{"id":"69113c93-a832-4028-83b3-e28f11823fa7","arxiv_id":"2509.00304","paper_version":2,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":2.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"AI-based PET reconstruction and resolution enhancement methods are surveyed for long axial field of view scanners and multiplexed multi-tracer imaging, with key challenges identified.","lead":"This review surveys how artificial intelligence is being used to reconstruct and enhance PET images, including on new long-field-of-view scanners that image the whole body at once. It maps recent AI methods for resolution improvement and multi-tracer imaging, and where they fall short before clinical use.","discovery_kind":"review","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Citation mismatch undercuts the review's one concrete LAFOV AI-reconstruction performance claim: the cited ref [27] (FAST-PET simulator) is not the FastPET-on-Quadra study, so the 7-min→20-s / 2.3% figures are unsupported as referenced.","rationale":"The reader's weakest-assumption analysis lands exactly on the FastPET-on-Quadra attribution. I checked the reference list: [27] is the FAST-PET analytical simulator, not a deep-learning reconstruction study. This is the one concrete quantitative claim in the AI reconstruction section, so it carries more weight than a mere typo. The proposed test is a direct citation-content check. If the correct source exists, the issue is a fixable citation error; if not, the quantitative example is unsubstantiated and the sentence should be removed. The broad thesis—AI is increasingly important in PET reconstruction and LAFOV imaging—remains plausible and is supported by many other cited works (e.g., [26], [28], [30]), so no rejection is warranted. CONDITIONAL (i.e., revise before relying on the review) remains the right disposition. My independent pass did not find a more load-bearing technical flaw.","tokens_in":10302,"tokens_out":4980,"duration_ms":60999,"concrete_test":"Download the full text of reference [27] via its Crossref DOI and search for the strings 'Quadra', '20 seconds', '2.3%', and 'FastPET'. If none of the quantitative terms occur in connection with a neural-network reconstruction, the citation is wrong. Then rerun the same search across PubMed/Google Scholar for 'FastPET Quadra reconstruction' to locate the actual source; if a source with the 7-min/20-s/2.3% numbers exists, update the reference; if not, delete or qualify the sentence. A negative result on both searches means the claim should not appear in its current form.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central value of this review is as a trustworthy secondary map of the PET-AI literature. The weakest load-bearing point is in 'Deep learned PET reconstruction': the sentence claiming that FastPET was extended to the Biograph Vision Quadra with a reconstruction-time reduction from 7 minutes to 20 seconds and an average absolute image difference of 2.3% is referenced to [27]. Reference [27] is Li et al., 'FAST (fast analytical simulator of tracer)-PET: an accurate and efficient PET analytical simulation tool' (Phys. Med. Biol. 2024). A simulation-tool paper is not the primary source for a neural-network reconstruction on Quadra, and the title/scope give no reason to expect the quoted timing or difference figures. Because these numbers are the only quantitative demonstration in the reconstruction section that AI can make LAFOV PET reconstruction clinically practical, the example fails unless a correct citation is supplied. If misattributions like this are widespread, the review's reliability as a field map is materially weakened. This is not an internal mathematical inconsistency, but a verifiable secondary-reporting error. It does not invalidate the broad narrative, but it must be fixed before the review can be relied on.