{"id":"190f2844-7fd0-4efa-860a-8f56da9a4499","arxiv_id":"2505.02843","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":2.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"A review connecting the physics of medical imaging acquisition to AI reconstruction and generation methods, arguing that physics-informed machine learning makes medical AI more trustworthy.","lead":"This paper reviews the physics behind every major medical imaging method, from X-rays and CT to MRI, ultrasound, and PET, and explains how AI researchers can use that physics to build more reliable medical AI. It is a teaching review aimed at AI developers who work with medical images but lack a physics background.","discovery_kind":"review","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The review asserts a universal benefit of physics-informed AI in medical imaging but provides no systematic comparative evidence; its own Section 3.4 lists conditions under which physics constraints hurt, so the abstract's claim is stronger than the review establishes.","rationale":"I agree with the reader's conditional verdict, but I locate the load-bearing weakness differently. The reader's weakest assumption was the accuracy of the physics tutorial, which is a genuine concern for a paper that aims to be a pedagogical reference. However, the central claim of the abstract is a causal/empirical claim about what physics-informed AI does: it enhances trustworthiness and robustness, especially with limited data. A review supporting that claim needs controlled evidence, ideally a systematic comparison showing that physics-informed methods outperform data-only baselines in the claimed regimes. This paper provides a taxonomy of PIML approaches and selected encouraging examples, but it does not synthesize quantitative evidence or reconcile the counterexamples it lists in Section 3.4, where over-regularization and incomplete physics knowledge are acknowledged. The paper's own caveats therefore undercut the strength of the abstract's language. The factual physics errors are important and should be corrected, but they concern the tutorial's reliability rather than the empirical support for the central claim. Since the reader's verdict is already CONDITIONAL and requires revision, this concern does not move the verdict; it sharpens the reason for the condition. I recommend keeping CONDITIONAL, with revision requirements covering both the physics corrections and the need to soften or more carefully qualify the general enhancement claim.","tokens_in":19148,"tokens_out":4218,"duration_ms":46810,"concrete_test":"Audit the PIML papers cited in Sections 2 and 3 ([29], [30], [70], [73], [86], [101], plus any medical-imaging examples in the surveys [9] and [48]). For each, tabulate: (1) whether a no-physics baseline is included; (2) whether the evaluation uses a limited-data or out-of-distribution setting; (3) whether the physics-informed variant significantly improves the reported metric. If fewer than a clear majority of applications satisfy all three, the abstract's universal 'enhances trustworthiness and robustness' should be revised to a conditional 'can enhance' with the supporting conditions stated.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The abstract's central claim is that integrating physics knowledge into AI algorithms 'enhances their trustworthiness and robustness in medical imaging, especially in scenarios with limited data availability.' For this claim to be supported by a review, the cited PIML applications need to show, in controlled comparisons, that physics-informed models beat purely data-driven baselines on robustness, out-of-distribution, and limited-data metrics. The paper instead groups methods by bias type (observational, learning, inductive, Sections 3.1-3.3), cites general surveys [9,48], and gives selected examples ([29,30,70,73,86,101]) without synthesizing effect sizes or baseline comparisons. The paper itself flags countervailing evidence in Section 3.4: excessive constraints can cause over-regularization, explainability and uncertainty remain limitations, and physics knowledge is incomplete. A reader cannot tell whether the claimed enhancement is a general property of physics-informed learning or an artifact of selected favorable cases. This is the load-bearing gap. The physics errors identified by the reader (Section 2.2's ionization description and the Hounsfield unit range) are real and matter for the paper's handbook purpose, but they are not the main threat to the abstract's causal claim; even a perfectly accurate tutorial would leave the enhancement claim unevidenced as stated.