Charge Au@Pt NPs combined with 3D STS-Net for adaptive and sensitized radiotherapy of hepatocellular carcinoma: Synergistic enhancement of therapeutic gain across physical and biological dimensions.

1
Introduction
Radiotherapy (RT), one of the three pillars of cancer treatment, derives its clinical value from its ability to precisely eradicate tumor cells while maximally sparing normal tissues [1]. The therapeutic gain ratio (TGR), a key metric for evaluating RT quality, reflects the capacity to maximize tumor control probability (TCP) while minimizing normal tissue complication probability (NTCP) [2]. A higher TGR indicates better therapeutic efficacy [3]. Current approaches to improving TGR fall into two major categories. The first is the physical dimension, represented by adaptive radiotherapy (ART), which optimizes dose distribution dynamically through real-time monitoring of anatomical changes to enhance irradiation accuracy [4]. The second is the biological dimension, involving radiosensitization strategies that use drugs or nanomaterials to increase tumor cell responsiveness to ionizing radiation, thereby improving RT effectiveness [[5], [6], [7], [8]].
Hepatocellular carcinoma (HCC), accounting for 85–90% of primary liver cancers, is among the most common malignancies [9]. Although advances in image-guided and precision RT have made it an important treatment option for unresectable or recurrent HCC, its TGR remains relatively low among solid tumors [10]. On the physical side, a major challenge of ART for HCC lies in tumor auto-contouring. The inherently low contrast between tumor and liver parenchyma on non-contrast CT requires multi-modal image fusion for accurate delineation, which is complex and difficult to meet the real-time requirements of online ART. While deep learning methods are widely used for automatic segmentation of organs at risk (OARs), their dice coefficient scores (DCS) for HCC lesion segmentation remain suboptimal and insufficient for clinical ART applications [11,12]. On the biological side, although various radiosensitizers can enhance radiation sensitivity, their significant toxicity and poor tumor selectivity limit safe and effective targeted sensitization [13,14]. Consequently, the dual challenges in HCC RT—precise dose delivery and enhanced radiobiological response—constrain further improvement in TGR.
Thus, innovation at both physical and biological levels remains fundamental for elevating TGR: the former enhances spatial dose precision through ART-based optimization, whereas the latter strengthens tumor radiosensitivity through targeted modulation. Despite their mechanistic complementarity, these two strategies have long evolved in parallel, lacking cross-dimensional integration or effective synergistic approaches. Establishing a physical–biological synergy model, whereby precise dose delivery and targeted radiosensitization are simultaneously amplified through cross-dimensional fusion, holds promise for substantially improving the therapeutic gain ratio of HCC RT. Such a synergistic paradigm leverages the advantages of both domains and enables system-level enhancement of radiotherapy efficacy, offering a new theoretical framework and practical direction for precision RT in HCC.
Guided by this concept, we developed a physical–biological synergistic system triggered by the tumor microenvironment and supported by intelligent imaging. Considering the mildly acidic microenvironment of HCC (pH 6.5–6.8), we designed charge-reversible gold–platinum nanoparticles (Charge Au@Pt NPs). These particles remain negatively charged under physiological pH to reduce nonspecific adsorption but undergo rapid charge reversal via amine protonation in acidic tumor regions, enabling acid-responsive accumulation and controllable aggregation. This behavior ensures sustained tumor retention for ≥5 days, significantly enhancing tumor-to-liver contrast on CT and providing a robust imaging foundation for reliable target delineation. Leveraging the enhanced CT quality, we developed a 3D STS-Net deep learning model incorporating a small-object data augmentation module (SODAM), a transformer-based encoder structure, and a cross-scale attention mechanism (CAM), achieving high-precision automated HCC segmentation (DCS = 0.86). This effectively overcomes the long-standing challenge of tumor auto-contouring in online ART for HCC. Concurrently, Charge Au@Pt NPs enhance radiation-induced reactive oxygen species (ROS) generation through platinum-mediated catalysis, markedly increasing HCC radiosensitivity from the biological perspective. Thus, the nanoplatform serves as a central hub linking “imaging enhancement–dose optimization–biological response amplification.”
In summary, this study proposes for the first time an integrated HCC RT strategy that unifies “imaging enhancement and intelligent segmentation” with “targeted radiosensitization.” By harnessing the synergistic action of nanomaterials and artificial intelligence (AI), we demonstrate the feasibility of simultaneously optimizing physical dose delivery and biological radiation response. This offers an innovative, integrated, and practical solution for enhancing the TGR in HCC RT.