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This manuscript is a narrative review of AI-guided PET image reconstruction and multi-tracer imaging, with particular emphasis on long axial field-of-view (LAFOV) PET/CT. It surveys direct AI reconstruction methods, unrolled iterative approaches, self-supervised and diffusion-model methods, multiplexed PET tracer separation, and AI-based resolution enhancement (positron range correction, depth-of-interaction, motion). The paper concludes that AI methodologies are increasingly central to quantitative PET reconstruction and that clinical deployment of LAFOV-enabled AI techniques still requires validation and scalable solutions.","tokens_in":10615,"tokens_out":5008,"duration_ms":62292,"significance":"If the survey's attributions are accurate, the paper is a useful and timely map of a fast-moving field. Table 1 consolidates architecture, training-pair count, and LAFOV scalability considerations in one place, and the separation of reconstruction families (direct, unrolled, self-supervised, diffusion) is pedagogically valuable. The review also connects LAFOV-specific opportunities (histo-images, dynamic total-body frames, multiplexed PET) to existing SAFOV methods. However, the value of a secondary review depends heavily on citation accuracy; the one clearly identifiable misattribution is therefore a substantive concern rather than a cosmetic one.","major_comments":[{"comment":"The sentence stating that FastPET was extended to the Biograph Vision Quadra 'achieving a reconstruction time reduction from 7 minutes to 20 seconds post-scan, with an average absolute image difference of only 2.3%' is cited to reference [27]. Reference [27] is Li et al., 'FAST (fast analytical simulator of tracer)-PET: an accurate and efficient PET analytical simulation tool' (Phys. Med. Biol. 2024). That paper describes a simulation tool and does not report FastPET-on-Quadra reconstruction timing or image-difference figures. This is the only concrete quantitative demonstration in the reconstruction section that AI can make LAFOV reconstruction clinically practical. The claim is unsupported as referenced and must be corrected by supplying the correct primary source or by removing the specific numbers.","section":"Deep learned PET reconstruction, p. 4"},{"comment":"The section interleaves simulation-based studies and clinical or preclinical studies without consistently stating the evidence level. For example, the Deep-PRC positron range correction work (refs. [60]–[61]) is presented as recovering 18F-compatible resolution for 68Ga, but the underlying methodology is trained and validated on Monte Carlo simulations; the clinical-scanner extension is a conference contribution. Because the review's message that AI can enhance PET resolution is a central claim, the text should explicitly label which results are simulation-based and which are demonstrated on measured data. This will prevent over-reading of the current evidence and is essential for the review to serve as a trustworthy field map.","section":"AI Methods for Image Resolution Enhancement, p. 6"}],"minor_comments":[{"comment":"Typo: 'multiplexed imagin g' should read 'multiplexed imaging'.","section":"Section heading, p. 4"},{"comment":"The phrase 'in the order of tens of 3D data/images' is ambiguous. Please specify whether this means tens of training subjects, tens of reconstructed volumes, or tens of patches/slices, and state the units in Table 1 consistently.","section":"Introduction to reconstruction section, p. 4"},{"comment":"The 'Training pairs used' column mixes different units (2D image pairs, 3D images, slices, subjects). Adding explicit units in each row or a footnote would improve comparability.","section":"Table 1"},{"comment":"Several conference-proceedings entries are incomplete, e.g., ref. [44] lists only 'SPIE; 2023' without title or page numbers. Please complete the bibliographic details for all references.","section":"References"},{"comment":"The phrase 'has a spatial resolution of an equivalent scanner with a shorter axial field of view' is ambiguous; it likely means 'comparable to' rather than identical. Consider rewording for clarity.","section":"Abstract / Clinical care points"}],"recommendation":"major_revision","confidential_remarks":"The paper appears to be a commissioned review for a clinical imaging journal. The self-citation density is high but not unusual for a review written by active researchers in the field. The central concern is citation accuracy: I was able to verify one clear misattribution of a quantitative performance claim, and I recommend that the editor ask the authors to audit all numerical claims and their references before publication. The scope and structure are otherwise appropriate, and the review is likely to be useful once the referencing is corrected."