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"This manuscript is a narrative review intended as a pedagogical handbook for AI researchers entering medical imaging. It surveys the physical principles behind each clinical imaging modality (visible-light, X-ray, CT, mammography, fluoroscopy, MRI, SPECT, PET, ultrasound, and combined systems), discusses image-quality challenges and artifacts, and then introduces physics-informed machine learning (PIML), grouping methods into observational, learning, and inductive biases. The paper's central claim, stated in the abstract and repeated in Section 4, is that integrating physics knowledge into AI algorithms enhances their trustworthiness and robustness, particularly when training data are scarce.","tokens_in":19334,"tokens_out":4409,"duration_ms":44575,"significance":"If the central claim were established, this review would be a useful orientation for AI researchers and could serve as a bridge between the medical imaging physics community and the machine learning community. The manuscript has genuine strengths: a modality-by-modality organization that is easy to navigate; concrete examples of physics-informed reconstruction and generation across modalities; an explicit taxonomy of PIML approaches; and a clearly written list of challenges and limitations in Section 3.4. However, the abstract's causal claim is presented as a general result while the supporting evidence is a curated set of examples rather than a systematic comparison; moreover, at least one substantive physics error appears in the tutorial portion. Both issues are fixable in revision, but they need to be addressed before the review can serve as a reliable reference.","major_comments":[{"comment":"The manuscript asserts that integrating physics knowledge into AI algorithms 'enhances their trustworthiness and robustness in medical imaging, especially in scenarios with limited data availability.' This claim is stated as a general result, but the review does not supply controlled comparisons or a quantitative synthesis: the cited examples, such as [29, 30, 70, 73, 86, 101], demonstrate feasibility on selected tasks, not superiority over purely data-driven baselines on out-of-distribution or limited-data metrics. Section 3.4 itself concedes that excessive constraints can cause over-regularization and that explainability and uncertainty remain limitations. Please soften the abstract and conclusion to 'can enhance' or add a systematic evidence table with baseline comparisons, so that the message matches the evidence presented.","section":"Abstract and Section 4"},{"comment":"The description of X-ray production is physically incorrect and uncited: the text states that 'electrons are produced due to the ionization of nitrogen and oxygen atoms, which attract positive ions to the cathode, and therefore inject electrons that are accelerated to the anode.' The standard account is that diagnostic X-ray tubes generate electrons via thermionic emission from a heated filament (cathode), and these electrons are then accelerated toward the anode. Because the paper's stated purpose is to provide authoritative physical foundations for AI researchers, this error is load-bearing and must be corrected and referenced, ideally to the manuscript's own primary source [18].","section":"Section 2.2"},{"comment":"The statement that CT voxel values 'ranging from -1000 to 1000' represent the Hounsfield Unit scale is an oversimplification that could mislead AI researchers who normalize or interpret CT data. Air is approximately -1000 HU and water 0 HU, but dense cortical bone and metal can exceed +1000 HU, commonly reaching values around +3000 HU depending on the scanner, reconstruction kernel, and object composition. Please replace this with a more precise statement about the conventional calibration points and the practical range of CT numbers.","section":"Section 2.2.2"},{"comment":"The challenges paragraph explicitly states that 'incorporating excessive constraints during training can lead to over-fitting and over-regularization' and that explainability, uncertainty, and incomplete physics knowledge remain limitations. These caveats are not carried into the abstract or Section 4, where the benefit of physics integration is stated without qualification. Please connect Section 3.4 explicitly to the central claim so that the review's overall message is internally consistent.","section":"Section 3.4"}],"minor_comments":[{"comment":"The paragraph beginning 'Optimizing US image quality involves selecting appropriate settings for the specific anatomical area being examined' appears twice with nearly identical wording later in the same section; please remove the duplicate.","section":"Section 2.5"},{"comment":"'Frecuency' should be 'Frequency' in the axis labels.","section":"Figures 1 and 2"},{"comment":"'the higher spacial frequencies are in the