2
Materials and methods
2.1
Study design
This study aimed to establish and validate a pioneering synergistic strategy that integrates both physical and biological dimensions to enhance the TGR in radiotherapy for hepatocellular carcinoma (HCC). The overall design consisted of three core components.①construction of Charge Au@Pt nanoparticles (NPs) with tumor microenvironment-responsive properties;

②development of an AI-based automatic segmentation model for adaptive radiotherapy using nanoparticle-enhanced CT imaging;

③evaluation of the radiosensitizing effects of Charge Au@Pt NPs.

As illustrated in the Graphical Abstract, ligand-mediated charge modulation enabled pH-triggered charge reversal and in situ aggregation of Charge Au@Pt NPs within the acidic tumor microenvironment. The nanoparticles were systematically characterized and further evaluated in an orthotopic small-HCC mouse model to determine whether they could provide long-lasting (several days), high-contrast CT enhancement. Subsequently, a three-dimensional Transformer–CNN hybrid automatic segmentation model (3D STS-Net) was constructed using nanoparticle-enhanced CT images. The model was trained and tested on CT datasets acquired from small HCC mice following Charge Au@Pt NP administration and compared with conventional models (3D U-Net and 3D U-Net++). Segmentation accuracy was assessed using standard quantitative metrics. Finally, the radiosensitizing potential of Charge Au@Pt NPs was validated through in vitro experiments and in vivo subcutaneous tumor models. Radiation-induced ROS generation, changes in cell viability, and tumor growth inhibition were quantitatively measured to evaluate whether the nanoparticles enhanced tumor response to ionizing radiation.
By integrating nanoparticle-enhanced imaging with AI-based automatic segmentation (physical dimension) and nanoparticle-mediated radiosensitization (biological dimension), this study establishes a comprehensive dual-dimensional synergistic strategy to improve the therapeutic efficacy of HCC radiotherapy.
Ethics statement: All animal experiments were conducted in accordance with the ARRIVE guidelines and approved by the Animal Ethics Committee of Zhejiang Chinese Medical University (Approval No. IACUC-20240611-20). All experimental procedures complied with national guidelines for the care and use of laboratory animals and the NIH Guide for the Care and Use of Laboratory Animals.
The sex of the animals used in this study was male, and sex-related effects on the study outcomes were not significant.

2.2
Nanoparticle synthesis and characterization
2.2.1
Single-factor experimental design
Au@Pt NPs were synthesized via a seed-mediated growth method. To optimize particle size and polydispersity index (PDI), a single-factor experimental design was applied, assessing the impact of various parameters:
Au/Pt molar ratios in AuNPs: K2PtCl6 (1.0 mmol L−1) at 1:1, 2:1, 3:1, 4:1, and 5:1; Reaction temperature: 40, 60, 80, 100, and 120 °C; Ascorbic acid dosage: 1, 5, 10, 15, and 20 mL of 5.0 mmol L−1 solution; Stirring time: 5, 15, 30, 60, and 90 min.
After the reactions, nanoparticle morphology was examined using TEM, while DLS quantified particle size and PDI to evaluate dispersity and uniformity.

