{"id":"c39edf91-6e0d-486b-b952-c1703e873bba","arxiv_id":"2504.14521","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"high","formal_verification":"none","parameter_count":2,"one_line_summary":"Stacked LSTM/BiLSTM networks improve correlation coefficients by up to 27.5% over single LSTM on selected simulated EIT test models, but the result lacks statistical support, code, and clinical validation.","lead":"Researchers stacked LSTM and BiLSTM neural networks to reconstruct lung electrical impedance tomography (EIT) images from simulated data, reporting higher correlation coefficients than single-layer networks. The gains appear in the tables but are presented as one-off maxima, with no error bars or code, so the general claim of superiority is not yet supported.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The stacking claim is not controlled: no capacity-matched baselines, no repeated runs, and Table 4 shows negative ratios, so the reported CC gains cannot yet be attributed to stacking.","rationale":"The reader's external-validity concern about noise-free synthetic data is legitimate, but it is secondary: the paper's central claim is about algorithm performance, and both baselines and stacked models are evaluated under the same simulated conditions. The more load-bearing problem is internal validity. The comparison is uncontrolled because the stacked models have more parameters and no equal-capacity baseline, and the evaluation is statistically weak because each number in Tables 2 and 3 is a single observation on a single test model. The paper's own Table 4 contains negative ratio values, directly contradicting a general stacking benefit. The paper does deserve credit for generating 60,000 lung FEM datasets and defining the CC metric in Eq. (8), which makes the intended comparison clear, but no training details, code, or repeated-seed statistics are provided, so the central claim cannot be assessed. I would keep the reader's rejection, but primarily for this internal confound rather than for clinical transferability.","tokens_in":8160,"tokens_out":6125,"duration_ms":57872,"concrete_test":"Re-run the six algorithms on the same circular and lung FEM data with (i) matched parameter budgets, e.g., a single-layer LSTM whose hidden width is chosen so total parameters equal the LSTM-LSTM model; (ii) per-model hyperparameter selection on a held-out validation split; and (iii) at least 10 random seeds per configuration. Report mean and standard deviation of CC over the 10 test models and paired differences with effect sizes. If the stacked models do not show consistent positive differences under this protocol, the claim that stacking 'effectively enhances' reconstruction quality is not supported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central claim (abstract, Section 5.2, Section 6) is that stacking LSTM/BiLSTM layers improves EIT reconstruction quality, with maximum CC increases of 27.5%/25.4% over LSTM and 11.7%/13.4% over BiLSTM. The load-bearing assumption is that those CC differences are attributable to stacking rather than to confounds. This is not supported by the reported experiment. Section 4.2 gives no training details: hidden size, number of epochs, learning rate, regularization, initialization, or random seeds. Tables 2 and 3 report a single CC per model and algorithm, with no error bars, no repeated runs, and no statistical testing. Table 4 is internally troubling: for lung models, LSTM-LSTB has ratio -0.025 on L3 and LSTM-LSTL has -0.004 on L4, so the text concedes 'the exception of models L3 and L4' yet still concludes that stacking 'effectively enhances' image quality. More importantly, the stacked models contain extra layers and therefore extra parameters and training cost; a single-layer LSTM with comparable parameter count or a wider hidden layer is never evaluated. Without matching capacity or optimizing hyperparameters separately, any improvement could be due to model size or optimization effort, not to the architectural operation of stacking. The central claim as phrased is therefore unverdictable from the reported experiment.