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The short version: this is a competent, readable review that will be useful to people entering the PET-AI space. The comparison table is genuinely helpful—parameter counts, training-pair counts, and LAFOV scalability notes are not compiled anywhere else at this granularity. The mPET section is current and fair, and the deep-learning positron-range-correction coverage is better than most reviews. I'd trust the broad narrative.\n\nThe soft spot is real and specific. The sentence in the deep-learned reconstruction section says FastPET was extended to Quadra with reconstruction time down from 7 min to 20 s and mean absolute image difference 2.3%, citing ref [27]. Ref [27] is Li et al., FAST-PET, a simulation-tool paper. That is not the source for that result. I checked the reference; it's about an analytical simulator, not a neural network on Quadra. The numbers might be right—they probably come from a different FastPET follow-up—but as cited, this is unsupported. Since this is the only quantitative 'AI makes LAFOV reconstruction practical' example in the section, the review needs a correct citation before it can be relied on as a precise map.\n\nOther than that, the weaknesses are typical for this genre: no structured search methodology, some selection bias toward the authors' own groups, and a few claims that are a bit sweeping (e.g., 'AI is now central'). None of that is disqualifying for a clinical review.\n\nVerdict: deserving of serious peer review. The error is correctable and doesn't invalidate the review's thesis. For a reader, I'd say worth reading for the table and the mPET overview; expect to double-check any specific number before quoting it.","headline":"Useful narrative review of AI in LAFOV PET with a concrete citation error in the one quantitative reconstruction claim; fix that and it's a solid field map.","tokens_in":11083,"tokens_out":1452,"would_cite":false,"duration_ms":17460,"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":"AI-guided PET reconstruction is becoming central to long-axial-field-of-view scanners, with deep learning promising faster, sharper, and multi-tracer images—if clinical validation catches up.","keywords":["long-axial field-of-view PET","artificial intelligence","PET image reconstruction","multiplexed PET imaging","image resolution enhancement","positron range correction","deep learning"],"falsifier":"Check reference [27] for the reported FastPET numbers; if they are absent, re-run FastPET on a held-out LAFOV patient cohort and compare reconstruction time and mean absolute difference against clinical OSEM. The review's central speed and quality claim is falsified if the time saving disappears or the image difference exceeds the reported roughly 2.3% on out-of-distribution data.","tokens_in":10253,"feed_emoji":"🩻","tokens_out":6379,"duration_ms":73026,"temperature":0.7,"pith_summary":"Long-axial field-of-view (LAFOV) PET/CT scanners collect far more counts than standard scanners, but they do not automatically produce sharper images: their spatial resolution matches shorter scanners, and their larger data volumes make reconstruction slower. This review argues that artificial intelligence, particularly trained neural networks, is becoming the central tool for closing that gap. AI methods can reconstruct LAFOV data in seconds instead of minutes, recover resolution lost to positron range and other physical effects, and separate signals from multiple injected tracers in a single scan. The paper does not claim these methods are ready for routine clinical use; it stresses that generalization outside training data, tracer-specific behavior, and scalability to large 3D volumes remain open problems. A sympathetic reader would take the review's thesis as: AI is the bridge that turns LAFOV's sensitivity advantage into practical clinical gains.","feed_headline":"AI can cut LAFOV PET reconstruction from 7 minutes to 20 seconds","feed_subtitle":"Deep learning also sharpens resolution and separates multiple tracers—if validation catches up.","key_machinery":"The key object is the trained neural network as a replacement for, or accelerator of, the PET reconstruction operator. FastPET's U-Net maps histo-images plus attenuation maps directly to OSEM-quality images; encoder-decoder networks map rebinned 2D sinograms to vendor reconstructions; unrolled networks fold iterative expectation-maximization steps into the network so that PET physics informs the learning; self-supervised deep image priors and model-informed diffusion methods work without paired training data. For multiplexed PET, kinetic-model-informed