periphery' should read 'the higher spatial frequencies are in the periphery.'","section":"Section 2.3"},{"comment":"'prosprocedural imaging evaluation' should be 'postprocedural imaging evaluation.'","section":"Section 2.2.4"},{"comment":"'Sef-adaptive PINNs' should be 'Self-adaptive PINNs.'","section":"Section 3.2"},{"comment":"'AI has holds significant potential' should be 'AI holds significant potential.'","section":"Section 2.2.3"},{"comment":"When naming the Hounsfield Unit, consider crediting Sir Godfrey Hounsfield explicitly, as the current phrasing 'after one of the main developers of this technology' is unnecessarily vague.","section":"Section 2.2.2"}],"recommendation":"major_revision","confidential_remarks":null},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Bottom line: this is a useful survey for AI researchers coming into medical imaging, but it is not yet the trusted physics reference it wants to be. The modality-by-modality structure is clear, the links to recent generative and physics-informed work are well chosen, and the taxonomy of observational, learning, and inductive biases is a sensible organizing device, even though it is borrowed from refs [9, 48]. Credit where due: the bibliography is relevant and mostly current, and the examples across CT, MRI, PET, and ultrasound give a reader a quick map of where physics-informed methods are being tried. The MRI and nuclear medicine sections are solid, and the authors do acknowledge in Section 3.4 that constraints can over-regularize and that explainability and uncertainty remain open. That honesty matters.\n\nThe soft spots are real, and they matter in proportion to the paper's stated purpose. Section 2.2's account of X-ray tube electron production via ionization of nitrogen and oxygen atoms is wrong, or at best nonstandard, and it is uncited. Thermionic emission from a heated cathode is the mechanism in a conventional X-ray tube; the text as written would actively mislead the audience the paper is trying to educate. The Hounsfield scale simplification to -1000 to 1000 is a minor over-simplification, because bone can go well above 1000, but when combined with the duplicated paragraph in the ultrasound section, it points to a lack of careful proofreading. None of this destroys the survey's value, but a handbook with an uncorrected physics error cannot be recommended without revision.\n\nThe bigger issue is the abstract's causal claim: integrating physics knowledge into AI algorithms enhances trustworthiness and robustness, especially under limited data. The review supports this with selected examples, not with controlled comparisons or effect sizes, and Section 3.4 itself lists conditions under which physics constraints hurt. So the conclusion is plausible and probably true in many settings, but as stated it is stronger than the evidence assembled. That is a framing problem, not a fatal flaw in the whole project.\n\nWho is this for? An AI graduate student or engineer entering medical imaging who wants a big-picture orientation. It is not for physicists or for readers needing systematic evidence. With corrections to Section 2.2 and a softer abstract, I would feel comfortable recommending it. As is, it deserves a serious referee to catch exactly these issues, so I would not desk reject it.","headline":"A useful but uneven pedagogical review of medical imaging physics for AI researchers; the central claim about physics improving trustworthiness is plausible but overgeneralized, and a few real physics errors need fixing before it can serve as the trusted handbook it aims to be.","tokens_in":19887,"tokens_out":2591,"would_cite":false,"duration_ms":26527,"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 review argues that embedding physics knowledge into medical imaging AI algorithms makes them more trustworthy and robust, especially when data are scarce.","keywords":["physics-informed machine learning","medical imaging","trustworthy AI","generative models","image reconstruction","imaging physics","limited data","explainability"],"falsifier":"Run a controlled head-to-head on a public low-dose CT or under-sampled MRI benchmark: train the same architecture with and without a physics-informed loss, forward model, or acquisition-matched noise schedule, and test both on an out-of-distribution set from a different scanner or dose level. If the physics-enhanced model is not consistently more robust or more accurate, the paper's central claim is not supported.","tokens_in":18917,"feed_emoji":"🩻","tokens_out":5776,"duration_ms":57929,"temperature":0.7,"pith_summary":"This paper sets out to establish that AI systems for medical imaging become more trustworthy—more robust, more explainable, and safer for clinical use—when they are built on an accurate