Swin UNETR-based prediction of [68Ga]Ga-FAPI-46 PET/CT dose rate maps in cancer patients: quantitative comparison with [18F]FDG PET/CT Authors: Saeideh Izadi-Yazdi, Farshid Babapour-Mofrad, Elmira Yazdani, Najme Karamzade-Ziarati, Hossein Arabi, Mahdi Sadeghi Published: 2026 ABSTRACT AbstractBackgroundFibroblast activation protein (FAP) is a promising theranostic target due to its high stroma expression in numerous malignancies. This study presents the first DL-based framework for predicting dose rate maps (DRMs) from [68Ga]Ga-FAPI-46 PET/CT, following a quantitative and dosimetric comparison with [18F]FDG PET/CT. The approach supports personalized pre-therapeutic planning in radiopharmaceutical therapies (RPTs).MethodsPET/CT scans with [68Ga]Ga-FAPI-46 and [18F]FDG from 22 cancer patients were retrospectively analyzed. DRMs were generated using the dose voxel kernel (DVK) method using the GATE toolkit. Absorbed doses (ADs) obtained from DVK-based DRMs for tumors and organs at risk (OARs) were compared with those derived from full Monte Carlo (MC), and local energy deposition (LED), based DRMs, and organ-level estimates calculated using OLINDA/EXM v1.1. Tracer uptake of [68Ga]Ga-FAPI-46 and [18F]FDG was compared using SUVmax, SUVmean, and tumor-to-background ratio (TBR). A shifted windows UNEt TRansformers (Swin UNETR) model was trained to predict DRMs and benchmarked against ResNet-32 using R², RMSE, and gamma index.Results[68Ga]Ga-FAPI-46 PET/CT demonstrated higher TBRs and lower OARs uptake and ADs compared to [18F]FDG, indicating its promise in enhancing lesion detectability. The Swin UNETR model achieved an RMSE of 0.0598 Gy2/s, R2 of 0.960, and a gamma pass rate of 98.71%, outperforming ResNet-32.ConclusionCompared to [18F]FDG, [68Ga]Ga-FAPI-46 PET offers higher image contrast, better lesion detectability, and improved dosimetric profiles, supporting personalized RPT planning. While Swin UNETR enables fast and accurate DRM prediction from [68Ga]Ga-FAPI-46, broader validation in larger, multicenter cohorts is needed to ensure reproducibility and clinical impact.Supplementary InformationThe online version contains supplementary material available at 10.1186/s12880-026-02489-8.Keywords: Deep learning, Fibroblast activation protein (FAP), FAPI, Swin UNETR transformer, Dosimetry, Theranostics, Radiopharmaceutical therapy (RPT) Background Advancements in nuclear medicine have enabled the development of fibroblast activation protein inhibitors (FAPIs) [1–4]. Unlike imaging modalities tailored to specific tumor types, such as prostate-specific membrane antigen (PSMA) ligand-based PET, FAPI-PET employs a more universal strategy by targeting activated fibroblasts, specifically cancer-associated fibroblasts (CAFs), within the tumor microenvironment [5, 6]. Fibroblast activation protein (FAP), a membrane-bound type II serine protease, is markedly overexpressed in these activated fibroblasts. By binding selectively to the enzymatic domain of FAP, FAPI-PET enables the visualization of tumor stromal structures as small as 2–3 mm in diameter [7, 8]. Compared to conventional techniques such as [18F]FDG, FAPI PET demonstrates superior image contrast, largely due to the minimal background uptake in normal tissues, attributed to the low FAP expression in quiescent fibroblasts. Gallium-68-labeled FAPI variants, particularly [68Ga]Ga-FAPI-04, [68Ga]Ga-FAPI-46, and [68Ga]Ga-FAPI-2286, have emerged as the most frequently utilized tracers for clinical imaging [2, 3, 9]. The incorporation of FAPI-targeted radioligands into image-guided therapeutic approaches holds promise for novel treatment paradigms across both oncologic and non-oncologic diseases [10, 11]. Comparative analyses between FAPI and conventional [18F]FDG remain essential to assess their respective clinical merits [11–13]. A key advantage of Gallium-68-labeled FAPI tracers lies in their suitability for theranostic applications. By substituting the diagnostic radionuclide Gallium-68 with therapeutic isotopes such as Lutetium-177, Actinium-225, or Yttrium-90, FAPI ligands can be utilized for targeted radionuclide therapy (TRT) or radiopharmaceutical therapy (RPT), an approach commonly referred to as “theranostics” [10, 14, 15]. Accurate internal dosimetry is essential to maximize efficacy and limit toxicity. Conventional MIRD methods estimate average absorbed doses (ADs) in a homogeneous medium, overlooking tissue heterogeneity [16–18]. Direct Monte Carlo (MC) simulations provide accurate dose maps, but are computationally intensive [19]. Pre-treatment imaging allows patient-specific adjustments, and PET-derived metrics have been explored to improve dose prediction by characterizing tracer biodistribution [20]. The integration of artificial intelligence (AI) and PET is enhancing functional imaging, improving precision, and expanding clinical applications. Deep learning (DL) is especially impactful, aiding image acquisition, enhancement, and clinical tasks like staging, dose estimation, and outcome prediction [21–26]. This study aimed to compare [68Ga]FAPI-46 and [18F]FDG in cancer patients through quantitative image analysis and AD calculations for organs at risk (OARs) and lesions, to evaluate the potential advantages of FAPI imaging. For dosimetric comparison, voxel-based MC dosimetry using the dose voxel kernel (DVK) method was benchmarked against full MC simulations, local energy deposition (LED), and organ-based approaches using OLINDA/EXM v1.1. The secondary objective was to reduce the computational cost of full MC simulations by using DL to predict DRMs from [68Ga]FAPI-46 images, trained on DVK-based DRMs. This is the first study to apply shifted windows UNEt TRansformers (Swin UNETR), a transformer-based architecture, for fast and accurate voxel-level FAPI-based DRM prediction to support personalized RPT planning [27]. Methods Study design The overall workflow of the study is illustrated in Fig. 1, outlining each major step in the pipeline from retrospective data collection to model evaluation. The process integrates advanced imaging, segmentation, quantitative analysis, dose estimation, and DL model development for comparative and predictive analysis in theranostic PET/CT imaging. The steps are summarized as follows: Study workflow: an end-to-end pipeline beginning with the acquisition of PET/CT scans and reports. Subsequent steps include segmentation, quantitative analysis, image preprocessing for dosimetry, dose estimation, image preprocessing for modeling, model construction, and model evaluation Data collection: Paired [68Ga]Ga-FAPI-46 and [18F]FDG PET/CT images and reports were acquired for each patient and stored for analysis. Data collection: Paired [68Ga]Ga-FAPI-46 and [18F]FDG PET/CT images and reports were acquired for each patient and stored for analysis. Segmentation: Targets were delineated using the TotalSegmentator tool, followed by manual validation by an expert to ensure anatomical accuracy. Segmentation: Targets were delineated using the TotalSegmentator tool, followed by manual validation by an expert to ensure anatomical accuracy. Quantitative analysis: SUVs were compared between the two tracers for each target to assess relative biodistribution. Quantitative analysis: SUVs were compared between the two tracers for each target to assess relative biodistribution. Dosimetry image preprocessing: CT images were down-sampled, and PET/CT images were used