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes four stacked recurrent architectures (LSTM-LSTM, LSTM-BiLSTM, BiLSTM-LSTM, BiLSTM-BiLSTM) for lung electrical impedance tomography (EIT) image reconstruction. The authors train and evaluate these models on simulated circular and lung cross-section finite-element models, report correlation coefficients (CC) between reconstructed and ground-truth conductivity distributions, and compute relative increases over single-layer LSTM and BiLSTM baselines. The central claim is that stacking LSTM/BiLSTM layers improves reconstruction quality, with maximum CC increases of 27.5%, 25.4%, 11.7%, and 13.4% over the corresponding baselines. The paper concludes that stacked neural networks are a promising approach for lung EIT image enhancement.","tokens_in":8411,"tokens_out":3493,"duration_ms":30039,"significance":"If the central claim is validated, the paper would offer a simple and potentially useful empirical contribution to deep-learning-based EIT reconstruction, which is an active area. The use of the correlation coefficient as a standard image-quality metric and the tabulation of per-model results are positive features. However, the experiments as reported lack the controls needed to attribute observed improvements to stacking rather than to confounds such as model capacity, optimization effort, or run-to-run variability. The paper also evaluates only noise-free synthetic data, so the clinical relevance suggested by the title and conclusion is not yet established. The work does not provide code, trained models, or machine-checked proofs, so the reproducibility of the numerical results cannot be independently assessed.","major_comments":[{"comment":"The text in Section 5.2 states that the BiLSTM-LSTM and BiLSTM-BiLSTM algorithms exhibit increase ratios of 109.7% and 113.4%, but Table 4 reports values of 0.117 and 0.134 (i.e., 11.7% and 13.4%). This is an order-of-magnitude inconsistency that makes the reported maximum increases unreliable. Additionally, Table 4 contains negative ratios (BiL-BiLB L4 = -0.047, LSTM-LSTB L3 = -0.025, LSTM-LSTL L4 = -0.004), directly contradicting the conclusion in Section 5.2 that the stacked algorithms exceed the baseline 'with the exception of models L3 and L4' and undermining the blanket claim in Section 6 that 'stacking neural networks effectively enhances the quality of the reconstructed images.' These exceptions must be acknowledged and analyzed, and the conclusions must be qualified accordingly.","section":"Section 5.2, Table 4"},{"comment":"The experimental protocol is critically under-specified. Section 4.2 describes the datasets but gives no training hyperparameters: hidden layer sizes, number of LSTM/BiLSTM layers, learning rate, number of epochs, optimizer, regularization, initialization, batch size, or random seeds. Tables 2 and 3 report a single correlation coefficient per model and algorithm, with no error bars, no repeated trials, and no statistical significance testing. Given the small differences between some configurations (e.g., BiL vs. BiLL on CIR1: 0.8818 vs. 0.9023; LSTM vs. LSTL on L4: 0.8643 vs. 0.8611), the reported gains may be indistinguishable from run-to-run variation. Repeated independent runs and reporting of mean, standard deviation, or confidence intervals are necessary before any claim of 'significant enhancement' is supportable.","section":"Sections 4.2 and 5.1"},{"comment":"The comparison is not capacity-controlled. Stacking layers as described (LSTM-LSTM, LSTM-BiLSTM, BiLSTM-LSTM, BiLSTM-BiLSTM) necessarily increases the number of trainable parameters relative to the single-layer LSTM and BiLSTM baselines. The manuscript never evaluates a single-layer LSTM or BiLSTM with a matched parameter count (e.g., a wider hidden state) or a matched training budget. Therefore the observed CC increases could be due to increased model capacity or additional optimization effort rather than to the architectural operation of stacking. To support the central claim, the authors must include capacity-matched baselines or otherwise control for model size and training cost.","section":"Section 3.4"},{"comment":"The training and evaluation are performed entirely on simulated, noise-free boundary voltages generated from 576-element finite-element models. The manuscript provides no validation with added measurement noise, electrode contact impedance, or anatomical/patient variability, despite the title's 'pulmonary' scope and the conclusion's promise to apply the algorithms to clinical lung monitoring and combine with CT images. If the paper's claim is limited to simulated EIT reconstruction, the wording should be revised accordingly. Otherwise, a noise-robustness study or experimental validation (phantom or clinical data) is required to justify transferability to real-world EIT measurements.","section":"Section 4.2 and Section 6"}],"minor_comments":[{"comment":"Equation numbering is inconsistent: after Eq. (1), the paper introduces Eqs. (3.1), (3.2), (3.3) for