AI-Driven Bone and Marrow Segmentation on FLT-PET/CT: Technical Multi-organ Validation in AML and HCT

AI-Driven Bone and Marrow Segmentation on FLT-PET/CT: Technical Multi-organ Validation in AML and HCT | Research Square window.SnipcartSettings = { analytics: { enabled: false } }; (function() { var accessVector = localStorage.getItem('access_vector') || ''; window.dataLayer = window.dataLayer || []; if (accessVector) { window.dataLayer.push({ user: { profile: { profileInfo: { snid: accessVector } } } }); } })(); (function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0],j=d.createElement(s),dl=l!='dataLayer'?'&l='+l:'';j.async=true;j.src='https://www.googletagmanager.com/gtm.js?id='+i+dl;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-K279D39R'); Browse Preprints In Review Journals COVID-19 Preprints AJE Video Bytes Research Tools Research Promotion AJE Professional Editing AJE Rubriq About Preprint Platform In Review Editorial Policies Our Team Advisory Board Help Center Sign In Submit a Preprint Cite Share Download PDF Research Article AI-Driven Bone and Marrow Segmentation on FLT-PET/CT: Technical Multi-organ Validation in AML and HCT Malakeh Malekzadeh, Hemendra Ghimire, Karteek Popuri, Kazuki Fujita, and 10 more This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-9077609/v1 This work is licensed under a CC BY 4.0 License Status: Under Review Version 1 posted 5 You are reading this latest preprint version Abstract Background [18F] 3'-deoxy-3'-fluorothymidine positron emission tomography (FLT-PET) is valuable for detecting acute myeloid leukemia (AML) and monitoring stem cell engraftment after hematopoietic stem cell transplant (HCT) by assessing cellular proliferation in marrow-rich tissues. Reliable marrow quantification is difficult to achieve, and manual segmentation is impractical in clinical workflows. Most automated tools focus on solid tumors and lack clinical validation for skeletal FLT-PET/CT. This study evaluates deep learning whole-body segmentation and cortical–trabecular marrow quantification on FLT-PET/CT in AML with HCT. Results Twenty refractory AML patients undergoing total marrow and lymphoid irradiation (TMLI) and transplantation were analyzed. From 134 predefined regions, five representative ROIs (spleen, liver, T6, L1, L3) validated agreement with manual segmentation. Automated and manual count measurements showed strong agreement, with a high correlation (r > 0.98, p < 0.0001). Consistent hotspot detection by both methods supports the AI tool’s accuracy and clinical applicability. Small liver/spleen differences and larger positive vertebral trabecular biases were observed. AI cut processing time by ~ 95%, markedly improving efficiency. Conclusion This study provides a technical validation of an AI-driven multi-organ segmentation platform for FLT-PET/CT in AML and HCT, including separate cortical bone and trabecular marrow compartments. The automated approach demonstrated high agreement, excellent reproducibility, and substantial efficiency gains in skeletal marrow and organ quantification. These findings establish a scalable framework for future studies that will correlate FLT-based bone marrow metrics with clinical response and transplant outcomes. Trial registration ClinicalTrials.gov NCT03422731. Registered 6 February 2018, https//www.cancer.gov/research/participate/clinicaltrialssearch/v?id=NCI201701778 [18F] 3'-deoxy-3'-fluorothymidine PET/CT Multiple Organ Segmentation Bone Marrow Examination Figures Figure 1 Figure 2 Figure 3 Figure 4 Figure 5 Introduction [18F] 3'-deoxy-3'-fluorothymidine (F-18 FLT) is a PET tracer of cellular proliferation that is trapped after phosphorylation by thymidine kinase-1 in cycling cells ( 1 , 2 ). It accumulates in hematopoietic bone marrow, spleen, and sites of leukemia involvement, and has been applied to detect acute myeloid leukemia (AML), assess treatment response, and monitor engraftment after hematopoietic cell transplantation (HCT) ( 3 ). However, reliable quantification of FLT uptake in bone marrow is challenging because trabecular spaces are small, anatomically complex, and show low contrast on CT, while manual delineation is time-consuming and impractical for routine clinical use or multi-timepoint analysis. Deep learning based multi organ segmentation has improved performance over classical thresholding, region growing, and atlas based methods, and is increasingly used in diagnostic and radiotherapy workflows ( 4 – 9 ). Existing AI tools for bone and marrow imaging have focused mainly on CT based bone mineral density, bone marrow lesions on MRI, or CT body composition analysis rather than tracer activity measurement, and they generally do not distinguish cortical from trabecular bone. This limitation is important because biologically meaningful marrow proliferation occurs primarily in trabecular compartments. In this study, we evaluated a deep learning based AI platform for automated multi-organ CT segmentation and cortical–trabecular marrow separation, coupled to FLT-PET quantification, in patients with AML undergoing HCT. We compared automated and expert manual segmentations for representative organs and trabecular vertebral regions and characterized skeletal, muscle, fat, and organ FLT uptake using the automated workflow. To our knowledge, this is the first clinical technical validation of whole body cortical and trabecular marrow segmentation for FLT-PET/CT in AML and HCT. Materials and methods Patient Selection Twenty patients with refractory acute myeloid leukemia (AML) scheduled to undergo total marrow and lymphoid irradiation (TMLI) combined with chemotherapy, followed by stem cell transplantation, were included in this analysis at the pre-treatment time point. All patients were enrolled on an ongoing clinical trial of FLT-PET/CT imaging (NCT03422731). Of these, 12 patients underwent whole-body FLT-PET/CT from head to toe, and 8 underwent upper-body imaging from neck to femur. The study was approved by the City of Hope Institutional Review Board (protocol 17222) and all participants provided written informed consent in accordance with institutional and regulatory guidelines. The mean age was 54.8 ± 13.6 years (range 20–75 years), mean height was 171 ± 8 cm, and mean weight was 80.0 ± 14.0 kg. [¹⁸F] FLT-FLT PET/CT Imaging Protocols All imaging was obtained using an integrated PET/CT scanner (Optima 560, GE Medical Systems, USA). [¹⁸F] FLT (2.06 ± 0.82 MBq/kg) in 2–5 mL of normal saline was injected intravenously. PET/CT (Optima 560, GE Medical Systems, USA) imaging was performed one hour after the [¹⁸F] FLT injection, covering the region from the vertex to the upper thigh. A standardized helical CT protocol (140 kVp, standard kernel) was used. The upper-body scan was acquired at ~ 270 mA with a voxel size of 1.38 × 1.38 × 3.27 mm³, while the lower-body scan used ~ 200 mA, a 0.875 pitch, and a voxel size of 0.98 × 0.98 × 3.75 mm³. PET data were then acquired immediately for 1 min per bed position. PET images comprised approximately 600 ± 40 slices for the whole body varying according to patient height, and voxel sizes of 3.65 × 3.65 × 3.27 mm 3 . Manual and automated segmentation For the detailed manual versus automated comparison, five representative ROIs (liver, spleen, T6, L1, and L3 trabecular compartments) were manually delineated in a subset of 8 patients, selected at random from the patient cohort. Manual segmentation and site-wise quantification were performed using Velocity ( Fig. 1 . A.b and Fig. 1 . B.a) , a routine image processing platform within the radiation oncology workflow (Varian Medical Systems, Palo Alto, CA), on CT-scan images. Manual segmentation was carried out by three observers, including two medical imaging experts with varying levels of segmentation experience and one radiologist. The operators performing manual segmentation did not require additional training to use the Velocity software, as it is a platform routinely employed in our research workflow. Automated segmentation was performed using the data analysis facilitation suite (DAFS; Voronoi Health, Canada), ( Fig. 1 . A.a and Fig. 1 . B.c). For DAFS input, no specific data preparation was required; CT and PET images (DICOM format) were simply imported into the software as a common folder. The segmentation covers muscles, fat compartments, major organs, bones (cortical and trabecular), vessels, glands, and pathological regions, with datasets processed with and without upper limbs for standardized quantification. The resulting masks provided volumetric and compositional data across multiple tissue and organ systems, enabling detailed morphometric and metabolic analyses. To this end, each CT annotation and segmentation was inspected via a sagittal, coronal, and axial view of each scan using the quick check quality option; mis-annotations were corrected using the CAST (CT Annotation and Segmentation Tool) feature from DAFS. Using DAFS, a total of 134 anatomical regions can be defined, categorized into skeletal muscle (n = 23), adipose tissue (n = 28), bone (n = 49), and organs, soft tissues, and glands (n = 34). Automated upper-body 3D renderings were generated to depict macroscopic FLT distribution across the trunk ( Fig. 2 . B, a-d) . Cortical bone and trabecular marrow were automatically separated to enable compartment-specific analysis. Trabecular-focused visualization highlighted FLT activity within marrow space, including detailed cortical-trabecular delineation in thoracic and lumbar vertebrae and microscopic FLT mapping of the L1 trabecular region in superior, anterior, and inferior views ( Fig. 2 . C) . To ensure an unbiased comparison between manual and automated segmentation and to avoid variables that could introduce error, Bq/mL values were used. Body-weight–normalized Standardized Uptake Value (SUV) measurements from all ROIs are reported as complementary results. Body-weight–normalized SUV (SUVbw) was calculated using Eq. 1 : \(\:SUV=\:\frac{\text{A}\text{c}\text{t}\text{i}\text{v}\text{i}\text{t}\text{y}\:(\text{B}\text{q}/\text{m}\text{L})\times\:\text{B}\text{o}\text{d}\text{y}\:\text{w}\text{e}\text{i}\text{g}\text{h}\text{t}\:\left(\text{k}\text{g}\right)\times\:1000\:(\text{g}/\text{k}\text{g})\:}{Injected\:Dose\:\left(Bq\right)*{2}^{-\left(\frac{Aquisition\:Time-StartTime}{Half\:Life}\right)}}\) (Eq. 1) In this formulation, the tissue activity concentration (Bq/mL) was obtained from the PET images, and the injected dose (Bq) was decay-corrected to the time of acquisition using the physical half-life of the radiotracer. Body weight (kg) was incorporated into the numerator to normalize uptake across subjects. All SUVbw calculations were performed using Microsoft Excel (Microsoft Corp., USA) formula to ensure consistent quantification across subjects. Statistical Analysis Pearson’s (rₚ) and Spearman’s (rₛ) correlation coefficients, accounting for data normality, along with Bland-Altman (B&A) analysis, were used to compare Bq/ml values across all ROIs between the two methods. Inter-operator reliability was evaluated using the intraclass correlation coefficient (ICC) between three independent operators (medical imaging experts and a radiologist) performing manual segmentation. Normality of cortical and trabecular SUV measurements was assessed using the Shapiro–Wilk test. As normality assumptions were not met, group differences were evaluated with the nonparametric Mann–Whitney U test (two-tailed). a two-sided p value < 0.05 was considered statistically significant. AI-Assisted Editing AI-assisted tools (ChatGPT, OpenAI) were used only for language editing and improving clarity in the manuscript text. All scientific content, analyses, and interpretations were performed and verified by the authors. Results The agreement between automated and manual segmentations was assessed using Pearson’s (rp) and Spearman’s (rs) correlation coefficients to evaluate their linear relationship ( Fig. 3 A, a–e ). Across the five representative ROIs (liver, spleen, and trabecular compartments of T6, L1, and L3), Pearson correlation coefficients (rₚ) ranged from 0.988 to 0.998 (p < 0.0001 for