deep learning and deep image priors carry the separation of multiple tracer signals. What these mechanisms share is replacing a hand-built sy","core_discovery":"The review's core claim is that AI-guided reconstruction is no longer a niche standard-field-of-view experiment but a central enabling technology for LAFOV PET/CT. It organizes the field into direct methods (e.g., FastPET, an image-to-image network that maps histo-images and attenuation maps to OSEM-quality reconstructions), unrolled iterative methods that embed PET physics into the network, and self-supervised methods that need no paired training data; it extends the same taxonomy to multiplexed PET, where kinetic-model-informed networks and deep image priors separate simultaneous tracers. For resolution, it points to deep-learning positron range correction as a practical alternative to com","pith_inferences":["The review's emphasis on out-of-distribution failure suggests a concrete extension: a standardized hard-case benchmark of LAFOV scans with unusual body habitus, high noise, or novel tracers, used to compare direct, unrolled, and self-supervised methods.","If AI positron range correction matures, it could uncouple radiotracer choice from physics-driven resolution limits, possibly shifting clinical preference toward 68Ga-labelled agents over 18F for some oncology applications.","The 20-second reconstruction figure, if it holds, implies a workflow shift: reconstruction would no longer gate scan duration, enabling online quality assurance and potentially real-time adaptive acquisition.","A testable extension is to re-verify the review's quantitative claims, such as the 2.3% image difference, against the primary sources before using them in clinical or regulatory planning."],"forward_implications":["If direct AI reconstruction performs as reported on LAFOV scanners, reconstruction time drops from roughly 7 minutes to about 20 seconds, making dynamic multi-frame whole-body studies clinically feasible.","If AI resolution enhancement, especially positron range correction, works across radionuclides, isotopes such as 68Ga and 82Rb could be used without accepting their usual resolution penalty, widening tracer choice.","If multiplexed PET separation matures, multiple tracers can be injected and imaged in one session, reducing repeated CT radiation exposure and image-registration errors from sequential scans.","If learning-based methods are relied on, operators must expect failures when inputs fall outside training distributions, so validation and fallback to conventional OSEM reconstruction remain necessary.","LAFOV's higher sensitivity combined with AI reconstruction could enable ultra-low-dose scans, opening PET to healthy control cohorts for building physiological baselines."],"supporting_citations":[{"why":"Supplies the taxonomy of direct, unrolled, and self-supervised AI reconstruction paradigms that organize the whole review.","marker":"[21]"},{"why":"Original FastPET method using histo-images and attenuation images as inputs to accelerate OSEM reconstruction.","marker":"[26]"},{"why":"Cited as the source for the claim that FastPET on a LAFOV scanner cuts reconstruction from 7 minutes to 20 seconds with 2.3% average difference; the cited paper is actually a PET analytical simulation tool, so this citation carries the speed claim but appears mismatched.","marker":"[27]"},{"why":"Demonstrates direct encoder-decoder AI reconstruction on LAFOV sinograms and shows image quality deteriorates outside the training distribution.","marker":"[28]"},{"why":"Proposes deep progressive learning, a two-stage unrolled plus enhancement approach applied to LAFOV total-body data.","marker":"[30]"},{"why":"Kinetic model-informed deep learning for multiplexed PET separation, giving the review its main example of combining kinetic priors with neural networks.","marker":"[45]"},{"why":"Self-supervised parametric map estimation with a deep image prior for multiplexed PET separation without pre-training.","marker":"[46]"},{"why":"Introduces Deep-PRC, a CNN strategy for positron range correction of 68Ga in preclinical scanners.","marker":"[60]"},{"why":"Updates Deep-PRC to clinical scanners, the key evidence that learned positron range correction can scale beyond preclinical systems.","marker":"[61]"},{"why":"Provides the LAFOV performance baseline, including sensitivity and resolution characteristics, that motivates the need for AI reconstruction