the candidate state, internal state, and external state, but the later quality metric is numbered Eq. (8). Please renumber all equations sequentially.","section":"Section 3.2"},{"comment":"The text reports 'maximum increases of 27.5% and 25.4%' for LSTM-LSTM and LSTM-BiLSTM over LSTM, but Table 4 also shows negative ratios for L3 and L4. Please add a sentence explicitly identifying the models where the stacked algorithm performs worse, and discuss possible reasons or whether these are within expected variation.","section":"Section 5.2"},{"comment":"Table 4's caption says 'increase ratio,' but the text interprets the values as percentages (e.g., 0.275 as 27.5%). Clarify the units and ensure all statements in the text consistently use either ratios or percentages. The phrase '109.7% and 113.4%' in Section 5.2 appears to be a typographical error for '11.7% and 13.4%'.","section":"Section 5.2"},{"comment":"Several typos and formatting issues remain: 'BILSTM' and 'BILSTM-BILSTM' in Section 6; 'Overall, the quality of EIT images is observed' is an incomplete sentence in Section 6; and the reference list contains an in-text citation [29] for static/dynamic imaging but no corresponding entry (only 28 references are listed).","section":"Throughout"},{"comment":"None of the figures (Fig. 1 through Fig. 11) has a caption in the submitted manuscript text. All figures should be captioned to explain their contents, especially the reconstruction results in Figures 10 and 11, which are central to the evaluation.","section":"Figures"}],"recommendation":"major_revision","confidential_remarks":"The paper presents a straightforward application of existing LSTM/BiLSTM stacking to simulated EIT data. The novelty is limited, and the evaluation is internally inconsistent and under-controlled. However, the core idea (stacking recurrent networks for EIT reconstruction) is not fundamentally flawed; the manuscript could become publishable if the authors substantially rework the experimental section, correct the numerical inconsistencies, add capacity-matched baselines and repeated trials, and temper the clinical claims. If the authors cannot provide the additional experiments requested, I would recommend rejection. I would also flag that the '109.7%' error in Section 5.2 and the negative Table 4 entries should be caught in any further review."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The useful thing to know about this paper is that it reports a purely simulated comparison of stacked LSTM/BiLSTM networks for EIT image reconstruction, and the headline numbers (27.5%, 25.4%, 11.7%, 13.4%) come from a handful of test models. The second thing: the central claim that stacking 'effectively enhances' image quality is not supported by the experiment as reported, because the stacked networks are never compared against baselines with matched capacity or optimization budget.\n\nWhat is new is modest: prior papers (refs 25–26) already apply LSTM and BiLSTM to electrical tomography, so the contribution here is four specific stacking combinations, tested on finite-element models with 576 elements. The paper describes the network components clearly and reports CC values for each test model in Tables 2–3, which is more transparent than many reconstruction papers. The idea of stacking is natural and worth a controlled test.\n\nThe controlling problem is experimental rigor. There are no training details—no hidden sizes, epochs, learning rates, initialization, or random seeds. Each CC is a single number with no repeated runs, error bars, or statistical testing. Table 4 even contains negative ratios (LSTM-LSTB on L3, BiL-BiLB on L4), which the text acknowledges in passing but then overrides with 'generally superior.' Because the stacked networks have extra layers and therefore more parameters and longer training, the observed gains could reflect model size or optimization effort rather than the architectural operation of stacking. A proper control would train a single-layer LSTM with comparable parameter count, or tune hyperparameters separately for each architecture. Without that, the 27.5% maximum increase is not attributable to stacking. There is also no external benchmark against a conventional EIT reconstruction algorithm, and the data is entirely synthetic and noiseless, so the abstract's suggestion that this is a 'valuable approach' for clinical lung monitoring is unsupported.