2.2.2
Au@Pt NPs synthesis and structural characterization
Au@Pt NPs were prepared via seed-mediated growth. Briefly, 30 mL of Au NPs were mixed with 10 mL of K2PtCl6 solution (1.0 mmol L−1) and heated to 80 °C. Ascorbic acid (10 mL, 5.0 mmol L−1) was then added dropwise under continuous stirring for 30 min. The solution color transitioned from ruby red to brownish-red, indicating Pt shell formation. After cooling to room temperature, ICP-OES quantified the Au/Pt elemental composition. Nanoparticles were stored at 4 °C in the dark for subsequent experiments.
For morphology analysis, diluted Au@Pt NPs were dropped onto ultrathin carbon-coated copper grids, air-dried for 15 min, and excess liquid removed with filter paper, followed by TEM imaging. Surface morphology and elemental distribution were further characterized via Scanning Electron Microscopy–Energy Dispersive X-ray Spectroscopy (SEM-EDS). Freeze-dried samples were mounted on conductive adhesive, sputter-coated to enhance conductivity, and imaged at appropriate accelerating voltage, with EDS mapping confirming uniformity and composition.

2.2.3
Catalytic oxygen generation of Au@Pt NPs
To evaluate H2O2 decomposition and oxygen generation, 5% Na2SO3 served as zero-oxygen calibration, and air-saturated water as full-oxygen reference. The dissolved oxygen probe was calibrated in air for 3 min, followed by saturation of distilled water (300–500 mL) via ≥15 min aeration at constant temperature. Au@Pt NPs were diluted to 200 μg/mL and added to an equivalent H2O2 solution, with oxygen generation monitored using a portable dissolved oxygen meter over 15 min under gentle stirring, comparing to H2O2 control to assess catalytic efficiency.

2.2.4
Charge Au@Pt NPs synthesis and aggregation evaluation
Charge Au@Pt NPs were prepared by ligand exchange. Briefly, 0.42 mL of 20 mM 11-mercaptoundecanoic acid (MUA) and 0.21 mL of 20 mM 11-mercaptoundecyl-N,N,N-trimethylammonium bromide (MABr) were added to preformed Au@Pt NPs and reacted overnight on a shaker. Nanoparticles were purified via centrifugation (12,000 rpm, 10 min × 3) and resuspended in water.
For morphology characterization, freeze-dried samples were resuspended, ultrasonicated, and dropped onto carbon-coated grids, air-dried for 15 min, with residual liquid absorbed at 45° using filter paper prior to TEM imaging. UV–Vis spectra were recorded, and pH responsiveness was evaluated by adding Charge Au@Pt NPs to PBS at pH 5.0, 6.3, and 7.4, monitoring color and spectral changes. Cellular uptake was assessed in H22 cells (1 × 104 cells/well) incubated with Au@Pt or Charge Au@Pt NPs for 24 h, followed by fixation with 4% PFA or methanol/ethanol, ultrathin sectioning, and TEM observation.

2.2.5
Cytotoxicity assay
CCK-8 assays assessed cytocompatibility of Au@Pt NPs, Charge Au@Pt NPs, and their liquid or freeze-dried forms in H22 cells. Cells (1000 cells/well) were seeded in 96-well plates and cultured 24 h, followed by treatment with varying nanoparticle concentrations for 24 h. CCK-8 reagent was added, absorbance measured at 450 nm, and cell viability calculated.

2.2.6
Raman microscopy imaging
H22 cells were seeded on sterile coverslips in 12-well plates at 5 × 106 cells/mL, cultured 24 h at 37 °C and 5% CO2, then treated with Au@Pt or Charge Au@Pt NPs at 0, 4, 12, 24, and 48 h. After PBS washes and fixation with paraformaldehyde for 15 min, coverslips were imaged under Raman microscopy to assess intracellular nanoparticle distribution and accumulation.

2.2.7
Histological and organ safety evaluation
Three healthy mice received respective treatments. At the end of the dosing cycle, animals were euthanized, and major organs (heart, liver, spleen, lung, kidney) were harvested, rinsed with saline, fixed in formalin, dehydrated, paraffin-embedded, sectioned, and H&E-stained. Microscopic imaging was performed to assess tissue necrosis and pathological alterations.

