless capsule endoscopy records hours of gut video that clinicians must inspect frame by frame, and locating the capsule once swallowed remains difficult without extra sensors. This paper proposes that Neural Cellular Automata (NCAs), a lightweight architecture where one learned rule is applied repeatedly across an image grid, can perform two clinically useful tasks directly on a microcontroller small enough to fit in a capsule: segmenting bleeding regions and estimating depth from single frames. The paper reports that the segmentation NCA beats other portable segmentation models on Dice score while storing more than 100 times fewer parameters, and that the depth NCA, trained by distilling a large pretrained depth model, produces visually detailed depth maps, sometimes more detailed than the teacher. If these results hold, on-capsule diagnosis becomes practical and sensor-free localization through visual odometry becomes a realistic next step.","feed_headline":"Bleeding checks and depth maps now run on a swallowed capsule","feed_subtitle":"Neural Cellular Automata squeeze diagnosis and localization onto hardware small enough for a capsule endoscope.","key_machinery":"The load-bearing object is the Neural Cellular Automaton itself: a shared MLP rule applied to every cell of an image grid, with each cell aggregating its Moore neighborhood through learned $3 \\times 3$ filter banks before the MLP computes a state update that is added back with a stochastic cell update. Because the same parameters are reused over hundreds of iterative steps, the model stays tiny, about 47,000 parameters for segmentation, while still refining spatial structure. For depth, the mechanism is distillation: the outputs of a large monocular depth model, accepted only when their normalized gradient magnitude exceeds a threshold so flat pseudo-labels are discarded, become the training signal for the student NCA. On the device, the paper adds an early-stopping rule based on the summed absolute change of hidden channels after a minimum of ten steps, which is what turns a three-second inference into one under a second.","core_discovery":"The central claim is that distillation and iterative cellular inference make it possible to compress bleeding segmentation and monocular depth estimation into NCA models of under 70 kB that run on an ESP32-S3 microcontroller, the kind of chip that fits inside a wireless capsule endoscope. On a held-out capsule-endoscopy test set, the segmentation model reaches 0.576 Dice, outperforming U-Net backbones that use millions of parameters, while the depth model is trained on pseudo ground truth produced by Depth Anything V2 and filtered by a flatness check; qualitatively it produces depth maps that the authors judge at least as realistic as the teacher's on difficult capsule images. A from-scratch C implementation with SIMD instructions reduces inference on the ESP32-S3 from roughly 9 s to 3 s per image, and a temporal early-stopping rule that halts inference when hidden-channel activity falls below a threshold cuts total NCA steps by a factor of 5. The paper concludes that this is the first demonstration of reliable bleeding segmentation and depth estimation on a miniaturized device, paving the way for visual odometry on the capsule.","pith_inferences":["The flatness-filtered pseudo-label curation is a transferable idea: any distillation pipeline facing severe domain shift could use a similar gradient-based or confidence-based filter on teacher outputs before training.","The paper does not implement visual odometry, but its absolute-depth claim is the key enabling step; an immediate test would be feeding the NCA depth maps into an existing monocular VO pipeline.","The temporal early-stop threshold and cooldown are likely domain-specific; adapting them to other video streams with mostly static content could yield similar speedups on other edge devices.","If the Dice advantage over much larger CNNs persists on multi-patient data, it would suggest that local texture cues dominate bleeding detection in capsule images, rather than large-scale spatial context."],"forward_implications":["Bleeding segmentation can run on the capsule at frame rates around 2–3 FPS, matching typical recording rates, so only frames with findings need to be transmitted.","On-device depth maps supply the missing input for visual odometry, opening a route to sensor-free localization of the capsule in the GI tract.","Model storage drops from millions of parameters to tens of thousands, so other diagnostic tasks could be embedded on the same class of microcontroller.","The early-stopping scheme lets users trade a small amount of accuracy for speed, which matters on