\n\nWho is this paper for? A reader curious about how recurrent nets might behave in EIT, but not as evidence for any method's superiority. I would not cite it for the claimed effect. The paper needs major additional work: matched-capacity baselines, repeated runs, error bars, proper statistics, and ideally noisy or experimental data. If I were the editor, I'd desk reject it with an invitation to resubmit after that work; it is too thin as is. That said, the question—whether stacking recurrent layers helps EIT—is legitimate and the paper gives a starting point.","headline":"A modest simulation study of stacked LSTM/BiLSTM for EIT whose headline gains are unsupported by the experiment as reported.","tokens_in":8969,"tokens_out":4383,"would_cite":false,"duration_ms":38207,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":false},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper establishes that stacking LSTM and BiLSTM layers improves lung EIT image reconstruction, raising correlation coefficients up to 27.5% over a single LSTM.","keywords":["Electrical impedance tomography","LSTM","BiLSTM","deep recurrent neural network","lung imaging","image reconstruction","conductivity distribution","correlation coefficient"],"falsifier":"Run the same stacks and baselines on boundary voltages from a physical thorax phantom with known conductivity inclusions, or from patients whose conductivity maps are obtainable from CT; if the stacked networks' correlation coefficients do not beat single-layer baselines on this noisy real-electrode data, the central claim fails. A minimal version is to add measurement noise and random contact impedances to the simulated voltages and re-measure the reported 27.5% and 25.4% gains.","tokens_in":7933,"feed_emoji":"🫁","tokens_out":12193,"duration_ms":99510,"temperature":0.7,"pith_summary":"This paper tries to establish that stacking long short-term memory (LSTM) and bidirectional LSTM (BiLSTM) layers noticeably improves the quality of electrical impedance tomography (EIT) images of the lungs. Using simulated boundary voltages from 576-element finite element chest and circular models, it trains four stacked networks and finds that they match or beat their single-layer baselines in most test cases, with the largest gains on difficult reconstructions. The strongest reported improvements are 27.5% for LSTM-LSTM over LSTM and 25.4% for LSTM-BiLSTM over LSTM on circular models, with 11.7% and 13.4% gains for BiLSTM-based stacks over BiLSTM. The authors conclude that recurrent depth suppresses artifacts and contour errors and that the two LSTM-fronted stacks are generally superior. The result matters because EIT is a radiation-free, real-time bedside lung monitor whose clinical usefulness is limited by reconstructed image quality.","feed_headline":"Stacked networks lift lung EIT image quality by up to 27.5%","feed_subtitle":"The LSTM-LSTM and LSTM-BiLSTM stacks beat single-layer baselines on simulated lung conductivity images.","key_machinery":"The engine of the argument is the deep recurrent stack. A single LSTM cell propagates an internal state $c_t$ under forget, input, and output gates; a BiLSTM layer runs two such cells in opposite directions over the same input. Stacking means feeding the hidden-state sequence of the first recurrent layer into a second recurrent layer, and the paper builds the four combinations in both orders. Each network maps 40 boundary voltages collected by an 8-electrode adjacent-excitation dynamic EIT system into 576 finite element conductivities, and the correlation coefficient between the reconstructed and true conductivity distributions is the score that carries the comparison.","core_discovery":"The central claim is that stacking LSTM and BiLSTM layers—LSTM-LSTM, LSTM-BiLSTM, BiLSTM-LSTM, BiLSTM-BiLSTM—reconstructs simulated lung conductivity distributions from boundary voltages more accurately than either network used alone. The simulations support this by correlation coefficients: on five circular models, LSTM-LSTM reaches 0.9159 and LSTM-BiLSTM 0.9111, while on five lung models the maxima are 0.9691 and 0.9714. The largest relative gains appear where the base algorithm is weakest: circular model CIR3 improves from 0.6650 with LSTM to 0.8477 with LSTM-LSTM (a 27.5% gain), and lung model L5 improves from 0.7365 to 0.9552 (about 29.7%). The paper interprets the gains as artifact reduction and contour correction and adds the caveat that stacking increases training time, so extra depth should be added deliberately.","pith_inferences":["The comparisons use clean simulated voltages with no electrode contact impedance or measurement noise, so the reported percentage gains should be read as upper bounds until the