to generate activity/density maps necessary for dosimetry simulations. Dosimetry image preprocessing: CT images were down-sampled, and PET/CT images were used to generate activity/density maps necessary for dosimetry simulations. Dose estimation: AD was estimated using the DVK approach and benchmarked against OLINDA/EXM, LED, and full MC simulations. Dose estimation: AD was estimated using the DVK approach and benchmarked against OLINDA/EXM, LED, and full MC simulations. Modeling image preprocessing: Data were prepared for DL by applying normalization, data augmentation, CT table removal, and image cropping. Modeling image preprocessing: Data were prepared for DL by applying normalization, data augmentation, CT table removal, and image cropping. Deep learning model construction: Predictive models were built using CNN architectures (e.g., ResNet) and transformer-based models (e.g., SWIN UNETR). Deep learning model construction: Predictive models were built using CNN architectures (e.g., ResNet) and transformer-based models (e.g., SWIN UNETR). Deep learning model evaluation: Model performance was assessed using regression metrics and the Gamma index to ensure accurate dose prediction. Deep learning model evaluation: Model performance was assessed using regression metrics and the Gamma index to ensure accurate dose prediction. The following subsections elaborate in detail on each step of this study design, providing methodology, implementation, and rationale for each component in the pipeline. Data collection A retrospective analysis was conducted on 22 patients (15 females, 7 males; age range: 10–74 years; see Table S1 in the Supplementary Tables) who underwent both [18F]FDG and [68Ga]Ga-FAPI-46 between July 2017 and August 2018 for routine clinical evaluation. All patients received a paired [⁶⁸Ga]Ga-FAPI-46 and [¹⁸F]FDG PET/CT scans within a mean time interval of 5 days (min-max: 1–18 days). No anticancer treatment was administered between the two imaging sessions. The patient selection process, including inclusion and exclusion criteria, is summarized in Fig. 2. Flowchart illustrating the patient selection process. Patients were retrospectively screened based on availability of paired [¹⁸F]FDG and [⁶⁸Ga]Ga-FAPI-46 PET/CT scans, imaging time interval, treatment status between scans, and data completeness. A total of 22 patients met all inclusion criteria and were included in the final analysis Mean injected activities were 7.10 mCi for [18F]FDG and 4.62 mCi for [68Ga]Ga-FAPI-46. [68Ga]Ga-API-46 scans were acquired approximately 60 min post-injection (range: 50–70 min) using time-of-flight (TOF) PET/CT. Whole-body scans were obtained with a 3-minute bed position. Low-dose CT (120 kV, 40–120 mA, 512 × 512 matrix size, 3 mm slice thickness) was used for attenuation correction and anatomical localization. Images were reconstructed using ordered subset expectation maximization (OSEM) (4 iterations, 12 subsets) with attenuation, dead time, scatter, and random corrections, plus sharpness recovery and Gaussian filtering (FWHM = 6.4 mm), yielding a 168 × 168 matrix. [68Ga]Ga-FAPI-46 was performed for further evaluation of suspicious regions not visualized on [18F]FDG scans. The cohort included 11 breast cancers, 6 adenocarcinomas, 2 thyroid cancers, 1 sarcoma, 1 liver tumor, and 1 head and neck cancer. The nuclear medicine physician reviewed all scans to verify the type of cancer and assess the image quality. Only patients with complete, paired [18F]FDG and [68Ga]Ga-FAPI-46 datasets were included in the analysis. Segmentation Image segmentation was performed on the CT image of each PET/CT using the TotalSegmentator, which provides high-quality whole-body organ and tissue delineation [28]. Automatic segmentations were reviewed and manually refined by a board-certified nuclear medicine physician (> 8 years’ experience). Segmented structures included the adrenal glands, liver, kidneys, colon, spleen, small intestine, pancreas, lungs, heart, brain, gluteus muscles (muscle reference), gallbladder, urinary bladder, and bone (vertebrae, ribs, and pelvic bones merged as “bone” for red marrow). Figure 3 shows representative segmentation results of OAR for both tracers in coronal, sagittal, and transaxial views. Representative segmentation of organs at risk (OARs) on [18F]FDG (left) and [68Ga]Ga-FAPI (right) PET/CT scans acquired from the same patient Lesion segmentation was performed using the following standardized workflow. First, all lesions were segmented on the [18F]FDG PET/CT scan by an experienced nuclear medicine clinician using the semi-automated PETTumorSegmentation module in 3D Slicer (v5.2.2). Subsequently, the CT component of the [68Ga]Ga-FAPI-46 PET/CT was registered to the corresponding CT of the FDG scan using the Transform and Rigid Registration modules in 3D Slicer (v5.2.2). The resulting transformation was then applied to the FDG-derived lesion masks, which were hardened and overlaid onto the [68Ga]Ga-FAPI-46 PET/CT. Lesion contours were visually inspected and manually modified when necessary to ensure anatomical and metabolic consistency between the two modalities. As illustrated in Fig. 4, some lesions required no modification (e.g., Patient 9, Lesion 1), whereas others required adjustment due to differences in lesion extent or conspicuity between tracers (e.g., Patient 9, Lesion 2; Patient 11 and 4, Lesion 1), which is reflected in the final lesion volumes. Importantly, certain lesions were visible exclusively on the FAPI scan (e.g., a small liver lesion in Patient 11); however, these lesions were excluded from analysis because they could not be reliably matched to a corresponding FDG lesion. Including such lesions would have precluded a valid paired comparison and introduced methodological inconsistency. Representative transaxial views of [68Ga]Ga-FAPI-46 and [18F]FDG PET/CT scans illustrating lesion segmentation in four cases: Patient 9 (Lesion 1), Patient 9 (Lesion 2), Patient 11 (Lesion 1), and Patient 4 (Lesion 1). The colored overlays indicate the segmented lesion volumes used for quantitative analysis on both imaging modalities Regarding the number of lesions analyzed, although the study aimed to include up to three lesions per patient, lesion availability varied across the cohort. Some patients had fewer than three evaluable lesions (e.g., Fig. 4, Patient 9 had two lesions, and Patient 4 had only one lesion that was visible on both scans). Therefore, restricting the analysis to a maximum of three lesions was not intended to preferentially select the largest lesions per patient, but rather to ensure methodological uniformity, reliable co-registration, and paired quantification across both imaging modalities. Image preprocessing and dosimetry Image data were preprocessed using 3D Slicer (version 5.2.2). PET images had a matrix size of 168 × 168 with a voxel size of 4.07 × 4.07 × 3 mm³. CT images had a matrix size of 512 × 512 and a voxel size of 1.36 × 1.36 × 3 mm³, acquired with a tube voltage of 120 kV, current of 55 mAs, and pitch of 1.5. A signal intensity range between − 1200 and 3000 HU was applied. CT images were downsampled to match the PET spatial resolution and matrix size using B-spline interpolation. MC simulations were applied in two distinct ways in this study for [68Ga]Ga-FAPI and [18F]FDG data, as outlined below, with 10⁹ primary events in each, achieving a mean statistical uncertainty of