2.3
In vivo CT imaging of orthotopic HCC
2.3.1
Orthotopic HCC mouse model
Female Balb/c mice (4–6 weeks old, 18.0 ± 2.0 g) were used. Initially, 1 × 106 logarithmic-phase H22 hepatocellular carcinoma (HCC) cells were subcutaneously injected into the dorsal region to establish a subcutaneous tumor model. Once tumors reached approximately 1 cm in diameter, they were excised and cut into 1 mm3 tissue blocks for later implantation. Subsequently, 40 mice were anesthetized, and abdominal hair removed and disinfected. A midline laparotomy was performed, and tumor tissue blocks were implanted beneath the hepatic capsule. Hemostasis was achieved using electrocautery, and the abdominal wall and skin were sutured. Postoperative monitoring indicated 32 mice survived after 7 days and were randomly assigned to either the Au@Pt NPs group or the Charge Au@Pt NPs group (n = 16 per group) for CT imaging studies.

2.3.2
In vivo CT enhancement imaging
Mice in each group received a tail vein injection of 200 μL Au@Pt NPs or Charge Au@Pt NPs at a concentration of 5 mg/mL. Imaging was performed using a MILabs U-CT system at baseline (0 h) and at 6, 24, 48, 72, 96, 120, and 144 h post-injection. Anesthesia was induced with 2–3% isoflurane and maintained at 1.5–2% during scanning. The scanning field covered approximately 1 cm from the superior diaphragm to the inferior liver margin. Acquisition parameters were set at 50 kV and 0.21 mA, with a reconstruction matrix of 1024 × 1024 × 800 and isotropic voxel size of 0.02 mm × 0.02 mm × 0.02 mm.

2.3.3
Computational environment and CT image preprocessing
CT image reconstruction and preprocessing were performed on a workstation equipped with an Intel i9-10900K CPU, NVIDIA RTX 3090 GPU, 64 GB RAM, and 1 TB SSD + 4 TB HDD, running Linux Mint 19.3 (64-bit). In Python (PyCharm V2.3.0), PyDicom V2.3.0 was used for image reconstruction. The original 3D CT volumes (1024 × 1024 × 800) were resampled to 0.2 mm × 0.2 mm × 0.2 mm isotropic spacing and resized to 512 × 512 × 400 voxels for quantitative analysis and automatic segmentation.
Reconstructed images were imported into RaySearch 9A (Stockholm, Sweden). Senior radiation oncologists manually delineated tumors and corresponding normal liver tissue of equal volume, located 2 cm from the tumor margin. The system calculated the mean CT values of Tumor and Liver regions as well as tumor volumes over time, enabling assessment of the targeting efficiency and in vivo imaging performance of Au@Pt and Charge Au@Pt NPs in small orthotopic HCC.

2.4
Development of the 3D STS-Net AI automatic segmentation model
2.4.1
Construction of the 3D STS-net automatic segmentation model
As illustrated in Fig. 4, the proposed 3D STS-Net model comprises three main components: a SODAM, a Tran-CAM, and a 3D U-Net decoder. The model is designed to achieve high-precision automatic segmentation of small HCC lesions under Charged Au@Pt NP-enhanced CT imaging.①SODAM

To address the challenges of small lesion detection and limited training samples, a 2.5D reconstruction strategy was employed, where all available consecutive 2D slices were used to create three-channel images, with each image formed by three consecutive slices. This approach ensured that the model learned from the full range of tumor depths. Additionally, diverse augmentations—including flipping, cropping, and scaling—were applied to the slices containing lesions. This SODAM was designed to enhance the model's sensitivity and recognition capability for small lesions.②Tran-CAM

The Tran-CAM module integrates a parallel Transformer and 3D U-Net structure to extract global semantic and local spatial features simultaneously. Multi-level feature fusion is achieved via a CAM, enhancing the representational capacity for small lesions. During the encoding stage, layer-wise interactions between the Transformer and U-Net encoders are implemented as described in Equation (1). The fused attention features are subsequently embedded into the U-Net encoder as formulated in Equation (2), yielding multi-scale semantic representations.③3D U-Net Decoder

During the decoding stage, features from each layer of the U-Net encoder {} are extracted and fused with the corresponding features of the current decoder layer {} via skip connections (Equation (3)), thereby integrating spatial details with semantic information:
Through the aforementioned fusion operations, the model effectively integrates multi-scale shallow spatial details with deep semantic information, thereby enhancing the accuracy of the final predicted masks and improving the localization of small HCC lesions in automatic segmentation.