all comparisons). Spearman correlation for the spleen was 0.983, indicating strong monotonic agreement. Moreover, the agreement between manual and automated was evaluated by employing B&A analysis ( Fig. 3 B, a-e ) , where the difference between the two methods (Manual-Automated) was plotted against the mean values of them (Manual+Automated)/2). The mean differences are − 354.5, -448.9, 1332.0 and 961.2 Bq/mL for a) liver, b) spleen, c) T6 trabecular, d) L1 trabecular and e) L3 trabecular, respectively. SUV values for all organ quantified ROIs are illustrated in Figs. 4 and 5 and Tables S1- S3 . Figure 4 a demonstrates distinct trabecular SUV values across the thoracic, lumbar, and sacral spine compared with cortical. The trabecular marrow showed significantly higher values compared with cortical bone (p < 0.0001), (Fig. 4 b ) . Figure 4 c presents skeletal bone SUVs. In our cohort, muscle and fat ROIs showed uniformly low FLT uptake, as expected for non-proliferative tissues ( Fig. 5 a, b, Table S2 ) , confirming that the AI workflow performs robustly across diverse tissue classes and functions as a true multi-organ tool rather than a marrow-only application. Notably, elevated SUVs in lower-body regions defined by − 150 to − 50 HU and in visceral adipose tissue (VAT) were present in approximately 50% of the cohort and are therefore detailed in Tables S2 rather than illustrated in the figure. Among organ SUVs ( Fig. 5 c ) , the bladder showed the highest uptake (65.450 ± 43.39), whereas the skin demonstrated the lowest values (0.299 ± 0.12). Owing to its disproportionately high uptake, the bladder was excluded from the organ plot in Fig. 5 c to avoid scaling distortion; corresponding values are provided in Tables S1–S3 . Manual segmentation required 25 ± 4 minutes on average to delineate five organs (≈ 5 minutes per organ). In the worst-case scenario, manual delineation extended to 96.03 ± 22 minutes, or approximately 19.2 minutes per organ. In contrast, the automated method completed segmentation of 134 ROIs in 115–120 minutes (scan number: 306.42 ± 18.83), equivalent to about 0.83 minutes per ROI. When normalized per region, automation reduced processing time by approximately 84% compared with the average manual workflow, and by up to 96% compared with the worst-case manual scenario. These findings highlight the substantial efficiency gains achieved through automation, especially given that delineation time increases nearly linearly with the number of slices. Moreover, overall inter-observer reliability was excellent, with an averaged-measures ICC = 0.97 (95% CI 0.95–0.98, n = 40 ROIs). These values showed excellent agreement, confirming the reproducibility and minimal operator dependence of the segmentation, establishing a reliable ground truth for validating the automated method. Moreover, manual inspection was performed to verify whether the software correctly segmented the targeted organs. All vertebral cancellous bones were correctly identified except for T9 and L5, which required manual correction due to aortic calcifications at T9 and increased angulation for L5, respectively. Discussion In this study, we evaluated manual versus automated CT-based segmentation methods to quantify activity counts from FLT-PET/CT images in patients with AML undergoing HCT. The AI-driven multi-organ platform showed excellent agreement with expert manual segmentation for both large soft-tissue organs and small trabecular vertebral ROIs, with Pearson correlation coefficients above 0.98 and narrow B&A limits of agreement. Manual inter-observer reliability was also excellent, providing a robust ground truth for validating the automated method. These findings support the feasibility of applying deep learning-based tools for skeletal FLT-PET/CT analysis in hematologic malignancies, consistent with prior advances in multi-organ segmentation and radiotherapy imaging applications.( 4 – 8 , 10 – 12 ) In this study, manual segmentation showed excellent correlation with the automated method across both small (T6, L1, L3 trabecular) and large organs (liver and spleen; r = 0.99-1) ( Fig. 3 A, a-e ) . B&A analysis demonstrated good agreement between methods ( Fig. 3 B, a-e ) , with most measurements falling within the limits of agreement and minimal overall bias. Agreement was strongest for large soft-tissue organs, whereas greater variability was observed in trabecular bone regions, suggesting that automated segmentation may be more challenged by fine bone structures despite overall consistency. In trabecular spinal regions (T6, L1, L3) showed a slight positive bias, likely reflecting sharper bone boundary definition on CT. Cortical thickness appeared greater in the automated segmentation than in the manual segmentation ( Fig. 1 A ) , suggesting cortical overestimation by the automated method, particularly at T6 (1332.0 Bq/mL) and L1/L3 (961.2 Bq/mL). Tighter automated contours yield smaller ROIs and higher apparent FLT uptake, underscoring the sensitivity of PET quantification to segmentation precision. The biological separation of cortical bone and trabecular marrow is essential for accurate interpretation of FLT uptake. Trabecular bone contains hematopoietically active marrow with high cellular proliferation, whereas cortical bone is largely mineralized and minimally proliferative; accordingly, FLT signal-retained after phosphorylation by thymidine kinase-1 during the S-phase-primarily reflects proliferating hematopoietic cells.( 1 – 3 ) Consistent with this mechanism, trabecular SUV measurements provide a more specific representation of marrow proliferative activity than cortical or composite vertebral values, with higher uptake demonstrated in Fig. 4 a, underscoring the clinical relevance of compartment-specific analysis in refractory acute myeloid leukemia, leukemia-niche characterization, and radiation-targeting strategies. ( 13 ) Given the marked radiosensitivity of bone marrow and its role as a dose-limiting organ, accurate assessment of active red marrow is critical for transplantation and dose- toxicity evaluation. ( 14 ). Prior FLT-PET studies, including McGuire et al., have shown dose-dependent suppression of marrow proliferation with vertebral uptake correlating with delivered radiation dose, supporting FLT-PET as a sensitive biomarker of hematopoietic activity.( 15 ) Precise trabecular-specific quantification therefore enables improved evaluation of marrow response, recovery, and clinical outcomes. Three-dimensional automated renderings further illustrate the macroscopic spatial distribution of FLT uptake across the upper-body skeleton, consistently visualizing the spine, hips, and proximal femora from multiple viewing angles ( Fig. 2 B, a-d ) . Global active marrow mapping enables skeleton-wide assessment of hematopoiesis, heterogeneity, disease niches, and post-transplant regeneration beyond localized biopsy. Trabecular-specific segmentation isolates FLT activity from cortical signal ( Fig. 2 C ) , improving biologically precise assessment of microscopic proliferation and reducing structural, partial-volume, and mineral attenuation confounding relevant to therapy and transplantation.( 3 , 7 – 9 , 16 ) In manual segmentation for multi organs, organ boundaries, such as the liver, are often drawn slightly within the true anatomical edge to reduce partial-volume effects. However, due to limited soft-tissue contrast in CT, operators place the contour several pixels within the actual margin (often exceeding the ideal ~ 3-pixel offset). As a result, manual segmentations can fall noticeably within the true organ boundary; consequently, manual segmentation tends to overestimate mean values in soft-tissue regions such as the liver and spleen, as derived from CT images. In contrast, DAFS identifies the external anatomical boundary algorithmically. Ideally, once the exact outer border is defined, the software should then apply a standardized inward offset (e.g., 3 pixels) to minimize partial volume effects while avoiding the operator-dependent inward bias seen in manual segmentation ( Fig. 1 . B.c). No study confirms outer-voxel selection in automated segmentation, but PET boundary voxels commonly show mixed tissue and underestimated activity. ( 17 ) Manual contouring is highly variable, while automated methods consistently capture complete organ boundaries ( 4 ), which may contribute to the lower liver and spleen uptake observed with automated segmentation. SUV values across regions are summarized in Figs. 4 – 5 and Tables S1–S3 . The platform also segmented muscle, adipose tissue, and major organs, with uniformly low FLT uptake in muscle and fat confirming robust multi-organ performance. The prostate exhibited the highest SUV (4.403 ± 4.23) among the organ category, with the substantial variation potentially attributable to size-related differences. Organ segmentation is essential for imaging and therapy but remains limited by traditional thresholding, ( 18 ), graph-cut ( 19 ), region-growing ( 20 ), and atlas-based methods ( 21 ), which show variability, computational burden, and poor performance in irregular or low-contrast anatomy. ( 4 ) Deep learning- particularly transformer models such as Swin UNETR-has improved accuracy, robustness, and generalizability in whole-body PET/CT segmentation. However, standardized evaluation of clinical efficiency, especially in non-solid tumors, is lacking. Current artificial intelligence applications in bone marrow imaging largely focus on magnetic resonance–based density or lesion assessment, ( 4 , 7 , 8 , 12 , 22 ). and DAFS has primarily been used for computed tomography body composition rather than FLT-PET/CT ( 23 ). Most CT segmentation tools cover limited organs, with some reaching ~ 120 structures( 10 ), whereas the proposed software segments 134–152. Because HU-based algorithms cannot separate cortical from trabecular bone, ( 11 ) essential for FLT marrow assessment (10), DAFS offers broader coverage, reproducibility, usability, and compartment-specific distinction. Finally, automated segmentation markedly improved efficiency compared with manual delineation. This highlights its superior scalability for volumetric datasets, where manual delineation time increases almost linearly with the number of slices. Beyond speed, reducing user interaction minimizes operator fatigue and variability, and common challenges in manual workflows. Collectively, these findings underscore the practicality of automated segmentation for large-scale or time-sensitive imaging studies, where accuracy and consistency throughput are critical for clinical and research applications. This work still has several limitations. Firstly, our validation was evaluated using data from a single institution and a single scanner. As a result, more studies using data from several other institutions are needed to demonstrate the generalizability of the results. The ROI mismatch between manual and the automatic software system for T9 and L5 trabecular and cortical segmentation is also a limitation in this study. Moreover, because the whole-body scan contains overlapping slices between the upper- and lower-body acquisitions, redundant slices can be removed in future studies, leaving only the selected non-overlapping slices for import into the DAFS software for quantification. Despite these limitations, our findings indicate that AI-based multi-organ segmentation with explicit cortical and trabecular marrow compartmentalization can provide accurate, reproducible, and efficient FLT-PET/CT quantification in patients with AML undergoing HCT. By extending prior multi-organ and bone marrow segmentation work to functional FLT imaging, this study establishes a technical foundation for future investigations that will link regional marrow activity to clinical and biological endpoints, including biopsy targeting, treatment response assessment, and risk stratification in leukemia and transplant care. Conclusion In this study, we implemented and validated a fully automated three-dimensional segmentation and quantification workflow using the DAFS platform to measure regional FLT activity (Bq/mL) on