approximately 3%. Full patient-specific MC simulations were conducted for five patients due to the extensive computational time required. Activity maps were first generated by converting SUVs to activity concentration (Bq/mL), accounting for body weight and injected dose. CT images were converted to mass density maps using the Schneider method [16], and together with PET-derived activity maps, served as inputs to the GATE platform (v9.2, Geant4 v11.0) to produce individualized dose rate maps (DRM, Gy/s) [29]. Simulations employed the DoseActor tool, the MersenneTwister random number generator, and included electromagnetic (emstandard_opt3) and radioactive decay physics. Dose estimates from the DVK method were benchmarked against these full MC results. Full patient-specific MC simulations were conducted for five patients due to the extensive computational time required. Activity maps were first generated by converting SUVs to activity concentration (Bq/mL), accounting for body weight and injected dose. CT images were converted to mass density maps using the Schneider method [16], and together with PET-derived activity maps, served as inputs to the GATE platform (v9.2, Geant4 v11.0) to produce individualized dose rate maps (DRM, Gy/s) [29]. Simulations employed the DoseActor tool, the MersenneTwister random number generator, and included electromagnetic (emstandard_opt3) and radioactive decay physics. Dose estimates from the DVK method were benchmarked against these full MC results. The DVK-based DRM, a voxelized MC–based dosimetry approach, was used as the ground truth for DL model development due to its higher computational efficiency compared with full MC simulations. Dose kernels (DKs) were generated in GATE using the same physics settings as for full MC DRM, simulating energy deposition from a 68Ga point source in homogeneous media representing various tissue types, with densities derived from Schneider-converted HU. The voxel S-value (VSV) kernels matched the PET voxel dimensions and quantified the spatial energy deposition per disintegration (Gy/MBq·s) under equilibrium conditions with isotropic radiation transport. DKs were voxel-wise convolved with patient-specific PET activity maps to create patient-specific DRMs simulating spatial dose deposition (as illustrated in Fig. 5). The DVK-based DRM, a voxelized MC–based dosimetry approach, was used as the ground truth for DL model development due to its higher computational efficiency compared with full MC simulations. Dose kernels (DKs) were generated in GATE using the same physics settings as for full MC DRM, simulating energy deposition from a 68Ga point source in homogeneous media representing various tissue types, with densities derived from Schneider-converted HU. The voxel S-value (VSV) kernels matched the PET voxel dimensions and quantified the spatial energy deposition per disintegration (Gy/MBq·s) under equilibrium conditions with isotropic radiation transport. DKs were voxel-wise convolved with patient-specific PET activity maps to create patient-specific DRMs simulating spatial dose deposition (as illustrated in Fig. 5). Dose voxel kernel (DVK)-based dose rate map (DRM) generation workflow. Dose kernels (DKs) were generated in GATE using a 68Ga point source in tissue-equivalent media, with voxel S-value (VSV) kernels matching PET voxel dimensions to quantify spatial energy deposition (Gy/MBq·s). These DKs were voxel-wise convolved with patient-specific PET activity maps to create DRMs ADs from DVK-based DRMs were evaluated in 22 patients and compared to full MC simulations (5 patients), a local energy deposition (LED) method via OpenDose in 3D Slicer v5.2.2, and organ-level calculations using OLINDA/EXM v1.1 [30, 31]. For voxel-based methods (full MC, DVK, and LED), mean ADs were calculated by applying VOIs to voxelized DRMs and multiplying by time-integrated activity coefficients (TIACs). In the organ-based dosimetry using OLINDA/EXM, ADs are calculated by selecting the radioisotope and phantom type (adult male or female) to access pre-calculated S-values, followed by inputting the TIACs to derive the ADs. ADs were reported in Gy and in Gy/GBq (normalized to administered activity). Quantitative analysis For each lesion and OAR within the defined VOIs, SUVmax and SUVmean were computed. Additionally, the tumor-to-background ratio (TBR) was determined by dividing the lesion SUVmax by the background SUVmean, as defined by Eq. 1. To calculate the TBR, the background VOI was also determined manually in a non-cancerous area of normal tissue selected according to the anatomical location of the tumors. Specifically, for lesions in the liver, kidney, or spleen, the background VOI was located in the adjacent normal parenchyma of the same organ, thereby preventing any visible tumor or vascular involvement or areas of heterogeneous uptake of tracer. This organ-matched background selection was chosen to ensure that the TBR reflects the local contrast between tumor uptake and the physiological uptake of the corresponding normal tissue. The same identification strategy was applied consistently for all patients and for both tracers. These metrics enabled a functional comparison between [⁶⁸Ga]Ga-FAPI and [¹⁸F]FDG in terms of lesion detectability and off-target uptake. Image preprocessing and deep learning model construction B-spline interpolated PET/CT images with a voxel size of 2.20 mm3 and a 128 × 128 matrix size were used as input, with both modalities normalized to a [0, 1] scale. The final input size for all models was standardized to a 3D VOI of 96 × 96 × 96 voxels with two channels (PET and CT). A suite of stochastic data augmentations was applied during training using MONAI’s transformation pipeline to mitigate overfitting and improve generalizability. These included probabilistic transformations such as flipping (RandFlipd), 90-degree rotations (RandRotate90d), intensity scaling (RandScaleIntensityd), and shifting (RandShiftIntensityd), with probabilities ranging from 10% to 20%. This study investigated DL–based prediction of voxel-wise DRMs from PET/CT imaging by developing and comparing two model architectures: a transformer-based network and a convolutional neural network (CNN). The primary model was a Swin UNETR, which integrates vision transformers’ representational power with convolutional decoders’ spatial localization capabilities [27]. To provide a performance baseline, a 3D ResNet32 architecture was implemented as the CNN benchmark. Both models were developed using PyTorch (v1.13) and MONAI (v0.9) frameworks and trained on an NVIDIA RTX 3090 GPU equipped with 24 GB of memory. Swin UNETR: transformer-based architecture The main model employed was a Swin UNETR, which integrates a Swin Transformer encoder with a convolutional decoder. The architecture of the Swin UNETR remains structurally unchanged to be able to use the pretrained SSL weights from the PET/CT segmentation task [32], and fine-tuning it on the current Ga-FAPI PET/CT dose prediction task. The encoder uses hierarchical attention mechanisms with shifted windows to effectively capture long-range dependencies in 3D medical images. According to Fig. 6, the input volume is partitioned into non-overlapping patches, and attention is computed locally before propagating information globally. Decoder blocks reconstruct the spatial output while leveraging skip connections from the encoder at multiple