understanding of the physics that produces each image. It argues that the gap between AI research and clinical translation is partly due to developers lacking this physical background, and it supplies a modality-by-modality tutorial of that physics, from visible-light and X-ray imaging through MRI, nuclear medicine, and ultrasound. It then connects each modality's acquisition physics to the current AI toolkit, especially generative models and reconstruction algorithms, and reviews physics-informed machine learning, where physical laws are inserted as data constraints, loss terms, or architectural inductive biases. The payoff the paper claims is that these physics-based constraints make AI models behave more reliably precisely in the settings where medical AI struggles: scarce labeled data, low-dose or accelerated acquisitions, and out-of-distribution inputs. A sympathetic reader would take the paper's central thesis as a handbook-level argument that imaging physics is not optional background for medical AI but a load-bearing component of trustworthiness.","feed_headline":"Physics-informed AI can make medical imaging more trustworthy","feed_subtitle":"A review argues that physics constraints make imaging AI more robust, especially when data are scarce.","key_machinery":"The machinery that carries the argument is the mapping of every clinical imaging modality onto its governing physical process, combined with a taxonomy of physics-informed machine learning. For each modality the paper identifies the physical effect that forms the image—absorption and scattering for X-ray, Hounsfield-unit attenuation for CT, T1/T2 relaxation and k-space sampling for MRI, gamma emission and coincidence detection for PET/SPECT, echo propagation for ultrasound—and then classifies ways of injecting that physics into a learning algorithm: observational bias (the data themselves reflect physics), learning bias (physics-based penalty terms in the loss), and inductive bias (physics hard-wired into the architecture). This two-part structure is what lets the review move from 'physics describes the image' to 'physics can regularize, constrain, and explain the model'.","core_discovery":"The paper's central claim is that the trustworthiness of AI in medical imaging is substantially determined by how faithfully the model respects the physical processes that create the image. Radiographs and CT images are records of X-ray attenuation and scattering; MRI images are reconstructed from spatial-frequency data in k-space whose sampling pattern is governed by gradient physics; PET images are formed from coincidence detection of annihilation photons; ultrasound images are built from reflected acoustic pulses. The review argues that AI developers who ignore these processes can be misled by artifacts, overtrust synthetic images, and produce models that fail on out-of-distribution clinical data. The discovery it offers, as a review synthesis, is that the same physics that constrains image formation can be turned into algorithmic constraints—through synthetic data that mimics acquisition, loss functions that penalize physics violations, or architectures that encode physical invariances—and that this is a concrete route to robustness and explainability in limited-data regimes. In short, the paper claims that medical imaging physics is not a static background fact but an exploitable resource for making AI models more reliable.","pith_inferences":["The paper leaves implicit that the same logic implies physics should become a standard for evaluating synthetic medical images: a generative model that violates known acquisition physics could be rejected before clinician review.","A testable extension is to use modality-specific physics as a zero-shot or few-shot prior, so a model trained on one scanner could be adapted to another scanner with almost no labeled data.","If the physics tutorial is meant to be a handbook, it invites a companion set of worked examples that convert each modality's equations into code-level constraints, turning the review's thesis into directly actionable recipes.","The argument also suggests an educational consequence: medical imaging AI curricula should treat imaging physics not as a prerequisite nicety but as a core component of trustworthy-model design."],"forward_implications":["Generative models for medical images can be made physically plausible by embedding acquisition physics, such as a noise schedule that mimics ultrasound echo attenuation, so synthetic images are less likely to mislead clinicians.","Image reconstruction algorithms that include a physical forward model can recover high-quality images from lower-dose or under-sampled data, supporting reductions in radiation exposure and scan time.","Physics-based