and enhancement.","marker":"[5]"}],"fun_headline_variants":["AI-guided PET: 7-min LAFOV scans to 20 sec","AI sharpens LAFOV PET while cutting time to 20 sec","LAFOV PET: AI reconstruction cuts time and sharpens images","Deep learning speeds LAFOV PET, sharpens multi-tracer","AI reconstructs LAFOV PET in 20 sec, sharpens multi-tracer"],"cache_read_input_tokens":2688,"weakest_assumption_plain":"The review's map of the field is only as trustworthy as its secondary reporting; for example, it attributes to reference [27] a FastPET result (7 minutes to 20 seconds, 2.3% average difference) that the cited source—a simulation tool—does not appear to contain, so if such misattributions are widespread, the claimed gains could be overstated.","fun_headline_variants_meta":{"raw":{"variants":["AI-guided PET: 7-min LAFOV scans to 20 sec","AI sharpens LAFOV PET while cutting time to 20 sec","LAFOV PET: AI reconstruction cuts time and sharpens images","Deep learning speeds LAFOV PET, sharpens multi-tracer","AI reconstructs LAFOV PET in 20 sec, sharpens multi-tracer"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001017,"raw_usage":{"total_tokens":4053,"prompt_tokens":590,"completion_tokens":3463,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":334,"completion_tokens_details":{"reasoning_tokens":3365}},"tokens_in":334,"tokens_out":3463,"duration_ms":28594,"temperature":1.0,"reasoning_tokens":3365,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-05T13:43:09.432801+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Check reference [27] for the reported FastPET numbers; if they are absent, re-run FastPET on a held-out LAFOV patient cohort and compare reconstruction time and mean absolute difference against clinical OSEM. The review's central speed and quality claim is falsified if the time saving disappears or the image difference exceeds the reported roughly 2.3% on out-of-distribution data.","supporting_citations":[{"cited_title":"AI for PET image reconstruction","cited_arxiv_id":null,"evidence_quote":"Supplies the taxonomy of direct, unrolled, and self-supervised AI reconstruction paradigms that organize the whole review."},{"cited_title":"FastPET: Near Real -Time Reconstruction of PET Histo -Image Data Using a Neural Network","cited_arxiv_id":null,"evidence_quote":"Original FastPET method using histo-images and attenuation images as inputs to accelerate OSEM reconstruction."},{"cited_title":"FAST (fast analytical simulator of tracer)-PET: an accurate and efficient PET analytical simulation tool","cited_arxiv_id":null,"evidence_quote":"Cited as the source for the claim that FastPET on a LAFOV scanner cuts reconstruction from 7 minutes to 20 seconds with 2.3% average difference; the cited paper is actually a PET analytical simulation tool, so this citation carries the speed claim but appears mismatched."},{"cited_title":"An encoder -decoder network for direct image reconstruction on sinograms of a long axial field of view PET","cited_arxiv_id":null,"evidence_quote":"Demonstrates direct encoder-decoder AI reconstruction on LAFOV sinograms and shows image quality deteriorates outside the training distribution."},{"cited_title":"PET image reconstruction with deep progressive learning","cited_arxiv_id":null,"evidence_quote":"Proposes deep progressive learning, a two-stage unrolled plus enhancement approach applied to LAFOV total-body data."},{"cited_title":"Kinetic model-informed deep learning for multiplexed PET image separation","cited_arxiv_id":null,"evidence_quote":"Kinetic model-informed deep learning for multiplexed PET separation, giving the review its main example of combining kinetic priors with neural networks."},{"cited_title":"Self -supervised parametric map estimation for multiplexed PET with a deep image prior","cited_arxiv_id":null,"evidence_quote":"Self-supervised parametric map estimation with a deep image prior for multiplexed PET separation without pre-training."},{"cited_title":"Deep-learning based positron range correction of PET images","cited_arxiv_id":null,"evidence_quote":"Introduces Deep-PRC, a CNN strategy for positron range correction of 68Ga in preclinical scanners."},{"cited_title":"Deep-PRC: A Positron Range Correction Tool for preclinical and clinical PET/CT images","cited_arxiv_id":null,"evidence_quote":"Updates Deep-PRC to clinical scanners, the key evidence that learned positron range correction can scale beyond preclinical systems."}],"review_version":1}