resolutions, facilitating high-resolution reconstruction of the output dose distribution. Gradient checkpointing was enabled to optimize memory usage during training, allowing deeper backpropagation through the transformer layers. Key architectural settings for the model included two input channels corresponding to PET and CT images, a single output channel representing the predicted dose, a feature size of 48, and 3D spatial dimensions. A dropout rate of 0.2 was applied during training to reduce the risk of overfitting by randomly deactivating a subset of neurons, thereby encouraging the model to learn more robust and generalized features. Gradient checkpointing was also enabled to reduce memory consumption and support deeper model training. The encoder was initialized with self-supervised weights pre-trained on a [68Ga]Ga-PSMA-11 PET/CT segmentation task [32]. This pretraining involved tasks such as rotation prediction and volume inpainting, enabling the encoder to learn robust anatomical features without requiring extensive manual annotations. The decoder was trained from scratch and fine-tuned along with the encoder on 22 labeled PET/CT volumes. Training was performed for 1000 epochs using stochastic gradient descent (SGD) with Nesterov momentum (0.99), an initial learning rate of 1e − 4, and weight decay regularization. A warm-up cosine annealing schedule adjusted the learning rate over time. Additionally, an early stopping strategy with a patience of 50 epochs was implemented based on validation loss to prevent overfitting and ensure model convergence. The loss function was mean squared error (MSE), chosen for its suitability in voxel-wise regression tasks. A sliding window inference strategy was applied with a 50% overlap for patch-wise prediction. Schematic illustration of the proposed Swin UNETR framework for voxel-level dose rate map (DRM) estimation using FAPI PET/CT data. The dual-modality FAPI PET/CT input volume is divided into non-overlapping three-dimensional patches and processed by a multi-stage Swin Transformer encoder with hierarchical feature extraction and patch merging. Spatial resolution is progressively recovered in the decoder via deconvolution layers and encoder–decoder skip connections. Multi-scale feature refinement is achieved using residual blocks, while the final three-dimensional DRM is generated by a CNN-based output head. The Swin Transformer encoder employs both window-based multi-head self-attention (W-MSA) and shifted-window MSA (SW-MSA) mechanisms to efficiently capture local and contextual information Given the limited size of the labeled dataset and the architectural complexity of the Swin UNETR model, overfitting posed a significant challenge in this study. To address this, this study adopted a self-supervised learning (SSL) strategy previously introduced for PET/CT segmentation, in which a Swin UNETR encoder was pretrained on 652 unlabeled [68Ga]Ga-PSMA-11 PET/CT scans using three proxy tasks: image rotation prediction, volume inpainting, and contrastive learning [32]. This pretraining enabled the model to capture meaningful anatomical and functional representations from PET/CT data. Only the encoder was trained during the SSL phase, after which it was coupled with a decoder and fine-tuned on task-specific labeled data. By leveraging this large pool of unlabeled data, the model demonstrated improved stability and generalization during fine-tuning, thereby reducing the risk of overfitting and enhancing dose prediction performance. ResNet32: convolutional baseline To benchmark the performance of the transformer-based model, a 3D ResNet32 architecture was implemented as a CNN baseline. The architecture follows a standard residual learning framework adapted for volumetric data and includes multiple convolutional blocks with skip connections. Identical preprocessing steps and training configurations (epochs, learning rate, ROI size, and loss function) were used for the ResNet32 model to ensure a fair comparison. The total number of trainable parameters was logged for reference, and training was also performed using MSE loss. This model served as a reference point to assess the added value of transformer-based attention mechanisms for dose estimation tasks. Deep learning model evaluation Both Swin UNETR and ResNet32 models were evaluated under a consistent stratified 5-fold cross-validation (CV) protocol. Model performance was tracked throughout 1000 training epochs. The evaluation employed a range of voxel-level statistical metrics, including the coefficient of determination (R2), relative error (RE), mean absolute error (MAE), relative absolute error (RAE), root mean square error (RMSE), and peak signal-to-noise ratio (PSNR). Additionally, spatial agreement between the DL-predicted and reference dose maps was examined using Gamma analysis, applying a distance-to-agreement (DTA) of 2.2 mm and a dose difference (DD) threshold of 1%. Gamma pass rates and corresponding maps were computed to quantify the overlap between predicted and reference dose distributions. The pass rates were calculated using Eq. 2: As an additional region-of-interest–level evaluation, Bland–Altman (BA) analysis was performed on AD estimates for organs and lesions from 22 patients. The analysis was conducted for the best-performing model by comparing model-predicted doses with those obtained using the DVK reference method. BA plots were generated to assess mean bias and limits of agreement between the two approaches. Results Quantitative PET/CT analysis According to Table S2 in the Supplementary Tables, the mean ± standard deviation (SD) of SUVmean and SUVmax values were compared across various organs and tissues in 22 cancer patients who underwent both [68Ga]Ga-FAPI and [18F]FDG imaging. Overall, [68Ga]Ga-FAPI demonstrated lower background uptake in most regions, especially in the brain, compared to [18F]FDG. For example, the SUVmean in the brain was substantially lower with [68Ga]Ga-FAPI (0.16) versus [18F]FDG (7.63). In contrast, organs such as the kidneys and urinary bladder exhibited notably higher SUVmax values with [68Ga]Ga-FAPI, likely due to renal clearance. These differences in physiological distribution underscore the potential clinical advantages of [68Ga]Ga-FAPI in detecting certain tumor types or anatomical regions. As reported in Table S3 in the Supplementary Tables, [68Ga]Ga-FAPI demonstrated higher uptake and superior lesion contrast compared to [18F]FDG. For Lesion 1, [68Ga]Ga-FAPI showed increased SUVmean (3.33 vs. 2.60) and SUVmax (10.18 vs. 6.39). Similar trends were observed in Lesion 2 and Lesion 3, with [68Ga]Ga-FAPI consistently providing higher SUV values. Notably, Lesion 3 exhibited the highest SUVmax and SUVmean values for both tracers, with [68Ga]Ga-FAPI reaching 49.36 and 11.29, respectively, compared to 20.68 and 6.17 with [18F]FDG. These findings highlight the enhanced tumor detectability and contrast achievable with [68Ga]Ga-FAPI, particularly in smaller lesions. Moreover, Table S4 (Supplementary Tables) presents site-specific TBRs for [68Ga]Ga-FAPI-46 and [18F]FDG PET/CT and paired differences (ΔTBR) within the same lesions. Across all metastatic sites, including liver, bone, lymph nodes, breast, and head and neck, FAPI consistently showed higher TBR than FDG, with ΔTBR ranging from 3.1 ± 1.6 (liver) to 4.3 ± 2.8 (breast), demonstrating improved lesion contrast while accounting for