constraints act as a regularizer, which should reduce overfitting and improve generalization when labeled medical data are scarce.","Models with explicit physical structure are more explainable, because failures and predictions can be traced back to a physical quantity such as attenuation or relaxation time.","Physics-informed methods offer a route to robustness against out-of-distribution acquisitions, since the model already knows how scanner settings and patient anatomy affect the image."],"supporting_citations":[{"why":"The primary reference for the physical principles of every imaging modality; the review's tutorial sections rest on it.","marker":"[18]"},{"why":"Defines physics-informed machine learning and the taxonomy of observational, learning, and inductive biases used to organize Section 3.","marker":"[48]"},{"why":"Introduces physics-informed neural networks, the key learning-bias mechanism the review highlights.","marker":"[75]"},{"why":"Surveys physics-informed neural networks for medical image analysis, supporting the paper's claim that these methods apply broadly across modalities.","marker":"[10]"},{"why":"Provides the ultrasound diffusion-model example where a physics-based noise scheduler models echo attenuation.","marker":"[29]"},{"why":"Provides the low-dose PET example where physics-based uncertainty-aware multimodal learning improves robustness to out-of-distribution data.","marker":"[86]"},{"why":"Provides the MRI example of physics-informed motion correction using a physics-informed loss to exclude corrupted k-space lines.","marker":"[30]"},{"why":"Provides the MRI example of k-space acquisition optimization conditioned on imaging physics with a neural ODE.","marker":"[70]"}],"fun_headline_variants":["Physics constraints make medical imaging AI more trustworthy","Leverage imaging physics to build robust medical AI","Physics-informed AI: key to trustworthy medical imaging","Turn imaging physics into AI constraints for reliability"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The argument rests on the assumption that the physics tutorial the paper provides is accurate enough to serve as a trusted reference for AI developers; if a core physical account is wrong, as with its uncited explanation of X-ray electron production by ionizing nitrogen and oxygen molecules, the handbook misleads its intended readers and the claim that physics knowledge improves AI loses its foundation.","fun_headline_variants_meta":{"raw":{"variants":["Physics constraints make medical imaging AI more trustworthy","Leverage imaging physics to build robust medical AI","Physics-informed AI: key to trustworthy medical imaging","Turn imaging physics into AI constraints for reliability"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000222,"raw_usage":{"total_tokens":1431,"prompt_tokens":897,"completion_tokens":534,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":513,"completion_tokens_details":{"reasoning_tokens":476}},"tokens_in":513,"tokens_out":534,"duration_ms":5242,"temperature":1.0,"reasoning_tokens":476,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T05:46:58.429770+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run a controlled head-to-head on a public low-dose CT or under-sampled MRI benchmark: train the same architecture with and without a physics-informed loss, forward model, or acquisition-matched noise schedule, and test both on an out-of-distribution set from a different scanner or dose level. If the physics-enhanced model is not consistently more robust or more accurate, the paper's central claim is not supported.","supporting_citations":[{"cited_title":"Diffusion as sound propagation: Physics-inspired model for ultrasound image generation","cited_arxiv_id":null,"evidence_quote":"Provides the ultrasound diffusion-model example where a physics-based noise scheduler models echo attenuation."},{"cited_title":"Towards lower-dose pet using physics-based uncertainty-aware multimodal learning with robustness to out-of-distribution data","cited_arxiv_id":null,"evidence_quote":"Provides the low-dose PET example where physics-based uncertainty-aware multimodal learning improves robustness to out-of-distribution data."},{"cited_title":"Physics-informed deep learning for motion-corrected reconstruction of quantitative brain mri","cited_arxiv_id":null,"evidence_quote":"Provides the MRI example of physics-informed motion correction using a physics-informed loss to exclude corrupted k-space lines."},{"cited_title":"Learning optimal k-space acquisition and reconstruction using physics-informed neural networks","cited_arxiv_id":null,"evidence_quote":"Provides the MRI example of k-space acquisition optimization conditioned on imaging physics with a neural ODE."}],"review_version":1}