2.4.2
Ablation study and ten-fold cross-validation of the 3D STS-net model
To optimize the performance and generalizability of the 3D STS-Net algorithm, ablation experiments and ten-fold cross-validation were conducted. In the ablation study, the SODAM + Tran-CAM modules were removed to yield a standard 3D U-Net, while in another variant only the Tran-CAM module was removed. Each model variant was retrained with adjusted hyperparameters to assess the contribution of SODAM and Tran-CAM to overall performance, guiding model optimization. Subsequently, ten-fold cross-validation was performed to further evaluate and enhance the model's generalization across the dataset. Performance metrics included Accuracy, Sensitivity, Specificity, and the Area Under the Receiver Operating Characteristic Curve (AUC), which are defined as follows:
Metrics are defined as follows: Accuracy: proportion of correct predictions. Sensitivity: ability to detect positive lesions. Specificity: ability to exclude negative samples.AUC: overall classification performance; closer to 1 indicates better performance.

2.4.3
Geometric and dosimetric evaluation of 3D STS-net automatic segmentation accuracy
To assess the accuracy of automatic segmentation, classical models 3D U-Net and 3D U-Net++ were included for comparison. The target volumes segmented by 3D STS-Net were further used for radiotherapy planning, with manual segmentation serving as the reference standard for dosimetric accuracy evaluation.①Deep Learning Model Training and Segmentation:

Tumor regions of interest (Tumor) in CT images of in situ tumor-bearing mice enhanced with Charge Au@Pt NPs were manually annotated and randomly split into training/validation (80%) and test (20%) sets. The training set was used to train 3D U-Net, 3D U-Net++, and the self-developed 3D STS-Net models, while the test set was used to evaluate the segmentation performance of these three models for small hepatocellular carcinoma.②Geometric Accuracy Evaluation:

Using manual segmentation as the reference, the geometric accuracy of automatic segmentation was evaluated by Dice Similarity Coefficient (DSC), Union over Intersection (UoI), and Hausdorff Distance (HD). The definitions of these metrics are as follows:Here, A and B represent the automatically segmented region and the reference standard, respectively. In Equations (8), (9), DSC and UoI range from 0 to 1, with higher values indicating closer agreement with the reference; a value of 1 represents perfect overlap. Equation (10) reflects the maximum boundary deviation between automatic and true segmentation, with smaller values indicating better boundary alignment.③Dosimetric Accuracy Evaluation:

Based on clinical radiotherapy protocols for small HCC, the automatically segmented tumor region by 3D STS-Net (tumor-A) and the manual segmentation (tumor-M) were each expanded outward by 5 mm to generate planning target volumes PTV-A and PTV-M (Fig. 5D). The radiotherapy prescription was 5000 cGy delivered in 10 fractions, with 95% of the target volume required to receive the prescribed dose. Using the RaySearch 9.1 system with the built-in CCC algorithm, IMRT plans and dose calculations were performed with four different beam angles (Fig. 5D and E). Key dosimetric parameters—target coverage (V5000), mean dose (Dmean), conformity index (CI), and homogeneity index (HI)—were extracted, with the latter two calculated according to Equations (11) and (12).In Equation (11), TVRI is the target volume receiving the reference dose, TV is the total target volume, and VRI is the total volume receiving the reference dose or higher. In Equation (12), D2%, D98%, and D50% represent the doses received by 2%, 98%, and 50% of the target volume, respectively.