organ- and size-specific variability. Radiation dosimetry As shown in Table S5 in the Supplementary Tables, mean ADs (mGy/MBq) in organs and tumors were compared across methods: DVK (22 and 5 patients), full MC (5 patients), LED, and OLINDA/EXM, for [68Ga]Ga-FAPI-46 and [18F]FDG. Although full patient-specific MC simulations are considered the reference standard for voxel-level dosimetry, their substantial computational cost limited their feasibility to five patients in this study and restricted their role to validation only. Across both tracers, DVK-based estimates in these five patients closely match full MC results, generally differing only in the third decimal place, which indicates strong agreement and robustness of the DVK approach. Tumor dose estimates derived from DVK differed from full MC by only a few percent, underscoring DVK’s reliability in capturing lesion-level dose distributions. This close concordance justified the selection of DVK as a consistent and computationally efficient surrogate ground truth for DL model training, enabling scalable generation of dose maps across the full cohort of 22 patients while preserving the essential physics of radiation transport. LED consistently yielded slightly lower ADs than both DVK and MC, whereas OLINDA/EXM showed larger deviations, often under- or overestimating organ doses depending on the tissue, reflecting the limitations of phantom-based organ-level dosimetry. Overall, these findings demonstrate that DVK provides accurate and patient-specific tumor dose estimates without the prohibitive computational burden of full MC, making it particularly well-suited for clinical and research applications in RPT planning. Table S6 in the Supplementary Tables further quantifies these differences by reporting the relative percentage error (RPE) between methods. For [68Ga]Ga-FAPI-46 PET/CT, DVK showed excellent agreement with full MC across most organs and tumors, with RPEs typically below 5% in multiple organs. A slightly higher but still moderate discrepancy was observed for FAPI in the pancreas. For [18F]FDG PET/CT, the RPEs between DVK and MC were generally larger than for FAPI, with several organs exceeding 4% (e.g., kidneys: 4.76%, pancreas: 4.90%, brain: 4.41%), though most remained below 5%. Between DVK and LED, moderate discrepancies were observed (RPE range: 1–38%), particularly in the brain (20.70%), muscle (21.40%), heart (20.80%), lungs (25.32%), and kidneys (37.77%) for [¹⁸F]FDG, likely reflecting the limitations of the LED assumption. The largest deviations occurred when comparing DVK with OLINDA/EXM, where RPEs frequently exceeded 40–50% in several organs (e.g., liver: 48.00% for FAPI, 56.50% for FDG; tumor: 44.40% for FAPI, 56.90% for FDG), highlighting the impact of stylized phantoms and simplified geometry on organ-level dosimetry. Deep learning models Two DL models, Swin Transformer and ResNet-32, were trained for personalized voxel-level dose estimation. Table 1 presents voxel-wise mean quantitative metrics for the DL-based dosimetry methods, using MC results as the reference for Ga-FAPI images, respectively. In both cases, the SWIN UNETR + SSL model achieved the lowest error metrics (R2, RE, MAE, RAE, RMSE, and PSNR) and the highest similarity metrics (PSNR and gamma pass rate), outperforming ResNet32. Figure 7 shows averaged RMSE vs. epochs for ResNet32 (blue) and Swin UNETR (green), averaged over 20-epoch bins; Swin UNETR outperforms ResNet32 with faster reduction and stable values around 0.05–0.06 versus 0.07–0.10. Table 1Voxel-wise quantitative metrics comparing three DL-based dosimetry approaches against MC reference calculations for [68Ga]Ga-FAPI PET/CT imagesModelsR2 ScoreRE (%)MAE (Gy/s)RAE (%)RMSE (Gy2/s)PSNR (Gy/s)Gamma pass rateResNet320.830.3630.02390.4580.079919.6990.12%SWIN UNETR + SSL0.9600.1870.01010.2810.059850.47498.71% Voxel-wise quantitative metrics comparing three DL-based dosimetry approaches against MC reference calculations for [68Ga]Ga-FAPI PET/CT images Comparison of validation RMSE between ResNet32 and Swin UNETR over validation epochs (grouped every 20 epochs) Figure 8 compares Ga-FAPI PET/CT, reference DVK_DRM, DL-predicted DRM, and Gamma maps; the DL method achieved a 98% pass rate, showing strong spatial and dosimetric agreement. Two views of the [68Ga]Ga-FAPI PET/CT image, ground truth DRM (obtained via DVK approach), SWIN UNETR-based DRM, and Gamma index maps for a case with pass rates of 98%. Voxel values ≤ 1 were considered a pass for this evaluation Figure 9 presents BA analyses comparing AD estimates obtained with the SWIN-UNETR model and the DVK reference method for organs and tumors. Bland–Altman plots comparing absorbed dose estimates obtained with SWIN-UNETR and DVK for (a) organs and (b) tumors. The solid markers represent individual cases, the red dashed line indicates the mean bias, and the green dashed lines denote the upper and lower limits of agreement (± 1.96 SD). The organ analysis demonstrates minimal bias and tight agreement, whereas the tumor analysis shows a positive bias with increased variability across the dose range For organs (Fig. 9A), the mean dose differences are centered close to zero, indicating minimal systematic bias between the two methods. Most data points fall within the limits of agreement, with relatively narrow dispersion across the range of mean doses. This suggests good agreement and consistency of SWIN-UNETR–derived organ dose estimates compared to DVK, with no strong evidence of proportional bias. For tumors (Fig. 9B), the BA analysis shows a positive mean bias, indicating that SWIN-UNETR generally estimates higher absorbed doses than DVK. The limits of agreement are wider compared to the organ analysis, reflecting increased variability in tumor dose estimation. Nevertheless, the majority of data points remain within the limits of agreement, demonstrating acceptable concordance between the two methods across the evaluated tumor dose range. Overall, the BA results indicate strong agreement for organ dosimetry and moderate but systematic differences for tumor dosimetry, with SWIN-UNETR tending to overestimate tumor doses relative to DVK. Discussion FAPI theranostics represent a promising advancement in targeted cancer treatment, offering both diagnostic imaging and therapeutic potential [10]. Timely and precise estimation of ADs in tumors and organs is essential for optimizing therapeutic efficacy while minimizing toxicity. Such predictive information supports individualized treatment planning by facilitating the adjustment of RPT doses prior to treatment, aiming to increase tumor irradiation efficacy while minimizing adverse effects on healthy tissues [33, 34]. To the best of our knowledge, this is the first study to employ advanced DL models for individualized voxel-level dosimetry prediction using DRMs derived from [68Ga]Ga-FAPI-46 PET/CT imaging in cancer patients. This study first developed highly accurate voxel-based DRMs through MC simulations using the DVK method, specifically tailored for [68Ga]Ga-FAPI-46. These DRMs were then used to train a state-of-the-art transformer-based DL model, Swin UNETR, for rapid and reliable prediction of voxel-wise dosimetry. An additional objective of this study was to compare [68Ga]Ga-FAPI-46 with the commonly used radiopharmaceutical [18F]FDG, with the aim of elucidating differences in quantitative parameters and AD distribution within the