PET/CT in patients with AML undergoing HCT. To our knowledge, this is the first technical multi-organ validation of automated cortical bone and trabecular marrow segmentation for FLT-PET/CT in this setting. The AI-based approach showed excellent agreement with expert manual segmentation while enabling rapid, reproducible, and largely operator-independent assessment of skeletal marrow and organ activity. Among currently available multi-organ segmentation platforms, DAFS is notable for its explicit separation of cortical bone and trabecular marrow, a feature that is essential for biologically specific evaluation of active bone marrow. By accelerating PET analysis and standardizing marrow and organ quantification, this validated workflow provides a scalable foundation for future studies that will link FLT-based marrow metrics to clinical and biological endpoints and may ultimately support noninvasive assessment of functional bone marrow in leukemia and transplant care. Abbreviations FLT-PET [¹⁸F] 3'-deoxy-3'-fluorothymidine positron emission tomography AML acute myeloid leukemia HCT hematopoietic stem cell transplant TMLI total marrow and lymphoid irradiation ROIs regions of interest DAFS data analysis facilitation suite ICC intraclass correlation coefficient TK1 thymidine kinase-1 BM bone marrow CAST CT annotation and segmentation tool SUV standardized uptake value SUVbw body-weight–normalized SUV UB upper body LB lower body All SKM all skeletal muscles All SKM [-29,150] all skeletal muscles within HU -29 to 150 LPECMJR/RPECMJR left/right pectoralis major LPECMNR/RPECMNR left/right pectoralis minor LTEMPORALIS/RTEMPORALIS left/right temporalis LMASSETER/RMASSETER left/right masseter LSCM/RSCM left/right sternocleidomastoid LILIOPSOAS/RILIOPSOAS left/right iliopsoas LUPLGSKM/RUPLGSKM left/right upper limb muscles LLWLGSKM/RLWLGSKM left/right lower limb muscles LASKM/RASKM left/right abdominal muscles AOC aortic calcification CAAC cardiac aggregate calcium VAT visceral adipose tissue VAT [-150, -50] visceral fat within HU -150 to -50 EPAT epicardial/pericardial fat PAAT periaortic fat THAT thoracic fat SAT subcutaneous fat SAT [-190, -30] subcutaneous fat within HU -190 to -30 LASAT/RASAT left/right abdominal subcutaneous fat IMAT intramuscular adipose tissue All IMAT [-190, -30] IMAT within HU -190 to -30 LUPLGIMAT/RUPLGIMAT left/right upper limb intramuscular adipose tissue LLWLGIMAT/RLWLGIMAT left/right lower limb intramuscular adipose tissue LAIMAT/RAIMAT left/right abdominal intramuscular adipose tissue Declarations Ethics approval and consent to participate: Patients were enrolled on an ongoing correlative imaging study (ClinicalTrials.gov NCT03422731, City of Hope protocol 17222). The study was approved by the City of Hope Institutional Review Board, and all participants provided written informed consent. Consent for publication No Availability of data and material The data that support the findings of this study are avail­able from the corresponding author upon reasonable request. Competing interests M.F.B. and K.P. are professors at their respective academic institutions and disclose that they are also directors of Voronoi Health Analytics Inc., the developer of the DAFS AI-based automated software platform used in this study. Their contributions to this work were to provide technical guidance on software use, data extraction, and analytic workflow support. They were not involved in the study design, interpretation of results or in decisions regarding manuscript conclusions. Voronoi Health Analytics Inc. as a corporate entity had no role in study design, data collection, manuscript preparation, or publication decisions. The remaining authors declare no competing interests. Funding This work has been supported by NIH grants 2R01CA154491 (PI: S.K.H.) and ONCOTEST (Ghent, Belgium, PI: S.K.H.). The content is solely the responsibility of the authors and does not necessarily represent the official views of the National Institutes of Health. Author’s Contributions Authorship: Conceptualization, M.M., S.K.H.; Methodology, M.M., K.P. M.F.B; Patient Recruitment: M.A.M., J.W., A.S, D.Y.; Formal Analysis, M.M, HG, BC, KP; Investigation: MM, S.K.H.; Manuscript Preparation, M.M.; Manuscript Editing, M.M., S.K.H., M.F.B., GS, JF, KF; Visualization, M.M. and S.K.H.; Supervision, S.K.H; Project Administration, S.K.H.; Funding Acquisition, S.K.H. All authors have read and agreed to the published version of the manuscript. Acknowledgements The authors would like to acknowledge the study participants and the technologists and staff of the imaging center in radiation oncology department (City of Hope National Medical Center) for their valuable assistance in data acquisition and support throughout this research. References Salskov A, Tammisetti VS, Grierson J, Vesselle H, editors. FLT: measuring tumor cell proliferation in vivo with positron emission tomography and 3′-deoxy-3′-[18F] fluorothymidine. Seminars in nuclear medicine. Elsevier; 2007. Shields AF, Grierson JR, Dohmen BM, Machulla H, Stayanoff JC, Lawhorn-Crews JM et al. Imaging proliferation in vivo with [F-18] FLT and positron emission tomography. 1998;4(11):1334–6. Han EJ, Lee B-h, Kim J-A, Park YH, Choi WHJE. Early assessment of response to induction therapy in acute myeloid leukemia using 18F-FLT PET/CT. 2017;7(1):75. Liu X, Qu L, Xie Z, Zhao J, Shi Y, Song ZJBEO. Towards more precise automatic analysis: a systematic review of deep learning-based multi-organ segmentation. 2024;23(1):52. Liu Z, Liu F, Chen W, Liu X, Hou X, Shen J et al. Automatic segmentation of clinical target volumes for post-modified radical mastectomy radiotherapy using convolutional neural networks. 2021;10:581347. Huang D, Bai H, Wang L, Hou Y, Li L, Xia Y et al. The application and development of deep learning in radiotherapy: a systematic review. 2021;20:15330338211016386. Niu X, Huang Y, Li X, Yan W, Lu X, Jia X et al. Development and validation of a fully automated system using deep learning for opportunistic osteoporosis screening using low-dose computed tomography scans. 2023;13(8):5294. Ponnusamy R, Zhang M, Wang Y, Sun X, Chowdhury M, Driban JB et al. Automatic segmentation of bone marrow lesions on MRI using a deep learning method. 2024;11(4):374. Jimenez-Pastor A, Alberich-Bayarri A, Fos-Guarinos B, Garcia-Castro F, Garcia-Juan D, Glocker B, et al. Automated vertebrae localization and identification by decision forests and image-based refinement on real-world. CT data. 2020;125(1):48–56. Trägårdh E, Borrelli P, Kaboteh R, Gillberg T, Ulén J, Enqvist O, et al. RECOMIA—a cloud-based Platf Artif Intell Res nuclear Med Radiol. 2020;7(1):51. Wasserthal J, Breit H-C, Meyer MT, Pradella M, Hinck D, Sauter AW et al. TotalSegmentator: robust segmentation of 104 anatomic structures in CT images. 2023;5(5):e230024. Yazdani E, Karamzadeh-Ziarati N, Cheshmi SS, Sadeghi M, Geramifar P, Vosoughi H et al. Automated segmentation of lesions and organs at risk on [68Ga] Ga-PSMA-11 PET/CT images using self-supervised learning with Swin UNETR. 2024;24(1):30. Magome T, Froelich J, Holtan SG, Takahashi Y, Verneris MR, Brown K, et al. Whole-Body Distribution of Leukemia and Functional Total Marrow Irradiation Based on FLT-PET and Dual-Energy CT. Mol Imaging. 2017;16:1536012117732203. Gulec SA, Mesoloras G, Dezarn WA, McNeillie P, Kennedy AS. Safety and efficacy of Y-90 microsphere treatment in patients with primary and metastatic liver cancer: the tumor selectivity of the treatment as a function of tumor to liver flow ratio. J Transl Med. 2007;5:15. McGuire SM, Menda Y, Boles Ponto LL, Gross B, Buatti J, Bayouth JE. 3'-deoxy-3'-[¹⁸F]fluorothymidine PET quantification of bone marrow response to radiation dose. Int J Radiat Oncol Biol Phys. 2011;81(3):888–93. Jeraj R, Duan F, Mattison RJ, Romanoff J, Kostakoglu L, Arber DA, et al. Early assessment of treatment response in AML using FLT PET/cT: a trial of the ECOG-ACRIN Cancer Research Group (EAI141). Leuk Lymphoma. 2025;66(14):2765–73. Bettinardi V, Castiglioni I, De Bernardi E, Gilardi MJC, Imaging T. PET quantification: strategies for partial volume correction. 2014;2(3):199–218. Saranathan AM, Parente M, editors. Threshold based segmentation method for hyperspectral images. 2013 5Th workshop on hyperspectral image and signal processing: evolution in remote sensing (WHISPERS). IEEE; 2013. Shi J. Malik JJITopa, intelligence m. Normalized cuts image segmentation. 2000;22(8):888–905. Vyavahare AJ, Thool R, editors. Segmentation using region growing algorithm based on CLAHE for medical images. Fourth International Conference on Advances in Recent Technologies in Communication and Computing (ARTCom2012); 2012: IET. Isgum I, Staring M, Rutten A, Prokop M, Viergever MA, Van BJItomi G. Multi-atlas-based segmentation with local decision fusion—application to cardiac and aortic segmentation in CT scans. 2009;28(7):1000–10. Ecabert O, Peters J, Schramm H, Lorenz C, von Berg J, Walker MJ, et al. Automatic model-based segmentation heart CT images. 2008;27(9):1189–201. Anyene I, Caan B, Williams GR, Popuri K, Lenchik L, Giri S et al. Body composition from single versus multi-slice abdominal computed tomography: concordance and associations with colorectal cancer survival. 2022;13(6):2974–84. Supplementary Files Summlementary1.Malekzadehetal.pdf Cite Share Download PDF Status: Under Review Version 1 posted Reviewers agreed at journal 11 Apr, 2026 Reviewers invited by journal 09 Apr, 2026 Editor invited by journal 23 Mar, 2026 Editor assigned by journal 19 Mar, 2026 First submitted to journal 18 Mar, 2026 You are reading this latest preprint version Research Square lets you share your work early, gain feedback from the community, and start making changes to your manuscript prior to peer review in a journal. As a division of Research Square Company, we’re committed to making research communication faster, fairer, and more useful. We do this by developing innovative software and high quality services for the global research community. Our growing team is made up of researchers and industry professionals working together to solve the most critical problems facing scientific publishing. Also discoverable on Platform About Our Team In Review Editorial Policies Advisory Board Help Center Resources Author Services Accessibility API Access RSS feed Manage Cookie Preferences © Research Square 2026 | ISSN 2693-5015 (online) Privacy Policy Terms of Service Do Not Sell My Personal Information {"props":{"pageProps":{"initialData":{"identity":"rs-9077609","acceptedTermsAndConditions":true,"allowDirectSubmit":false,"archivedVersions":[],"articleType":"Research Article","associatedPublications":[],"authors":[{"id":620532857,"identity":"e95fe75b-6b19-4f24-a73c-465329da859c","order_by":0,"name":"Malakeh Malekzadeh","email":"","orcid":"","institution":"City of Hope National Medical Center","correspondingAuthor":false,"prefix":"","firstName":"Malakeh","middleName":"","lastName":"Malekzadeh","suffix":""},{"id":620532858,"identity":"6d3da428-3537-4a9e-9e6b-ed132063c67b","order_by":1,"name":"Hemendra Ghimire","email":"","orcid":"","institution":"City of Hope National Medical Center","correspondingAuthor":false,"prefix":"","firstName":"Hemendra","middleName":"","lastName":"Ghimire","suffix":""},{"id":620532859,"identity":"55cab876-fde9-4feb-a8df-3b237825913e","order_by":2,"name":"Karteek Popuri","email":"","orcid":"","institution":"Memorial University: Memorial University of Newfoundland","correspondingAuthor":false,"prefix":"","firstName":"Karteek","middleName":"","lastName":"Popuri","suffix":""},{"id":620532860,"identity":"0690e7d2-bd33-44cf-8159-9b91c82eb048","order_by":3,"name":"Kazuki Fujita","email":"","orcid":"","institution":"City of Hope National Medical Center","correspondingAuthor":false,"prefix":"","firstName":"Kazuki","middleName":"","lastName":"Fujita","suffix":""},{"id":620532861,"identity":"74f8b133-0507-41e1-bd03-2bba56824677","order_by":4,"name":"Amandeep Salhotra","email":"","orcid":"","institution":"City of Hope National Medical