same stacks are tested on noisy physical data; that test is the natural next experiment.","Since BiLSTM layers read the voltage sequence in both directions, stacked BiLSTM variants may show larger advantages on time-resolved ventilation sequences than on the static conductivity snapshots evaluated here, a hypothesis a temporal correlation metric could test.","The 27.5% gain on CIR3 comes from a case where baseline LSTM performs worst, which suggests stacking may be most valuable as a corrective stage attached to a weak base rather than as a universal improvement; ablating each added layer on easy and hard cases would test this.","Transferring from 60,000 synthetic finite element datasets to bedside monitoring would likely require simulated noise augmentation or fine-tuning on measured voltages, because the training distribution contains none of the artifacts of real electrodes."],"forward_implications":["Single-layer LSTM or BiLSTM reconstruction software can be upgraded by stacking an extra recurrent layer without any change to the 8-electrode measurement protocol, so the reported quality gain is available at the algorithm level.","The largest relative gains occur where the single-layer baseline is weakest, so stacking primarily repairs artifact-heavy, hard reconstructions rather than polishing already good images.","Because the four stacked variants cluster closely in maximum lung-model correlation (0.9691 to 0.9736), choice among them can be driven by training time or implementation convenience rather than by a decisive accuracy gap.","The training-time caveat implies a design rule: add recurrent depth only while the extra correlation repays the compute cost, which is exactly the trade-off the paper endorses."],"supporting_citations":[{"why":"Provides the direct precedent of a double-stage LSTM network for electrical tomography that the paper's stacked architectures extend.","marker":"Klosowski et al 2023b"},{"why":"Establishes LSTM as a usable machine-learning reconstruction approach in electrical tomography, motivating the recurrent structure.","marker":"Klosowski et al 2023a"},{"why":"Earlier multilayer neural network EIT reconstruction from the same line of work that the deep recurrent stacking builds on.","marker":"Song et al 2024"},{"why":"Supplies the standard that EIT images are assessed against CT-derived regional ventilation, used to justify the correlation-coefficient metric.","marker":"Katayama et al 2024"},{"why":"Supplies the correlation-coefficient evaluation convention for comparing reconstructed tomographic images with the true conductivity distribution.","marker":"Deng et al 2023"}],"fun_headline_variants":["LSTM stacks boost lung EIT quality by up to 27.5%","Stacked recurrent nets sharpen lung EIT images","Deep LSTM pairing lifts EIT correlation 27.5%","Lung EIT imaging improved 27.5% with stacked LSTMs","Recurrent network stacks enhance lung EIT reconstruction"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that clean synthetic boundary voltages from a 576-element finite element model stand in for real clinical lung EIT; if electrode noise, contact impedance, or patient anatomy shift the voltage-to-conductivity map, the stacked networks' correlation gains may not reproduce outside simulation.","fun_headline_variants_meta":{"raw":{"variants":["LSTM stacks boost lung EIT quality by up to 27.5%","Stacked recurrent nets sharpen lung EIT images","Deep LSTM pairing lifts EIT correlation 27.5%","Lung EIT imaging improved 27.5% with stacked LSTMs","Recurrent network stacks enhance lung EIT reconstruction"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000692,"raw_usage":{"total_tokens":3133,"prompt_tokens":946,"completion_tokens":2187,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":562,"completion_tokens_details":{"reasoning_tokens":2099}},"tokens_in":562,"tokens_out":2187,"duration_ms":14423,"temperature":1.0,"reasoning_tokens":2099,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T11:46:17.924541+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the same stacks and baselines on boundary voltages from a physical thorax phantom with known conductivity inclusions, or from patients whose conductivity maps are obtainable from CT; if the stacked networks' correlation coefficients do not beat single-layer baselines on this noisy real-electrode data, the central claim fails. A minimal version is to add measurement noise and random contact impedances to the simulated voltages and re-measure the reported 27.5% and 25.4% gains.","supporting_citations":[],"review_version":1}