body. [68Ga]Ga-FAPI-46 targets FAP, which is overexpressed in cancer-associated fibroblasts, enabling the detection of primary tumors and metastases that may not be visualized using FDG. FAPI-46 demonstrates tumor uptake levels that are three to six times higher than those of FDG. Furthermore, while FDG accumulates in a wide range of normal tissues, FAPI uptake is more tumor-specific, allowing for improved delineation of malignant lesions [35]. This selective uptake not only enhances image contrast but also contributes to reduced radiation burden in OARs. In addition to its tumor-selective biodistribution, FAPI-46 offers practical advantages over FDG, including shorter uptake time and no need for fasting or blood glucose monitoring, simplifying patient preparation [36]. These distinctions underline the diagnostic and dosimetric benefits of FAPI-based imaging, as supported by the results of Table S6 in the Supplementary Tables. Elboga et al. [36] compared [68Ga]Ga-FAPI and [18F]FDG for lesion detection in multiple myeloma. Their results demonstrated that [68Ga]Ga-FAPI identified bone lesions in two patients that were missed by [18F]FDG, and it showed higher SUVmax values in the lungs and presacral lymph nodes. The authors highlighted that [68Ga]Ga-FAPI exhibited greater tumor uptake than [18F]FDG. This study demonstrates that the DVK method achieves high accuracy in AD estimation, with tumor and organ doses closely matching those obtained from full MC simulations. The excellent agreement between DVK and MC, with RPEs typically below 5%, highlights DVK as a robust surrogate that balances computational efficiency with precision. Importantly, tumor ADs estimated by DVK were nearly identical to MC results, underscoring its reliability in lesion-level dosimetry, a key requirement for individualized treatment planning in RPTs. Tumors received the highest ADs, with values of 0.25 mGy/MBq for [68Ga]Ga-FAPI-46 and 0.43 mGy/MBq for [18F]FDG (Table S5 in the Supplementary Tables), exceeding normal tissue ADs. The urinary bladder showed the highest ADs for both tracers, followed by the kidneys, reflecting their clearance pathways. In contrast, the brain exhibited the lowest ADs, except for the markedly elevated brain dose for [18F]FDG due to its physiological glucose metabolism. These results confirm DVK’s accurate capture of tracer-specific biodistribution across a 22-patient cohort. When comparing the AD results (Table S5 in the Supplementary Tables) with the previously published OLINDA/EXM results by Meyer et al. [37] for [68Ga]Ga-FAPI-46, both studies identified the urinary bladder wall and kidneys as the organs receiving the highest ADs, confirming the urinary clearance pathway as dose-limiting. The agreement between these two OLINDA-based estimates supports their consistency across studies. In contrast, DVK (22 patients) gave a bladder dose slightly higher than OLINDA, and a kidney dose slightly lower. For most parenchymal organs, DVK estimates in 22 patients were consistently lower than OLINDA. The downward shift in DVK estimates reflects voxel-based, patient-specific dosimetry, capturing individual organ geometry and tracer distribution more realistically than phantom-based OLINDA. DVK reported tumor ADs of 0.252 mGy/MBq in 22 patients, versus 0.364 mGy/MBq from OLINDA (Table S5 in the Supplementary Tables). Phantom-based methods, such as Meyer et al. [37] lack tumor dosimetry, highlighting DVK’s advantage in accurate, patient-specific dose estimates and showing OLINDA’s tendency to overestimate. MC simulations are highly accurate for RPT dose estimation, but their intensive computational demands limit clinical use [38, 39]. Various approaches have been proposed to balance accuracy and computational efficiency, including machine learning (ML) algorithms and deterministic models like the dose point kernel and VSV techniques. More recently, DL has shown promise in predicting patient-specific dose distributions from pretherapy PET/CT scans using tracers such as Ga-68 or F-18 [40, 41]. DL-based approaches have been extensively explored in studies using [68Ga]Ga-PSMA and [68Ga]Ga-DOTATATE tracers [33, 42, 43]. However, their application to [68Ga]Ga-FAPI has been largely unexplored. This study addresses this gap by implementing a DL-based model for dose prediction using [68Ga]Ga-FAPI. PET-based MC simulations can generate realistic dose maps for training DL models when SPECT data are unavailable. Yazdani et al. [44, 45] demonstrated that pretherapy [68Ga]Ga-PSMA PET/CT–based simulations capture biologically meaningful spatial heterogeneity in ADs for kidneys and tumors, supporting their use in predictive modeling. With the expanding use of RPTs, incorporating [68Ga]Ga-FAPI-46 into early dose estimation strategies may help streamline clinical workflows and advance personalized treatment planning [33, 43, 46–48]. As a pioneer study, Lee et al. [49] proposed a deep CNN model for voxel-level dose prediction in personalized dosimetry, offering a fast and accurate alternative to direct MC simulations. Using PET/CT image patches as input, their CNN approach achieved significantly lower voxel dose rate errors (2.54% ± 2.09%) compared to the VSV method (9.97% ± 1.79%) and OLINDA/EXM (34.22%), and outperformed both methods in whole-body organ dose estimation. Götz et al. [50] introduced a novel patient-specific dosimetry method for [¹⁷⁷Lu]-based RPTs, utilizing a neural network trained on over 52,000 MC-derived DVKs to predict organ-specific dose distributions. Focusing on the kidneys as critical organs, the method demonstrated improved accuracy over standard homogeneous kernel convolution approaches, with no added computational burden. Akhavanallaf et al. [42] introduced a novel DL-based framework for whole-body, patient-specific dosimetry that accounts for heterogeneous activity distribution, anatomical variation, and tissue composition. In their study, a deep neural network (DNN) was trained to predict patient-specific voxel-level S-value kernels using density maps from CT images, with MC simulations providing the ground truth reference. These predicted kernels were convolved with activity distributions to generate 3D dose maps following the voxel-based MIRD formalism. The DNN-predicted dose maps demonstrated strong concordance with MC-derived results, achieving a mean relative absolute error (MRAE) of 2.6%, which outperformed both the single S-value (SSV) and multiple S-value (MSV) MIRD methods, as well as the OLINDA/EXM software. Li et al. [51] developed a residual deep CNN (DblurDoseNet) for patient-specific voxel-level dosimetry, addressing the limitations of MC simulations and DVK convolution under SPECT resolution constraints. Trained on MC-generated dose-rate maps from virtual phantoms, the CNN learned the residual between DVK and MC outputs and demonstrated significantly improved accuracy, resolution, and noise performance. Specifically, the CNN reduced mean dose-rate errors by up to 66% and NRMSE by up to 18% compared to DVK and MC in both lesion and kidney regions, and generated full patient dose-rate maps in approximately 30 s. Dewaraja et al. [52] developed and validated an automated voxel-level dosimetry pipeline for patient-specific [177Lu]-DOTATATE therapy, incorporating CNNs for organ segmentation, contour intensity-based image registration, MC dose-rate estimation, and voxel-wise curve fitting. Applied to 20 patients with