Center","correspondingAuthor":false,"prefix":"","firstName":"Amandeep","middleName":"","lastName":"Salhotra","suffix":""},{"id":620532862,"identity":"7fb9de32-0073-4baa-ba66-c1435136d4be","order_by":5,"name":"Dave Yamauchi","email":"","orcid":"","institution":"City of Hope National Medical Center","correspondingAuthor":false,"prefix":"","firstName":"Dave","middleName":"","lastName":"Yamauchi","suffix":""},{"id":620532863,"identity":"254c2b56-51d2-4eb5-bbc0-0af7774f1c8a","order_by":6,"name":"Bihong Chen","email":"","orcid":"","institution":"City of Hope National Medical Center","correspondingAuthor":false,"prefix":"","firstName":"Bihong","middleName":"","lastName":"Chen","suffix":""},{"id":620532864,"identity":"e7240d49-7669-4c11-9241-283d2741d405","order_by":7,"name":"Jerry Froelich","email":"","orcid":"","institution":"University of Minnesota","correspondingAuthor":false,"prefix":"","firstName":"Jerry","middleName":"","lastName":"Froelich","suffix":""},{"id":620532865,"identity":"0baef282-1b29-4735-9d91-90a4fefe3e73","order_by":8,"name":"Guy Storme","email":"","orcid":"","institution":"Universitair Ziekenhuis Brussel","correspondingAuthor":false,"prefix":"","firstName":"Guy","middleName":"","lastName":"Storme","suffix":""},{"id":620532866,"identity":"c3abb31d-659d-4343-b390-fe19e54b4c68","order_by":9,"name":"Anthony Stein","email":"","orcid":"","institution":"City of Hope National Medical Center","correspondingAuthor":false,"prefix":"","firstName":"Anthony","middleName":"","lastName":"Stein","suffix":""},{"id":620532867,"identity":"82fc2914-dd22-4f12-bcf7-45d4f7b90cab","order_by":10,"name":"Mirza Faisal Beg","email":"","orcid":"","institution":"Simon Fraser University","correspondingAuthor":false,"prefix":"","firstName":"Mirza","middleName":"Faisal","lastName":"Beg","suffix":""},{"id":620532868,"identity":"12da2a2f-a670-492d-8881-0885027c5e3b","order_by":11,"name":"Jeffrey Wong","email":"","orcid":"","institution":"City of Hope National Medical Center","correspondingAuthor":false,"prefix":"","firstName":"Jeffrey","middleName":"","lastName":"Wong","suffix":""},{"id":620532869,"identity":"45556d5c-590f-483c-920d-be06f63ad606","order_by":12,"name":"Monzr M Al Malki","email":"","orcid":"","institution":"City of Hope National Medical Center","correspondingAuthor":false,"prefix":"","firstName":"Monzr","middleName":"M Al","lastName":"Malki","suffix":""},{"id":620532870,"identity":"8419251a-f749-4fa4-bd8e-6b04f064a87b","order_by":13,"name":"Susanta K Hui","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAAAnElEQVRIiWNgGAWjYDACCQaGAx8YGBJI03JwBslamHlI0iI/u8fwsE2NTR4De+/jF0RpMbhzxuBwzrG0Ygae42YWxGmRyDE4nNtwOLFBIo3NgDiHzQBqsSRJC8MNoBZGiBbmB8Q57EZawcGeY2mJbTzH2IizRH5G8uYPP2psEvvZ25g/EKcHBoBWsEmQpgUISLVlFIyCUTAKRgoAAF2ZLm1FDkaEAAAAAElFTkSuQmCC","orcid":"https://orcid.org/0000-0002-1394-3724","institution":"City of Hope National Medical Center","correspondingAuthor":true,"prefix":"","firstName":"Susanta","middleName":"K","lastName":"Hui","suffix":""}],"badges":[],"createdAt":"2026-03-10 01:00:14","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-9077609/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-9077609/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":107256936,"identity":"520eba61-d42f-4c79-9b80-a377aa3a8606","added_by":"auto","created_at":"2026-04-19 12:25:12","extension":"png","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":2610216,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eOverall segmentation of vertebral structures and Thoracoabdominal organs. A)\u003c/strong\u003e Vertebral cortical and trabecular bone segmentation shown in axial, sagittal, and coronal views, including (a) automated segmentation and (b) manual segmentation. \u003cstrong\u003eB)\u003c/strong\u003eLiver and spleen segmentation on CT images, including (a) manual segmentation, (b) PET-CT map in the Velocity workspace, (c) automated segmentation, and (d) PET-CT map in the DAFS software.\u003c/p\u003e","description":"","filename":"Fig1.png","url":"https://assets-eu.researchsquare.com/files/rs-9077609/v1/499fe4fddc04f42d52776b03.png"},{"id":107256942,"identity":"800b3492-a329-410a-9989-a4c0b2d82f44","added_by":"auto","created_at":"2026-04-19 12:25:13","extension":"png","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":612475,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003e(A) Segmentation of thoracic and lumbar vertebrae and the liver.\u003c/strong\u003e \u003cstrong\u003e(a)\u003c/strong\u003e Manual segmentation performed in Velocity and \u003cstrong\u003e(b)\u003c/strong\u003eautomated segmentation generated by DAFS in whole-body FLT PET/CT images.\u003cstrong\u003e (B) Macroscopic FLT distribution (3D):\u003cbr\u003e\nAutomated upper-body 3D renderings illustrating the spatial distribution of FLT uptake across the skeleton, with multi-view visualization of the spine, hips, and proximal femora: \u003c/strong\u003e(a) anterior, (b) anterior oblique, (c) posterior, and (d) posterior oblique. \u003cstrong\u003e(C) Compartment-specific visualization highlighting FLT activity within trabecular bone marrow, enabling detailed assessment of marrow proliferation independent of cortical bone signal:\u003c/strong\u003e (a) central skeletal segmentation; (b) cortical (purple arrow) and trabecular (orange arrow) compartments of the thoracic and lumbar vertebrae, including a 3D rendering of the trabecular compartment; (c) microscopic-level FLT-PET mapping of the trabecular region of the L1 vertebra (superior view); (d) anterior view; and (e) inferior view.\u003c/p\u003e","description":"","filename":"Fig2.png","url":"https://assets-eu.researchsquare.com/files/rs-9077609/v1/795c980b253e70293a8d8eee.png"},{"id":107256943,"identity":"57071a43-97d2-46e4-9c2d-e3c95eae9224","added_by":"auto","created_at":"2026-04-19 12:25:13","extension":"png","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":1332597,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eCorrelation (Pearson's \u003c/strong\u003e\u003cem\u003e\u003cstrong\u003e𝑟\u003c/strong\u003e\u003c/em\u003e\u003csub\u003e\u003cem\u003e\u003cstrong\u003e𝑝\u003c/strong\u003e\u003c/em\u003e\u003c/sub\u003e\u003cstrong\u003e and Spearman's \u003c/strong\u003e\u003cem\u003e\u003cstrong\u003er\u003c/strong\u003e\u003c/em\u003e\u003csub\u003e\u003cem\u003e\u003cstrong\u003es\u003c/strong\u003e\u003c/em\u003e\u003c/sub\u003e\u003cstrong\u003e) and Bland-Altman (B\u0026amp;A) analyses of mean Bq/mL values comparing automated and manual segmentations. (A) \u003c/strong\u003eCorrelation plots for manual segmentation from 8 patients in (a) liver, (b) spleen, (c) T6 trabecular bone, (d) L1 trabecular bone, and (e) L3 trabecular bone regions versus automated segmentation.\u003cstrong\u003e (B) \u003c/strong\u003eB\u0026amp;A analysis, showing the difference between the two segmentation methods (Manual − Automated), plotted against the mean values (Manual+Automated)/2) in (a) liver, (b) spleen, (c) T6 trabecular bone, (d) L1 trabecular bone, and (e) L3 trabecular. The mean difference (solid line) and limits of agreement (dotted line) are shown. B\u0026amp;A, Bland \u0026amp; Altman; r\u003csub\u003ep\u003c/sub\u003e, Pearson’s correlation coefficient, r\u003csub\u003es, \u003c/sub\u003eSpearman’s correlation coefficient.\u003c/p\u003e","description":"","filename":"Fig3.png","url":"https://assets-eu.researchsquare.com/files/rs-9077609/v1/9adcdb468e075eea288e9ce1.png"},{"id":107256897,"identity":"b26362e1-8f4f-4162-911b-938e0fc4366e","added_by":"auto","created_at":"2026-04-19 12:25:02","extension":"png","order_by":4,"title":"Figure 4","display":"","copyAsset":false,"role":"figure","size":575874,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eSkeletal SUVs from FLT-PET/CT. (a) \u003c/strong\u003eSpinal cortical, and trabecular compartments across thoracic levels T1–T12, lumbar levels L1–L5, and the sacrum.\u003cstrong\u003e (b) \u003c/strong\u003eViolin plots showing SUV distributions across cortical bone and trabecular marrow compartments. The trabecular marrow shows significantly higher values than cortical bone.**** (p \u0026lt; 0.0001), \u003cstrong\u003e(c)\u003c/strong\u003e Skeletal bone groups. UB=Upper body; Lower body=LB.\u003c/p\u003e","description":"","filename":"Fig4.png","url":"https://assets-eu.researchsquare.com/files/rs-9077609/v1/298d90062a3d407163bc3a0c.png"},{"id":107256938,"identity":"d8099d53-c8a7-4358-b04c-77198d193f10","added_by":"auto","created_at":"2026-04-19 12:25:12","extension":"png","order_by":5,"title":"Figure 5","display":"","copyAsset":false,"role":"figure","size":1323273,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eSUVs from FLT-PET/CT across (a) skeletal muscle, (b) adipose tissue, and (c) other organs\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eUB=Upper body; Lower body=LB, All SKM = all skeletal muscles; All SKM[-29,150] = all skeletal muscles within HU -29 to 150; No Arms = excluding upper limbs; LPECMJR/RPECMJR = left/right pectoralis major; LPECMNR/RPECMNR = left/right pectoralis minor; LTEMPORALIS/RTEMPORALIS = left/right temporalis; LMASSETER/RMASSETER = left/right masseter; LSCM/RSCM = left/right sternocleidomastoid; LILIOPSOAS/RILIOPSOAS = left/right iliopsoas; LUPLGSKM/RUPLGSKM = left/right upper limb muscles; LLWLGSKM/RLWLGSKM = left/right lower limb muscles; LASKM/RASKM = left/right abdominal muscles; AOC-U-CAAC= enables AOC-CAAC-based agatston score; VAT = visceral adipose tissue; VAT[-150,-50] = visceral fat within HU -150 to -50; EPAT = epicardial/pericardial fat; PAAT = periaortic fat; THAT = thoracic fat; SAT = subcutaneous fat; SAT[-190,-30] = subcutaneous fat within HU -190 to -30; LASAT/RASAT = left/right abdominal subcutaneous fat. IMAT = intramuscular adipose tissue; All_IMAT [-190, -30] = IMAT within HU -190 to -30; LUPLGIMAT/RUPLGIMAT = left/right upper limb IMAT; LLWLGIMAT/RLWLGIMAT = left/right lower limb IMAT; LAIMAT/RAIMAT = left/right abdominal IMAT.\u003c/p\u003e","description":"","filename":"Fig5.png","url":"https://assets-eu.researchsquare.com/files/rs-9077609/v1/2e1c6e969979829425458302.png"},{"id":107705300,"identity":"5f7e38e3-9fa5-446b-8028-26b51931c508","added_by":"auto","created_at":"2026-04-24 09:11:13","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":10179774,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-9077609/v1/00585a76-c37a-4e97-9090-e97d4aeec877.pdf"},{"id":107256941,"identity":"347f71b8-d473-4059-943f-fa52782d5913","added_by":"auto","created_at":"2026-04-19 12:25:13","extension":"pdf","order_by":1,"title":"","display":"","copyAsset":false,"role":"supplement","size":80361,"visible":true,"origin":"","legend":"","description":"","filename":"Summlementary1.Malekzadehetal.pdf","url":"https://assets-eu.researchsquare.com/files/rs-9077609/v1/ccc8c0002e93324dab36e3ec.pdf"}],"financialInterests":"","formattedTitle":"AI-Driven Bone and Marrow Segmentation on FLT-PET/CT: Technical Multi-organ Validation in AML and HCT","fulltext":[{"header":"Introduction","content":"\u003cp\u003e[18F] 3'-deoxy-3'-fluorothymidine (F-18 FLT) is a PET tracer of cellular proliferation that is trapped after phosphorylation by thymidine kinase-1 in cycling cells (\u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e, \u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e). It accumulates in hematopoietic bone marrow, spleen, and sites of leukemia involvement, and has been applied to detect acute myeloid leukemia (AML), assess treatment response, and monitor engraftment after hematopoietic cell transplantation (HCT) (\u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e3\u003c/span\u003e). However, reliable quantification of FLT uptake in bone marrow is challenging because trabecular spaces are small, anatomically complex, and show low contrast on CT, while manual delineation is time-consuming and impractical for routine clinical use or multi-timepoint analysis.