battery-limited devices.","If the student can match a much larger teacher on this underrepresented domain, distillation from general-purpose models becomes a viable recipe for other niche medical imaging tasks."],"supporting_citations":[{"why":"Supplies the NCA segmentation approach and prior evidence that NCAs are robust and lightweight for medical images, including the batch-duplication training trick.","marker":"[9]"},{"why":"Defines the core NCA architecture of learned filters, MLP update rule, and stochastic cell updates that eNCApsulate adapts.","marker":"[16]"},{"why":"Teacher model whose outputs are treated as pseudo ground truth for training the depth-estimation NCA.","marker":"[18]"},{"why":"Provides the training frames and held-out test images used for bleeding segmentation and for deriving pseudo-labeled depth samples.","marker":"[17]"},{"why":"Provides an unseen benchmark set used for the qualitative depth evaluation across categories such as blood, bubbles, folds, debris, and foreign bodies.","marker":"[10]"},{"why":"Prior work running simple CNN models on a capsule-sized chip, serving as the baseline that this work extends beyond.","marker":"[13]"},{"why":"Motivates depth estimation and visual odometry for capsule endoscopy, establishing the downstream localization goal.","marker":"[6]"}],"fun_headline_variants":["Bleeding and depth AI now run on a swallowed capsule","NCA squeezes diagnosis and depth estimation onto capsule hardware","First on-chip bleeding and depth mapping for capsule endoscopy","Distilled NCA makes capsule endoscopy efficient with 100x fewer params","Capsule cameras get on-board bleeding detection and depth estimates"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The depth-estimation claim rests entirely on the assumption that the pseudo ground truth produced by the large teacher model is accurate enough for capsule endoscopy; the paper has no independent depth ground truth, so a wrong teacher would make the student's apparent depth quality an artifact.","fun_headline_variants_meta":{"raw":{"variants":["Bleeding and depth AI now run on a swallowed capsule","NCA squeezes diagnosis and depth estimation onto capsule hardware","First on-chip bleeding and depth mapping for capsule endoscopy","Distilled NCA makes capsule endoscopy efficient with 100x fewer params","Capsule cameras get on-board bleeding detection and depth estimates"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000221,"raw_usage":{"total_tokens":1507,"prompt_tokens":1059,"completion_tokens":448,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":675,"completion_tokens_details":{"reasoning_tokens":361}},"tokens_in":675,"tokens_out":448,"duration_ms":5046,"temperature":1.0,"reasoning_tokens":361,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T04:59:14.358626+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Measure eNCApsulateD's output against known geometry from a simulated capsule environment or a calibrated phantom; if absolute depth error is large, or if the flatness filter removes most informative frames, the depth claim fails. A patient-disjoint bleeding segmentation test would also settle whether the Dice advantage over lightweight U-Nets holds beyond the data used.","supporting_citations":[{"cited_title":"In: International Conference on 11 Information Processing in Medical Imaging, pp","cited_arxiv_id":null,"evidence_quote":"Supplies the NCA segmentation approach and prior evidence that NCAs are robust and lightweight for medical images, including the batch-duplication training trick."},{"cited_title":"Distill 5(2), 23 (2020)","cited_arxiv_id":null,"evidence_quote":"Defines the core NCA architecture of learned filters, MLP update rule, and stochastic cell updates that eNCApsulate adapts."},{"cited_title":"Endosc Int Open 5(6), 477–483 (2017)","cited_arxiv_id":null,"evidence_quote":"Provides the training frames and held-out test images used for bleeding segmentation and for deriving pseudo-labeled depth samples."},{"cited_title":": Kvasir- capsule, a video capsule endoscopy dataset","cited_arxiv_id":null,"evidence_quote":"Provides an unseen benchmark set used for the qualitative depth evaluation across categories such as blood, bubbles, folds, debris, and foreign bodies."},{"cited_title":"Scientific reports 12(1), 13723 (2022)","cited_arxiv_id":null,"evidence_quote":"Prior work running simple CNN models on a capsule-sized chip, serving as the baseline that this work extends beyond."},{"cited_title":"Medical image analysis 71, 102058 (2021)","cited_arxiv_id":null,"evidence_quote":"Motivates depth estimation and visual odometry for capsule endoscopy, establishing the downstream localization goal."}],"review_version":1}