neuroendocrine tumors, the workflow demonstrated robust dose fitting and considerable inter-patient dose variability across organs and lesions. Mansouri et al. [53] developed a hybrid transformer-based DL model that integrates an MSV approach for fast and accurate voxel-level dosimetry in [¹⁷⁷Lu]Lu-DOTATATE therapy. Their model, trained to predict the difference between MC and MSV dose maps, achieved superior accuracy (voxel relative absolute error of 5.28% ± 1.32) compared to MSV and other methods, while drastically reducing computation time from days to seconds. The approach demonstrated high gamma pass rates (~ 99%) and organ-level dose estimation accuracy close to MC simulations. When contextualized against previously published DL-based dosimetry studies, the performance of the proposed Swin UNETR model is comparable to, and in some cases exceeds, that reported for CNN-based approaches using other radiotracers. Prior studies in [¹⁷⁷Lu]Lu-DOTATATE or [¹⁷⁷Lu]Lu-PSMA therapies have typically reported voxel-level relative errors or RMSE values in the range of 5–10%, with gamma pass rates between 95% and 99% depending on dose thresholds and evaluation criteria [42, 45, 49, 53]. In the present study, Swin UNETR achieved a mean voxel-level RMSE corresponding to a dose prediction error below 6% and a gamma pass rate of 98.71%, which falls well within, and in some cases improves upon, the accuracy reported in these earlier works. Notably, many prior studies relied on CNN-based architectures trained from scratch or on relatively shallow networks, whereas the present work leverages a transformer-based architecture with hierarchical self-attention and extensive self-supervised pretraining on a large unlabeled PET/CT dataset. This design appears particularly advantageous for capturing the spatial heterogeneity and long-range contextual dependencies inherent to voxel-level dose distributions, as reflected by the smoother convergence behavior and consistently lower prediction errors observed for Swin UNETR compared with a ResNet-based baseline. Furthermore, direct comparison across studies must consider differences in radionuclide physics and tracer biodistribution. [⁶⁸Ga]Ga-FAPI-46 exhibits distinct pharmacokinetics and tumor-to-background characteristics, posing different challenges for dose prediction. Despite these differences, the comparable accuracy achieved in this work suggests that transformer-based DL models can generalize effectively across tracer types and imaging contexts, supporting their broader applicability in personalized dosimetry. This study demonstrates that Swin UNETR outperformed the ResNet32-based model across voxel-wise performance metrics, including dosimetric accuracy (MAE, RAE, RMSE, Table 1) and image fidelity (PSNR, gamma pass rate). These findings suggest that transformer-based architectures, particularly those leveraging SSL and hybrid attention mechanisms, are better suited for capturing the spatial complexity and heterogeneity present in dose distributions. Swin UNETR showed lower prediction errors and smoother RMSE convergence (Fig. 7), indicating better generalization. The RMSE corresponds to a mean dose prediction error of less than 6%, which is well within the tolerances typically considered acceptable in clinical practice for personalized RPTs. Clinical studies have suggested that dosimetric errors of up to 10% are generally tolerable in treatment planning [54], but the lower errors observed here (in the range of ~ 6%) provide confidence in the reliability of the model for clinical use. Additionally, the gamma pass rate of 98.71% (Table 1) indicates excellent spatial agreement between the predicted dose maps and the reference dose maps derived from the DVK approach. In clinical dosimetry, a gamma pass rate of 95% or higher is often regarded as clinically acceptable [45, 53]. Given that the Swin UNETR model achieved a pass rate exceeding 98%, this suggests that the model’s dose predictions are highly reliable and likely to be clinically actionable for treatment planning. This level of accuracy is crucial for measuring radiation toxicity to OARs and optimizing the tumor dose distribution in personalized RPT protocols. It is important to note that the DVK method, while highly accurate, has inherent uncertainties due to its simplified assumptions about tissue heterogeneity and its reliance on pre-determined dose kernels. These approximations could introduce minor errors in the ground truth dose maps that were used for training and evaluating the model. As such, the reported RMSE and gamma pass rate may also reflect the limitations of the DVK method itself. Despite this, the Swin UNETR model demonstrated remarkable robustness, with performance metrics suggesting that the model can still achieve highly reliable dosimetry even when considering the uncertainties in the DVK-based reference. This study focused on [68Ga]Ga-FAPI-46; however, emerging tracers such as [68Ga]Ga-FAP-2286 have shown promising tumor uptake and retention across various cancers, together with favorable safety profiles [55, 56]. Future research should utilize DL-based voxel-level dosimetry models to further characterize the dosimetric properties of [68Ga]Ga-FAP-2286. From a clinical workflow perspective, the proposed framework is well aligned with current [177Lu]Lu-FAPI RPT practice, as PET/CT imaging is already routinely acquired during the pre-therapeutic evaluation phase to assess tracer biodistribution and target expression. By enabling rapid, voxel-level prediction of DRMs directly from PET/CT images, the DL-based approach substantially reduces the computational burden and turnaround time associated with conventional MC–based dosimetry. This near–real-time dosimetric feedback could facilitate patient-specific treatment planning, including individualized activity prescription, evaluation of tumor-to-organ dose trade-offs, and early identification of dose-limiting OARs. Moreover, such rapid dose estimation may support adaptive treatment strategies across therapy cycles, allowing dose modulation based on predicted absorbed dose distributions, while seamlessly integrating into existing clinical timelines without adding operational complexity. The B-A analysis is an invaluable approach to quantify and better explain the differences between the measured and predicted methods, which provides a statistical and visual perspective into their agreement. Based on the B-A plot in Fig. 9A for organs, an important observation is the minimal systematic bias between ADRM estimates obtained through the SWIN-UNETR model and the DVK reference method. However, for lesions, the positive mean bias indicates that predicting lesion ADs remains more challenging compared to organs. The higher variability might be due to various reasons such as lesion heterogeneity, smaller size, partial volume effects, segmentation and registration errors, suggesting that while the proposed models are relatively accurate, there is still room for improving consistency and precision. In this regard, increasing the data size in future studies not only improves the accuracy and reliability of the model but also makes it possible to evaluate BA analysis based on lesion size and locations. While the results of this study demonstrate the potential of Swin UNETR for rapid and accurate voxel-level dose prediction, the limited cohort size of 22 patients represents an important