\u003c/p\u003e \u003cp\u003eDeep learning based multi organ segmentation has improved performance over classical thresholding, region growing, and atlas based methods, and is increasingly used in diagnostic and radiotherapy workflows (\u003cspan additionalcitationids=\"CR5 CR6 CR7 CR8\" citationid=\"CR4\" class=\"CitationRef\"\u003e4\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR9\" class=\"CitationRef\"\u003e9\u003c/span\u003e). Existing AI tools for bone and marrow imaging have focused mainly on CT based bone mineral density, bone marrow lesions on MRI, or CT body composition analysis rather than tracer activity measurement, and they generally do not distinguish cortical from trabecular bone. This limitation is important because biologically meaningful marrow proliferation occurs primarily in trabecular compartments.\u003c/p\u003e \u003cp\u003eIn this study, we evaluated a deep learning based AI platform for automated multi-organ CT segmentation and cortical\u0026ndash;trabecular marrow separation, coupled to FLT-PET quantification, in patients with AML undergoing HCT. We compared automated and expert manual segmentations for representative organs and trabecular vertebral regions and characterized skeletal, muscle, fat, and organ FLT uptake using the automated workflow. To our knowledge, this is the first clinical technical validation of whole body cortical and trabecular marrow segmentation for FLT-PET/CT in AML and HCT.\u003c/p\u003e"},{"header":"Materials and methods","content":"\u003cdiv id=\"Sec3\" class=\"Section2\"\u003e \u003ch2\u003ePatient Selection\u003c/h2\u003e \u003cp\u003eTwenty patients with refractory acute myeloid leukemia (AML) scheduled to undergo total marrow and lymphoid irradiation (TMLI) combined with chemotherapy, followed by stem cell transplantation, were included in this analysis at the pre-treatment time point. All patients were enrolled on an ongoing clinical trial of FLT-PET/CT imaging (NCT03422731). Of these, 12 patients underwent whole-body FLT-PET/CT from head to toe, and 8 underwent upper-body imaging from neck to femur. The study was approved by the City of Hope Institutional Review Board (protocol 17222) and all participants provided written informed consent in accordance with institutional and regulatory guidelines. The mean age was 54.8\u0026thinsp;\u0026plusmn;\u0026thinsp;13.6 years (range 20\u0026ndash;75 years), mean height was 171\u0026thinsp;\u0026plusmn;\u0026thinsp;8 cm, and mean weight was 80.0\u0026thinsp;\u0026plusmn;\u0026thinsp;14.0 kg.\u003c/p\u003e \u003c/div\u003e\n\u003ch3\u003e[¹⁸F] FLT-FLT PET/CT Imaging Protocols\u003c/h3\u003e\n\u003cp\u003eAll imaging was obtained using an integrated PET/CT scanner (Optima 560, GE Medical Systems, USA). [\u0026sup1;⁸F] FLT (2.06\u0026thinsp;\u0026plusmn;\u0026thinsp;0.82 MBq/kg) in 2\u0026ndash;5 mL of normal saline was injected intravenously. PET/CT (Optima 560, GE Medical Systems, USA) imaging was performed one hour after the [\u0026sup1;⁸F] FLT injection, covering the region from the vertex to the upper thigh. A standardized helical CT protocol (140 kVp, standard kernel) was used. The upper-body scan was acquired at ~\u0026thinsp;270 mA with a voxel size of 1.38 \u0026times; 1.38 \u0026times; 3.27 mm\u0026sup3;, while the lower-body scan used\u0026thinsp;~\u0026thinsp;200 mA, a 0.875 pitch, and a voxel size of 0.98 \u0026times; 0.98 \u0026times; 3.75 mm\u0026sup3;. PET data were then acquired immediately for 1 min per bed position. PET images comprised approximately 600\u0026thinsp;\u0026plusmn;\u0026thinsp;40 slices for the whole body varying according to patient height, and voxel sizes of 3.65 \u0026times; 3.65 \u0026times; 3.27 mm\u003csup\u003e3\u003c/sup\u003e.\u003c/p\u003e\n\u003ch3\u003eManual and automated segmentation\u003c/h3\u003e\n\u003cp\u003eFor the detailed manual versus automated comparison, five representative ROIs (liver, spleen, T6, L1, and L3 trabecular compartments) were manually delineated in a subset of 8 patients, selected at random from the patient cohort. Manual segmentation and site-wise quantification were performed using Velocity \u003cb\u003e(\u003c/b\u003eFig.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003e.\u003cb\u003eA.b and\u003c/b\u003e Fig.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003e.\u003cb\u003eB.a)\u003c/b\u003e, a routine image processing platform within the radiation oncology workflow (Varian Medical Systems, Palo Alto, CA), on CT-scan images. Manual segmentation was carried out by three observers, including two medical imaging experts with varying levels of segmentation experience and one radiologist. The operators performing manual segmentation did not require additional training to use the Velocity software, as it is a platform routinely employed in our research workflow.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eAutomated segmentation was performed using the data analysis facilitation suite (DAFS; Voronoi Health, Canada), \u003cb\u003e(\u003c/b\u003eFig.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003e.\u003cb\u003eA.a and\u003c/b\u003e Fig.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003e.\u003cb\u003eB.c).\u003c/b\u003e For DAFS input, no specific data preparation was required; CT and PET images (DICOM format) were simply imported into the software as a common folder. The segmentation covers muscles, fat compartments, major organs, bones (cortical and trabecular), vessels, glands, and pathological regions, with datasets processed with and without upper limbs for standardized quantification. The resulting masks provided volumetric and compositional data across multiple tissue and organ systems, enabling detailed morphometric and metabolic analyses. To this end, each CT annotation and segmentation was inspected via a sagittal, coronal, and axial view of each scan using the quick check quality option; mis-annotations were corrected using the CAST (CT Annotation and Segmentation Tool) feature from DAFS. Using DAFS, a total of 134 anatomical regions can be defined, categorized into skeletal muscle (n\u0026thinsp;=\u0026thinsp;23), adipose tissue (n\u0026thinsp;=\u0026thinsp;28), bone (n\u0026thinsp;=\u0026thinsp;49), and organs, soft tissues, and glands (n\u0026thinsp;=\u0026thinsp;34).\u003c/p\u003e \u003cp\u003eAutomated upper-body 3D renderings were generated to depict macroscopic FLT distribution across the trunk \u003cb\u003e(\u003c/b\u003eFig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003e. \u003cb\u003eB, a-d)\u003c/b\u003e. Cortical bone and trabecular marrow were automatically separated to enable compartment-specific analysis. Trabecular-focused visualization highlighted FLT activity within marrow space, including detailed cortical-trabecular delineation in thoracic and lumbar vertebrae and microscopic FLT mapping of the L1 trabecular region in superior, anterior, and inferior views \u003cb\u003e(\u003c/b\u003eFig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003e. \u003cb\u003eC)\u003c/b\u003e.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eTo ensure an unbiased comparison between manual and automated segmentation and to avoid variables that could introduce error, Bq/mL values were used. Body-weight\u0026ndash;normalized Standardized Uptake Value (SUV) measurements from all\u003c/p\u003e \u003cp\u003eROIs are reported as complementary results. Body-weight\u0026ndash;normalized SUV (SUVbw) was calculated using \u003cb\u003eEq.\u0026nbsp;1\u003c/b\u003e:\u003c/p\u003e \u003cp\u003e \u003cspan class=\"InlineEquation\"\u003e \u003cspan class=\"mathinline\"\u003e\\(\\:SUV=\\:\\frac{\\text{A}\\text{c}\\text{t}\\text{i}\\text{v}\\text{i}\\text{t}\\text{y}\\:(\\text{B}\\text{q}/\\text{m}\\text{L})\\times\\:\\text{B}\\text{o}\\text{d}\\text{y}\\:\\text{w}\\text{e}\\text{i}\\text{g}\\text{h}\\text{t}\\:\\left(\\text{k}\\text{g}\\right)\\times\\:1000\\:(\\text{g}/\\text{k}\\text{g})\\:}{Injected\\:Dose\\:\\left(Bq\\right)*{2}^{-\\left(\\frac{Aquisition\\:Time-StartTime}{Half\\:Life}\\right)}}\\)\u003c/span\u003e \u003c/span\u003e (Eq.\u0026nbsp;1)\u003c/p\u003e \u003cp\u003e In this formulation, the tissue activity concentration (Bq/mL) was obtained from the PET images, and the injected dose (Bq) was decay-corrected to the time of acquisition using the physical half-life of the radiotracer. Body weight (kg) was incorporated into the numerator to normalize uptake across subjects. All SUVbw calculations were performed using Microsoft Excel (Microsoft Corp., USA) formula to ensure consistent quantification across subjects.\u003c/p\u003e \u003cdiv id=\"Sec6\" class=\"Section2\"\u003e \u003ch2\u003eStatistical Analysis\u003c/h2\u003e \u003cp\u003ePearson\u0026rsquo;s (rₚ) and Spearman\u0026rsquo;s (rₛ) correlation coefficients, accounting for data normality, along with Bland-Altman (B\u0026amp;A) analysis, were used to compare Bq/ml values across all ROIs between the two methods. Inter-operator reliability was evaluated using the intraclass correlation coefficient (ICC) between three independent operators (medical imaging experts and a radiologist) performing manual segmentation. Normality of cortical and trabecular SUV measurements was assessed using the Shapiro\u0026ndash;Wilk test. As normality assumptions were not met, group differences were evaluated with the nonparametric Mann\u0026ndash;Whitney U test (two-tailed). a two-sided \u003cem\u003ep\u003c/em\u003e value\u0026thinsp;\u0026lt;\u0026thinsp;0.05 was considered statistically significant.\u003c/p\u003e \u003c/div\u003e\n\u003ch3\u003eAI-Assisted Editing\u003c/h3\u003e\n\u003cp\u003eAI-assisted tools (ChatGPT, OpenAI) were used only for language editing and improving clarity in the manuscript text. All scientific content, analyses, and interpretations were performed and verified by the authors.\u003c/p\u003e"},{"header":"Results","content":"\u003cp\u003eThe agreement between automated and manual segmentations was assessed using Pearson\u0026rsquo;s (rp) and Spearman\u0026rsquo;s (rs) correlation coefficients to evaluate their linear relationship \u003cb\u003e(\u003c/b\u003eFig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003eA, a\u0026ndash;e\u003cb\u003e).\u003c/b\u003e Across the five representative ROIs (liver, spleen, and trabecular compartments of T6, L1, and L3), Pearson correlation coefficients (rₚ) ranged from 0.988 to 0.998 (p\u0026thinsp;\u0026lt;\u0026thinsp;0.0001 for all comparisons). Spearman correlation for the spleen was 0.983, indicating strong monotonic agreement. Moreover, the agreement between manual and automated was evaluated by employing B\u0026amp;A analysis \u003cb\u003e(\u003c/b\u003eFig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003eB, a-e\u003cb\u003e)\u003c/b\u003e, where the difference between the two methods (Manual-Automated) was plotted against the mean values of them (Manual+Automated)/2). The mean differences are \u0026minus;\u0026thinsp;354.5, -448.9, 1332.0 and 961.2 Bq/mL for a) liver, b) spleen, c) T6 trabecular, d) L1 trabecular and e) L3 trabecular, respectively.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eSUV values for all organ quantified ROIs are illustrated in Figs.\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e4\u003c/span\u003e and \u003cspan refid=\"Fig5\" class=\"InternalRef\"\u003e5\u003c/span\u003e and \u003cb\u003eTables S1- S3\u003c/b\u003e. Figure\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e4\u003c/span\u003ea demonstrates distinct trabecular SUV values across the thoracic, lumbar, and sacral spine compared with cortical. The trabecular marrow showed significantly higher values compared with cortical bone (p\u0026thinsp;\u0026lt;\u0026thinsp;0.0001), (Fig.\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e4\u003c/span\u003eb\u003cb\u003e)\u003c/b\u003e. Figure\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e4\u003c/span\u003ec presents skeletal bone SUVs. In our cohort, muscle and fat ROIs showed uniformly low FLT uptake, as expected for non-proliferative tissues \u003cb\u003e(\u003c/b\u003eFig.\u0026nbsp;\u003cspan refid=\"Fig5\" class=\"InternalRef\"\u003e5\u003c/span\u003ea, b, \u003cb\u003eTable \u003cspan refid=\"MOESM2\" class=\"InternalRef\"\u003eS2\u003c/span\u003e)\u003c/b\u003e, confirming that the AI workflow performs robustly across diverse tissue classes and functions as a true multi-organ tool rather than a marrow-only application. Notably, elevated SUVs in lower-body regions defined by \u0026minus;\u0026thinsp;150 to \u0026minus;\u0026thinsp;50 HU and in visceral adipose tissue (VAT) were present in approximately 50% of the cohort and are therefore detailed in \u003cb\u003eTables S2\u003c/b\u003e rather than illustrated in the figure. Among organ SUVs \u003cb\u003e(\u003c/b\u003eFig.\u0026nbsp;\u003cspan refid=\"Fig5\" class=\"InternalRef\"\u003e5\u003c/span\u003ec\u003cb\u003e)\u003c/b\u003e, the bladder showed the highest uptake (65.450\u0026thinsp;\u0026plusmn;\u0026thinsp;43.39), whereas the skin demonstrated the lowest values (0.299\u0026thinsp;\u0026plusmn;\u0026thinsp;0.12). Owing to its disproportionately high uptake, the bladder was excluded from the organ plot in Fig.\u0026nbsp;\u003cspan refid=\"Fig5\" class=\"InternalRef\"\u003e5\u003c/span\u003ec to avoid scaling distortion; corresponding values are provided in \u003cb\u003eTables S1\u0026ndash;S3\u003c/b\u003e.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eManual segmentation required 25\u0026thinsp;\u0026plusmn;\u0026thinsp;4 minutes on average to delineate five organs (\u0026asymp;\u0026thinsp;5 minutes per organ). In the worst-case scenario, manual delineation extended to 96.03\u0026thinsp;\u0026plusmn;\u0026thinsp;22 minutes, or approximately 19.2 minutes per organ. In contrast, the automated method completed segmentation of 134 ROIs in 115\u0026ndash;120 minutes (scan number: 306.42\u0026thinsp;\u0026plusmn;\u0026thinsp;18.83), equivalent to about 0.83 minutes per ROI. When normalized per region, automation reduced processing time by approximately 84% compared with the average manual workflow, and by up to 96% compared with the worst-case manual scenario. These findings highlight the substantial efficiency gains achieved through automation, especially given that delineation time increases nearly linearly with the number of slices.\u003c/p\u003e \u003cp\u003eMoreover, overall inter-observer reliability was excellent, with an averaged-measures ICC\u0026thinsp;=\u0026thinsp;0.97 (95% CI 0.95\u0026ndash;0.98, n\u0026thinsp;=\u0026thinsp;40 ROIs). These values showed excellent agreement, confirming the reproducibility and minimal operator dependence of the segmentation, establishing a reliable ground truth for validating the automated method. Moreover, manual inspection was performed to verify whether the software correctly segmented the targeted organs. All vertebral cancellous bones were correctly identified except for T9 and L5, which required manual correction due to aortic calcifications at T9 and increased angulation for L5, respectively.\u003c/p\u003e"},{"header":"Discussion","content":"\u003cp\u003eIn this study, we evaluated manual versus automated CT-based segmentation methods to quantify activity counts from FLT-PET/CT images in patients with AML undergoing HCT. The AI-driven multi-organ platform showed excellent agreement with expert manual segmentation for both large soft-tissue organs and small trabecular vertebral ROIs, with Pearson correlation coefficients above 0.98 and narrow B\u0026amp;A limits of agreement. Manual inter-observer reliability was also excellent, providing a robust ground truth for validating the automated method. These findings support the feasibility of applying deep learning-based tools for skeletal FLT-PET/CT analysis in hematologic malignancies, consistent with prior advances in multi-organ segmentation and radiotherapy imaging applications.(\u003cspan additionalcitationids=\"CR5 CR6 CR7\" citationid=\"CR4\" class=\"CitationRef\"\u003e4\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR8\" class=\"CitationRef\"\u003e8\u003c/span\u003e, \u003cspan additionalcitationids=\"CR11\" citationid=\"CR10\" class=\"CitationRef\"\u003e10\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR12\" class=\"CitationRef\"\u003e12\u003c/span\u003e)\u003c/p\u003e \u003cp\u003eIn this study, manual segmentation showed excellent correlation with the automated method across both small (T6, L1, L3 trabecular) and large organs (liver and spleen; r\u0026thinsp;=\u0026thinsp;0.99-1) \u003cb\u003e(\u003c/b\u003eFig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003eA, a-e\u003cb\u003e)\u003c/b\u003e. B\u0026amp;A analysis demonstrated good agreement between methods \u003cb\u003e(\u003c/b\u003eFig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003eB, a-e\u003cb\u003e)\u003c/b\u003e, with most measurements falling within the limits of agreement and minimal overall bias. Agreement was strongest for large soft-tissue organs, whereas greater variability was observed in trabecular bone regions, suggesting that automated segmentation may be more challenged by fine bone structures despite overall consistency.\u003c/p\u003e \u003cp\u003eIn trabecular spinal regions (T6, L1, L3) showed a slight positive bias, likely reflecting sharper bone boundary definition on CT. Cortical thickness appeared greater in the automated segmentation than in the manual segmentation \u003cb\u003e(\u003c/b\u003eFig.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003eA\u003cb\u003e)\u003c/b\u003e, suggesting cortical overestimation by the automated method, particularly at T6 (1332.0 Bq/mL) and L1/L3 (961.2 Bq/mL). Tighter automated contours yield smaller ROIs and higher apparent FLT uptake, underscoring the sensitivity of PET quantification to segmentation precision.\u003c/p\u003e \u003cp\u003eThe biological separation of cortical bone and trabecular marrow is essential for accurate interpretation of FLT uptake. Trabecular bone contains hematopoietically active marrow with high cellular proliferation, whereas cortical bone is largely mineralized and minimally proliferative; accordingly, FLT signal-retained after phosphorylation by thymidine kinase-1 during the S-phase-primarily reflects proliferating hematopoietic cells.(\u003cspan additionalcitationids=\"CR2\" citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e3\u003c/span\u003e) Consistent with this mechanism, trabecular SUV measurements provide a more specific representation of marrow proliferative activity than cortical or composite vertebral values, with higher uptake demonstrated in Fig.\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e4\u003c/span\u003ea, underscoring the clinical relevance of compartment-specific analysis in refractory acute myeloid leukemia, leukemia-niche characterization, and radiation-targeting strategies. (\u003cspan citationid=\"CR13\" class=\"CitationRef\"\u003e13\u003c/span\u003e) Given the marked radiosensitivity of bone marrow and its role as a dose-limiting organ, accurate assessment of active red marrow is critical for transplantation and dose- toxicity evaluation. (\u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e14\u003c/span\u003e). Prior FLT-PET studies, including McGuire et al., have shown dose-dependent suppression of marrow proliferation with vertebral uptake correlating with delivered radiation dose, supporting FLT-PET as a sensitive biomarker of hematopoietic activity.(\u003cspan citationid=\"CR15\" class=\"CitationRef\"\u003e15\u003c/span\u003e) Precise trabecular-specific quantification therefore enables improved evaluation of marrow response, recovery, and clinical outcomes. Three-dimensional automated renderings further illustrate the macroscopic spatial distribution of FLT uptake across the upper-body skeleton, consistently visualizing the spine, hips, and proximal femora from multiple viewing angles \u003cb\u003e(\u003c/b\u003eFig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003eB, a-d\u003cb\u003e)\u003c/b\u003e. Global active marrow mapping enables skeleton-wide assessment of hematopoiesis, heterogeneity, disease niches, and post-transplant regeneration beyond localized biopsy. Trabecular-specific segmentation isolates FLT activity from cortical signal \u003cb\u003e(\u003c/b\u003eFig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003eC\u003cb\u003e)\u003c/b\u003e, improving biologically precise assessment of microscopic proliferation and reducing structural, partial-volume, and mineral attenuation confounding relevant to therapy and transplantation.(\u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e3\u003c/span\u003e, \u003cspan additionalcitationids=\"CR8\" citationid=\"CR7\" class=\"CitationRef\"\u003e7\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR9\" class=\"CitationRef\"\u003e9\u003c/span\u003e, \u003cspan citationid=\"CR16\" class=\"CitationRef\"\u003e16\u003c/span\u003e)\u003c/p\u003e \u003cp\u003eIn manual segmentation for multi organs, organ boundaries, such as the liver, are often drawn slightly within the true anatomical edge to reduce partial-volume effects. However, due to limited soft-tissue contrast in CT, operators place the contour several pixels within the actual margin (often exceeding the ideal\u0026thinsp;~\u0026thinsp;3-pixel offset). As a result, manual segmentations can fall noticeably within the true organ boundary; consequently, manual segmentation tends to overestimate mean values in soft-tissue regions such as the liver and spleen, as derived from CT images. In contrast, DAFS identifies the external anatomical boundary algorithmically. Ideally, once the exact outer border is defined, the software should then apply a standardized inward offset (e.g., 3 pixels) to minimize partial volume effects while avoiding the operator-dependent inward bias seen in manual segmentation \u003cb\u003e(\u003c/b\u003eFig.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003e.\u003cb\u003eB.c).\u003c/b\u003e No study confirms outer-voxel selection in automated segmentation, but PET boundary voxels commonly show mixed tissue and underestimated activity. (\u003cspan citationid=\"CR17\" class=\"CitationRef\"\u003e17\u003c/span\u003e) Manual contouring is highly variable, while automated methods consistently capture complete organ boundaries (\u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e4\u003c/span\u003e), which may contribute to the lower liver and spleen uptake observed with automated segmentation.\u003c/p\u003e \u003cp\u003eSUV values across regions are summarized in Figs.\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e4\u003c/span\u003e\u0026ndash;\u003cspan refid=\"Fig5\" class=\"InternalRef\"\u003e5\u003c/span\u003e and \u003cb\u003eTables S1\u0026ndash;S3\u003c/b\u003e. The platform also segmented muscle, adipose tissue, and major organs, with uniformly low FLT uptake in muscle and fat confirming robust multi-organ performance. The prostate exhibited the highest SUV (4.403\u0026thinsp;\u0026plusmn;\u0026thinsp;4.23) among the organ category, with the substantial variation potentially attributable to size-related differences.\u003c/p\u003e \u003cp\u003eOrgan segmentation is essential for imaging and therapy but remains limited by traditional thresholding, (\u003cspan citationid=\"CR18\" class=\"CitationRef\"\u003e18\u003c/span\u003e), graph-cut (\u003cspan citationid=\"CR19\" class=\"CitationRef\"\u003e19\u003c/span\u003e), region-growing (\u003cspan citationid=\"CR20\" class=\"CitationRef\"\u003e20\u003c/span\u003e), and atlas-based methods (\u003cspan citationid=\"CR21\" class=\"CitationRef\"\u003e21\u003c/span\u003e), which show variability, computational burden, and poor performance in irregular or low-contrast anatomy. (\u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e4\u003c/span\u003e) Deep learning- particularly transformer models such as Swin UNETR-has improved accuracy, robustness, and generalizability in whole-body PET/CT segmentation. However, standardized evaluation of clinical efficiency, especially in non-solid tumors, is lacking. Current artificial intelligence applications in bone marrow imaging largely focus on magnetic resonance\u0026ndash;based density or lesion assessment, (\u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e4\u003c/span\u003e, \u003cspan citationid=\"CR7\" class=\"CitationRef\"\u003e7\u003c/span\u003e, \u003cspan citationid=\"CR8\" class=\"CitationRef\"\u003e8\u003c/span\u003e, \u003cspan citationid=\"CR12\" class=\"CitationRef\"\u003e12\u003c/span\u003e, \u003cspan citationid=\"CR22\" class=\"CitationRef\"\u003e22\u003c/span\u003e). and DAFS has primarily been used for computed tomography body composition rather than FLT-PET/CT (\u003cspan citationid=\"CR23\" class=\"CitationRef\"\u003e23\u003c/span\u003e).