limitation, particularly for the DL task. Accordingly, this study should be interpreted as an initial feasibility and proof-of-concept investigation rather than a definitive clinical validation. Future direction in larger, multicenter cohorts will be crucial to reduce the risk of overfitting, improve statistical robustness, and confirm the generalizability and reproducibility of the model across diverse patient populations. Although the current DL model was trained to reproduce DVK-based DRMs, the observed small discrepancies between DVK and full MC simulations suggest that DVK-based dose distributions may serve as a practical surrogate for MC-based dosimetry. Nevertheless, direct training and validation against larger MC-derived datasets will be required to fully establish this capability and to quantify residual uncertainties. A potential limitation of this study is the use of transfer learning from PSMA PET/CT data, which may introduce tracer-specific bias due to differences in physiological uptake and tumor expression patterns between PSMA and FAPI. Nevertheless, PET imaging data share common low- and mid-level features, such as noise characteristics, spatial resolution, and anatomical correspondence with CT, particularly for Ga-labeled tracers, making transfer learning a pragmatic approach in the setting of limited Lu-FAPI datasets. Despite these shared characteristics, residual bias cannot be excluded and may affect generalization in organs or lesions with tracer-specific uptake. Future work will address this limitation through the inclusion of larger, FAPI-only multicenter cohorts, tracer-aware fine-tuning, and domain adaptation strategies to further reduce bias and improve robustness. Moreover, the lesion segmentation strategy for quantitative analysis between [18F]FDG PET/CT and [68Ga]Ga-FAPI-46 PET/CT may introduce selection bias. The lesion segmentation was performed initially on [18F]FDG PET/CT; therefore, FAPI-only lesions (visible only on FAPI) were excluded. This approach, although introduced by an expert nuclear medicine clinician, may affect SUV and TBR quantitative parameters, which can be considered another limitation. The future work needs to address other strategies and compare them on a larger cohort to reach a stable segmentation process for quantitative comparative analysis. To enhance clinical interpretability and adoption, the development of explainable AI (XAI) approaches will be essential to address the “black-box” nature of current models, increasing transparency and clinician trust [57]. Additionally, federated learning strategies could enable collaborative model training across multiple institutions while preserving patient privacy, facilitating access to larger and more heterogeneous datasets, and thereby improving model robustness and broader applicability [58]. Moreover, ongoing advances in quantitative PET/MR imaging, particularly in attenuation correction and simultaneous multiparametric acquisition, may further enhance patient-specific dosimetry [59]. As these methods mature, PET/MR-derived data could be increasingly leveraged by DL models to improve voxel-level dose prediction and personalized radiopharmaceutical therapy planning. Collectively, these directions will support the evolution of Swin UNETR–based dose prediction from a promising proof-of-concept to a clinically relevant tool for patient-specific radiopharmaceutical therapy planning. Conclusion This study demonstrates that [68Ga]Ga-FAPI-46 PET/CT offers superior tumor specificity, higher tumor-to-background ratios, and lower ADs to critical organs compared to conventional [18F]FDG. By leveraging MC–based DRMs, a Swin UNETR transformer architecture was developed to accurately and rapidly predict voxel-level DRMs, significantly outperforming CNN. These results highlight the potential of DL-enhanced FAPI imaging for personalized treatment planning, with further validation in larger, multicenter cohorts needed to confirm clinical utility. Supplementary Information Below is the link to the electronic supplementary material. Supplementary Material 1 Acknowledgements We sincerely thank Habibollah Dadgar for his efforts in providing the data from the Nuclear Medicine and Molecular Imaging Department, RAZAVI Hospital. Abbreviations Absorbed Dose Artificial Intelligence Bland–Altman Batch Normalize Cancer-Associated Fibroblast Convolutional Neural Network Cross-Validation Distance to Agreement Dose Difference Dose Kernel Fine-Tuning Full Width Half Maximum Fluorodeoxyglucose Fibroblast Activation Protein Inhibitors Geant4 Application for Tomographic Emission Hounsfield Unit International Commission on Radiological Protection Leaky Rectified LINEAR Unit Local Energy Deposition Monte Carlo Medical Internal Radiation Dose Mean Absolute Error Organ At Risk Ordered Subset Expectation Maximization Peak Signal-to-Noise Ratio Relative Error Relative Absolute Error Root Mean Square Error Residual Network Rectified Linear Unit Regions of Interest Self-Supervised Learning Shifted Windows UNEt Transformer Standard Deviation Stochastic Gradient Descent Standardized uptake values Tumor-to-Background Ratios Tumor Microenvironment Time-Integrated Activity Time of Flight Author contributions All authors contributed to the study conception and design. SIY, EY, HA, and NKZ performed material preparation, data collection, and analysis. FBM, MS, and HA supervised the project and provided overall methodological guidance. EY and SIY led the development, training, and evaluation of the deep learning models, including the Swin-UNETR architecture, while SIY and EY contributed to the Monte Carlo simulations, dose kernel generation, and voxel-level dosimetry analysis. NKZ performed clinical image interpretation, patient selection, and annotation. The first draft of the manuscript was written by SIY and EY, and all authors read, revised, and approved the final manuscript. Funding The authors declare that no funds, grants, or other support were received during the preparation of this manuscript. Data availability The datasets generated and/or analyzed during the current study are not publicly available due to institutional privacy regulations and ethical restrictions on sharing patient medical imaging and clinical data, but are available from the corresponding author on reasonable request. Declarations Ethics approval and consent to participate The study protocol was approved by the Ethics Committee of Tehran University of Medical Sciences (TUMS) [IR.TUMS.MEDICINE.REC.1400.348]. Due to the retrospective nature of the study, the requirement for informed consent to participate was waived by the institutional review board. All procedures were conducted in accordance with the Declaration of Helsinki. Consent for publication Not applicable. Competing interests The authors declare no competing interests. Footnotes Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. Contributor Information Farshid Babapour-Mofrad, Email: babapour@iau.ac.ir. Mahdi Sadeghi, Email: sadeghi.m@iums.ac.ir. References Associated Data Supplementary Materials Supplementary Material 1 Data Availability Statement The datasets generated and/or analyzed during the current study are not publicly available due to institutional privacy regulations and ethical restrictions on sharing patient medical imaging and clinical data, but are available from the corresponding author on reasonable request.