\u003c/p\u003e \u003cp\u003eMost CT segmentation tools cover limited organs, with some reaching\u0026thinsp;~\u0026thinsp;120 structures(\u003cspan citationid=\"CR10\" class=\"CitationRef\"\u003e10\u003c/span\u003e), whereas the proposed software segments 134\u0026ndash;152. Because HU-based algorithms cannot separate cortical from trabecular bone, (\u003cspan citationid=\"CR11\" class=\"CitationRef\"\u003e11\u003c/span\u003e) essential for FLT marrow assessment (10), DAFS offers broader coverage, reproducibility, usability, and compartment-specific distinction.\u003c/p\u003e \u003cp\u003eFinally, automated segmentation markedly improved efficiency compared with manual delineation. This highlights its superior scalability for volumetric datasets, where manual delineation time increases almost linearly with the number of slices.\u003c/p\u003e \u003cp\u003eBeyond speed, reducing user interaction minimizes operator fatigue and variability, and common challenges in manual workflows. Collectively, these findings underscore the practicality of automated segmentation for large-scale or time-sensitive imaging studies, where accuracy and consistency throughput are critical for clinical and research applications.\u003c/p\u003e \u003cp\u003eThis work still has several limitations. Firstly, our validation was evaluated using data from a single institution and a single scanner. As a result, more studies using data from several other institutions are needed to demonstrate the generalizability of the results. The ROI mismatch between manual and the automatic software system for T9 and L5 trabecular and cortical segmentation is also a limitation in this study. Moreover, because the whole-body scan contains overlapping slices between the upper- and lower-body acquisitions, redundant slices can be removed in future studies, leaving only the selected non-overlapping slices for import into the DAFS software for quantification.\u003c/p\u003e \u003cp\u003eDespite these limitations, our findings indicate that AI-based multi-organ segmentation with explicit cortical and trabecular marrow compartmentalization can provide accurate, reproducible, and efficient FLT-PET/CT quantification in patients with AML undergoing HCT. By extending prior multi-organ and bone marrow segmentation work to functional FLT imaging, this study establishes a technical foundation for future investigations that will link regional marrow activity to clinical and biological endpoints, including biopsy targeting, treatment response assessment, and risk stratification in leukemia and transplant care.\u003c/p\u003e"},{"header":"Conclusion","content":"\u003cp\u003eIn this study, we implemented and validated a fully automated three-dimensional segmentation and quantification workflow using the DAFS platform to measure regional FLT activity (Bq/mL) on PET/CT in patients with AML undergoing HCT. To our knowledge, this is the first technical multi-organ validation of automated cortical bone and trabecular marrow segmentation for FLT-PET/CT in this setting. The AI-based approach showed excellent agreement with expert manual segmentation while enabling rapid, reproducible, and largely operator-independent assessment of skeletal marrow and organ activity. Among currently available multi-organ segmentation platforms, DAFS is notable for its explicit separation of cortical bone and trabecular marrow, a feature that is essential for biologically specific evaluation of active bone marrow. By accelerating PET analysis and standardizing marrow and organ quantification, this validated workflow provides a scalable foundation for future studies that will link FLT-based marrow metrics to clinical and biological endpoints and may ultimately support noninvasive assessment of functional bone marrow in leukemia and transplant care.\u003c/p\u003e"},{"header":"Abbreviations","content":"\u003cdiv class=\"DefinitionList\"\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003eFLT-PET\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003e[\u0026sup1;⁸F] 3'-deoxy-3'-fluorothymidine positron emission tomography\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003eAML\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003eacute myeloid leukemia\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003eHCT\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003ehematopoietic stem cell transplant\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003eTMLI\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003etotal marrow and lymphoid irradiation\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003eROIs\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003eregions of interest\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003eDAFS\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003edata analysis facilitation suite\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003eICC\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003eintraclass correlation coefficient\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003eTK1\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003ethymidine kinase-1\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003eBM\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003ebone marrow\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003eCAST\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003eCT annotation and segmentation tool\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003eSUV\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003estandardized uptake value\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003eSUVbw\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003ebody-weight\u0026ndash;normalized SUV\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003eUB\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003eupper body\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003eLB\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003elower body\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003eAll SKM\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003eall skeletal muscles\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003eAll SKM [-29,150]\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003eall skeletal muscles within HU -29 to 150\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003eLPECMJR/RPECMJR\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003eleft/right pectoralis major\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003eLPECMNR/RPECMNR\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003eleft/right pectoralis minor\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003eLTEMPORALIS/RTEMPORALIS\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003eleft/right temporalis\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003eLMASSETER/RMASSETER\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003eleft/right masseter\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003eLSCM/RSCM\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003eleft/right sternocleidomastoid\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003eLILIOPSOAS/RILIOPSOAS\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003eleft/right iliopsoas\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003eLUPLGSKM/RUPLGSKM\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003eleft/right upper limb muscles\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003eLLWLGSKM/RLWLGSKM\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003eleft/right lower limb muscles\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003eLASKM/RASKM\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003eleft/right abdominal muscles\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003eAOC\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003eaortic calcification\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003eCAAC\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003ecardiac aggregate calcium\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003eVAT\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003evisceral adipose tissue\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003eVAT [-150, -50]\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003evisceral fat within HU -150 to -50\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003eEPAT\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003eepicardial/pericardial fat\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003ePAAT\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003eperiaortic fat\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003eTHAT\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003ethoracic fat\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003eSAT\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003esubcutaneous fat\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003eSAT [-190, -30]\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003esubcutaneous fat within HU -190 to -30\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003eLASAT/RASAT\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003eleft/right abdominal subcutaneous fat\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003eIMAT\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003eintramuscular adipose tissue\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003eAll IMAT [-190, -30]\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003eIMAT within HU -190 to -30\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003eLUPLGIMAT/RUPLGIMAT\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003eleft/right upper limb intramuscular adipose tissue\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003eLLWLGIMAT/RLWLGIMAT\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003eleft/right lower limb intramuscular adipose tissue\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003eLAIMAT/RAIMAT\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003eleft/right abdominal intramuscular adipose tissue\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003c/div\u003e"},{"header":"Declarations","content":"\u003cp\u003e\u003cstrong\u003eEthics approval and consent to participate:\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003ePatients were enrolled on an ongoing correlative imaging study (ClinicalTrials.gov NCT03422731, City of Hope protocol 17222). The study was approved by the City of Hope Institutional Review Board, and all participants provided written informed consent.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eConsent for publication\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eNo\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eAvailability of data and material\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe data that support the findings of this study are avail\u0026shy;able from the corresponding author upon reasonable request.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eCompeting interests\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eM.F.B. and K.P. are professors at their respective academic institutions and disclose that they are also directors of Voronoi Health Analytics Inc., the developer of the DAFS AI-based automated software platform used in this study. Their contributions to this work were to provide technical\u0026nbsp;guidance\u0026nbsp;on\u0026nbsp;software use, data extraction, and analytic workflow support. They were not involved in the study design, interpretation of results or in decisions regarding manuscript conclusions. Voronoi Health Analytics Inc. as a corporate entity had no role in study design, data collection, manuscript preparation, or publication decisions. The remaining authors declare no competing interests.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eFunding\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThis work has been supported by NIH grants 2R01CA154491 (PI: S.K.H.) and ONCOTEST (Ghent, Belgium, PI: S.K.H.). The content is solely the responsibility of the authors and does not necessarily represent the official views of the National Institutes of Health.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eAuthor\u0026rsquo;s Contributions\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eAuthorship: Conceptualization, M.M., S.K.H.; Methodology, M.M., K.P. M.F.B; Patient Recruitment: M.A.M., J.W., A.S, D.Y.; Formal Analysis, M.M, HG, BC, KP; Investigation: MM, S.K.H.; Manuscript Preparation, M.M.; Manuscript Editing, M.M., S.K.H., M.F.B., GS, JF, KF; Visualization, M.M. and S.K.H.; Supervision, S.K.H; Project Administration, S.K.H.; Funding Acquisition, S.K.H. All authors have read and agreed to the published version of the manuscript.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eAcknowledgements\u0026nbsp;\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe authors would like to acknowledge the study participants and the technologists and staff of the imaging center in radiation oncology department (City of Hope National Medical Center) for their valuable assistance in data acquisition and support throughout this research.\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\u003cli\u003e\u003cspan\u003eSalskov A, Tammisetti VS, Grierson J, Vesselle H, editors. FLT: measuring tumor cell proliferation in vivo with positron emission tomography and 3\u0026prime;-deoxy-3\u0026prime;-[18F] fluorothymidine. Seminars in nuclear medicine. Elsevier; 2007.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eShields AF, Grierson JR, Dohmen BM, Machulla H, Stayanoff JC, Lawhorn-Crews JM et al. Imaging proliferation in vivo with [F-18] FLT and positron emission tomography. 1998;4(11):1334\u0026ndash;6.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eHan EJ, Lee B-h, Kim J-A, Park YH, Choi WHJE. Early assessment of response to induction therapy in acute myeloid leukemia using 18F-FLT PET/CT. 2017;7(1):75.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLiu X, Qu L, Xie Z, Zhao J, Shi Y, Song ZJBEO. Towards more precise automatic analysis: a systematic review of deep learning-based multi-organ segmentation. 2024;23(1):52.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLiu Z, Liu F, Chen W, Liu X, Hou X, Shen J et al. Automatic segmentation of clinical target volumes for post-modified radical mastectomy radiotherapy using convolutional neural networks. 2021;10:581347.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eHuang D, Bai H, Wang L, Hou Y, Li L, Xia Y et al. The application and development of deep learning in radiotherapy: a systematic review. 2021;20:15330338211016386.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eNiu X, Huang Y, Li X, Yan W, Lu X, Jia X et al. Development and validation of a fully automated system using deep learning for opportunistic osteoporosis screening using low-dose computed tomography scans. 2023;13(8):5294.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003ePonnusamy R, Zhang M, Wang Y, Sun X, Chowdhury M, Driban JB et al. Automatic segmentation of bone marrow lesions on MRI using a deep learning method. 2024;11(4):374.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eJimenez-Pastor A, Alberich-Bayarri A, Fos-Guarinos B, Garcia-Castro F, Garcia-Juan D, Glocker B, et al. Automated vertebrae localization and identification by decision forests and image-based refinement on real-world. CT data. 2020;125(1):48\u0026ndash;56.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eTr\u0026auml;g\u0026aring;rdh E, Borrelli P, Kaboteh R, Gillberg T, Ul\u0026eacute;n J, Enqvist O, et al. RECOMIA\u0026mdash;a cloud-based Platf Artif Intell Res nuclear Med Radiol. 2020;7(1):51.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eWasserthal J, Breit H-C, Meyer MT, Pradella M, Hinck D, Sauter AW et al. TotalSegmentator: robust segmentation of 104 anatomic structures in CT images. 2023;5(5):e230024.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eYazdani E, Karamzadeh-Ziarati N, Cheshmi SS, Sadeghi M, Geramifar P, Vosoughi H et al. Automated segmentation of lesions and organs at risk on [68Ga] Ga-PSMA-11 PET/CT images using self-supervised learning with Swin UNETR. 2024;24(1):30.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eMagome T, Froelich J, Holtan SG, Takahashi Y, Verneris MR, Brown K, et al. Whole-Body Distribution of Leukemia and Functional Total Marrow Irradiation Based on FLT-PET and Dual-Energy CT. Mol Imaging. 2017;16:1536012117732203.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eGulec SA, Mesoloras G, Dezarn WA, McNeillie P, Kennedy AS. Safety and efficacy of Y-90 microsphere treatment in patients with primary and metastatic liver cancer: the tumor selectivity of the treatment as a function of tumor to liver flow ratio. J Transl Med. 2007;5:15.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eMcGuire SM, Menda Y, Boles Ponto LL, Gross B, Buatti J, Bayouth JE. 3'-deoxy-3'-[\u0026sup1;⁸F]fluorothymidine PET quantification of bone marrow response to radiation dose. Int J Radiat Oncol Biol Phys. 2011;81(3):888\u0026ndash;93.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eJeraj R, Duan F, Mattison RJ, Romanoff J, Kostakoglu L, Arber DA, et al. Early assessment of treatment response in AML using FLT PET/cT: a trial of the ECOG-ACRIN Cancer Research Group (EAI141). Leuk Lymphoma. 2025;66(14):2765\u0026ndash;73.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eBettinardi V, Castiglioni I, De Bernardi E, Gilardi MJC, Imaging T. PET quantification: strategies for partial volume correction. 2014;2(3):199\u0026ndash;218.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eSaranathan AM, Parente M, editors. Threshold based segmentation method for hyperspectral images. 2013 5Th workshop on hyperspectral image and signal processing: evolution in remote sensing (WHISPERS). IEEE; 2013.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eShi J. Malik JJITopa, intelligence m. Normalized cuts image segmentation. 2000;22(8):888\u0026ndash;905.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eVyavahare AJ, Thool R, editors. Segmentation using region growing algorithm based on CLAHE for medical images. Fourth International Conference on Advances in Recent Technologies in Communication and Computing (ARTCom2012); 2012: IET.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eIsgum I, Staring M, Rutten A, Prokop M, Viergever MA, Van BJItomi G. Multi-atlas-based segmentation with local decision fusion\u0026mdash;application to cardiac and aortic segmentation in CT scans. 2009;28(7):1000\u0026ndash;10.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eEcabert O, Peters J, Schramm H, Lorenz C, von Berg J, Walker MJ, et al. Automatic model-based segmentation heart CT images. 2008;27(9):1189\u0026ndash;201.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eAnyene I, Caan B, Williams GR, Popuri K, Lenchik L, Giri S et al. Body composition from single versus multi-slice abdominal computed tomography: concordance and associations with colorectal cancer survival. 2022;13(6):2974\u0026ndash;84.\u003c/span\u003e\u003c/li\u003e\u003c/ol\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":false,"hideJournal":false,"highlight":"","institution":"","isAcceptedByJournal":false,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":false,"isPdf":false,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"ejnmmi-research","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"ejre","sideBox":"Learn more about [EJNMMI Research](http://ejnmmires.springeropen.com/)","snPcode":"","submissionUrl":"https://www.editorialmanager.com/ejre/default.aspx","title":"EJNMMI Research","twitterHandle":"@officialEANM","acdcEnabled":true,"dfaEnabled":true,"editorialSystem":"em","reportingPortfolio":"BMC/SO AJ","inReviewEnabled":true,"inReviewRevisionsEnabled":true},"keywords":"[18F] 3'-deoxy-3'-fluorothymidine, PET/CT, Multiple Organ Segmentation, Bone Marrow Examination","lastPublishedDoi":"10.21203/rs.3.rs-9077609/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-9077609/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003ch2\u003eBackground\u003c/h2\u003e \u003cp\u003e[18F] 3'-deoxy-3'-fluorothymidine positron emission tomography (FLT-PET) is valuable for detecting acute myeloid leukemia (AML) and monitoring stem cell engraftment after hematopoietic stem cell transplant (HCT) by assessing cellular proliferation in marrow-rich tissues. Reliable marrow quantification is difficult to achieve, and manual segmentation is impractical in clinical workflows. Most automated tools focus on solid tumors and lack clinical validation for skeletal FLT-PET/CT. This study evaluates deep learning whole-body segmentation and cortical\u0026ndash;trabecular marrow quantification on FLT-PET/CT in AML with HCT.\u003c/p\u003e\u003ch2\u003eResults\u003c/h2\u003e \u003cp\u003eTwenty refractory AML patients undergoing total marrow and lymphoid irradiation (TMLI) and transplantation were analyzed. From 134 predefined regions, five representative ROIs (spleen, liver, T6, L1, L3) validated agreement with manual segmentation. Automated and manual count measurements showed strong agreement, with a high correlation (r\u0026thinsp;\u0026gt;\u0026thinsp;0.98, p\u0026thinsp;\u0026lt;\u0026thinsp;0.0001). Consistent hotspot detection by both methods supports the AI tool\u0026rsquo;s accuracy and clinical applicability. Small liver/spleen differences and larger positive vertebral trabecular biases were observed. AI cut processing time by ~\u0026thinsp;95%, markedly improving efficiency.\u003c/p\u003e\u003ch2\u003eConclusion\u003c/h2\u003e \u003cp\u003eThis study provides a technical validation of an AI-driven multi-organ segmentation platform for FLT-PET/CT in AML and HCT, including separate cortical bone and trabecular marrow compartments. The automated approach demonstrated high agreement, excellent reproducibility, and substantial efficiency gains in skeletal marrow and organ quantification. These findings establish a scalable framework for future studies that will correlate FLT-based bone marrow metrics with clinical response and transplant outcomes.\u003c/p\u003e\u003ch2\u003eTrial registration\u003c/h2\u003e \u003cp\u003eClinicalTrials.gov NCT03422731. Registered 6 February 2018, https//www.cancer.gov/research/participate/clinicaltrialssearch/v?id=NCI201701778\u003c/p\u003e","manuscriptTitle":"AI-Driven Bone and Marrow Segmentation on FLT-PET/CT: Technical Multi-organ Validation in AML and HCT","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2026-04-19 12:24:38","doi":"10.21203/rs.3.rs-9077609/v1","editorialEvents":[{"type":"communityComments","content":0},{"type":"reviewerAgreed","content":"","date":"2026-04-11T10:26:29+00:00","index":0,"fulltext":""},{"type":"reviewersInvited","content":"","date":"2026-04-09T19:24:03+00:00","index":"","fulltext":""},{"type":"editorInvited","content":"EJNMMI Research","date":"2026-03-23T11:43:21+00:00","index":"","fulltext":""},{"type":"editorAssigned","content":"","date":"2026-03-19T06:49:09+00:00","index":"","fulltext":""},{"type":"submitted","content":"EJNMMI Research","date":"2026-03-18T13:03:07+00:00","index":"","fulltext":""}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"ejnmmi-research","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"ejre","sideBox":"Learn more about [EJNMMI Research](http://ejnmmires.springeropen.com/)","snPcode":"","submissionUrl":"https://www.editorialmanager.com/ejre/default.aspx","title":"EJNMMI Research","twitterHandle":"@officialEANM","acdcEnabled":true,"dfaEnabled":true,"editorialSystem":"em","reportingPortfolio":"BMC/SO AJ","inReviewEnabled":true,"inReviewRevisionsEnabled":true}}],"origin":"","ownerIdentity":"231e20fe-37f2-40cc-9f77-ec2b04e5d591","owner":[],"postedDate":"April 19th, 2026","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"under-review","subjectAreas":[],"tags":[],"updatedAt":"2026-04-19T12:24:43+00:00","versionOfRecord":[],"versionCreatedAt":"2026-04-19 12:24:38","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-9077609","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-9077609","identity":"rs-9077609","version":["v1"]},"buildId":"XKTyCvWXoU3ODBz1xrDgd","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}