2.5
Radiotherapy sensitization experiments
In this study, in vitro experiments employed the human hepatocellular carcinoma cell line HepG2 to model the biological response of human tumor cells to nanoparticles and radiotherapy, while in vivo experiments used the murine H22 hepatoma cell line to establish a stable and reproducible solid tumor model for evaluating radiotherapy sensitization. The combination of these two models enables a comprehensive assessment of the radiosensitizing performance of Au@Pt-based nanoplatforms at both cellular and organismal levels. Radiotherapy was delivered using an Elekta Infinity™ linear accelerator (Stockholm, Sweden).
2.5.1
In vitro experiments
HepG2 cells (Chinese Academy of Sciences Cell Bank) were cultured in DMEM supplemented with 10% fetal bovine serum at 37 °C in 5% CO2, with medium renewal every 2–3 days. At ∼80% confluence, cells were digested with 0.25% trypsin–0.02% EDTA and passaged. The following radiosensitization assays were conducted:
Cell viability (CCK-8 assay): HepG2 cells were seeded at 5 × 104 cells/mL in 96-well plates (100 μL/well) and allowed to attach for 24 h. Cells were exposed to radiotherapy doses of 0, 2, 4, 6, 8, or 10 Gy to establish a dose–response curve, with simultaneous treatment with Au@Pt NPs or Charge Au@Pt NPs (0–80 μg/mL) to assess single-agent effects. Six experimental groups were designed: Control, Au@Pt NPs, Charge Au@Pt NPs, RT, Au@Pt NPs + RT, and Charge Au@Pt NPs + RT. After 24, 48, and 72 h, 0.5 mg/mL CCK-8 reagent was added and incubated for 1 h. Absorbance at 490 nm was measured to calculate cell viability.
Flow cytometry for apoptosis: HepG2 cells were seeded in 6-well plates (3 mL/well) and treated as above. After 48 h, cells were collected, washed with PBS, and incubated with Annexin V-FITC/PI for 15 min in the dark. Early and late apoptotic fractions were analyzed using flow cytometry.
Colony formation assay: Log-phase HepG2 cells were seeded at 1000 cells/well in 6-well plates and treated as above, with the radiotherapy group receiving 4 Gy X-ray (6 MV). After 14 days of incubation, colonies were fixed with 4% paraformaldehyde for 15–30 min, washed with PBS, stained with crystal violet for 1 h, air-dried, photographed, and quantified.

2.5.2
In vivo experiments
H22 cells were prepared as a 1 × 106 cells/mL suspension and subcutaneously injected (100 μL) into the right dorsal flank of 4–6-week-old BALB/c nude mice to establish a xenograft model. Tumor dimensions were measured every 2 days (length L, width W) and tumor volume calculated as V = L × W × 0.5. When tumor volumes reached 50–100 mm3, mice were randomized into four groups (n = 5): Control, RT (12 Gy), Charge Au@Pt NPs, and Charge Au@Pt NPs + RT. Nanoparticles were administered via tail vein injection (200 μL, 2 mg/mL). Radiotherapy consisted of a single 12 Gy dose of 6 MV X-rays. Tumor volume and body weight were monitored for 21 days post-treatment. At study endpoint, animals were euthanized, and tumors and major organs were harvested, weighed, and organ indices calculated to evaluate treatment response and systemic toxicity.

2.5.3
In vivo TEM observation
At 24 h after intravenous administration of Au@Pt NPs or Charge Au@Pt NPs (200 μL, 2 mg/mL), the subcutaneous tumor tissues were excised and cut into approximately 1 mm3 cubes. The tissue samples were subsequently fixed in 3% glutaraldehyde, followed by epoxy resin embedding. Ultrathin sections were prepared using an ultramicrotome and collected onto 150-mesh copper grids. The sections were then stained with 2% uranyl acetate in ethanol for 10 min and lead citrate for 5 min, and finally examined under a TEM.

2.5.4
Fluorescence imaging of tumor sections
At 24 h after intravenous injection of free Cy5 or Cy5-loaded Charge Au@Pt NPs at a Cy5-equivalent dose of 1.2 mg kg−1, the tumors were harvested and fixed in 4% paraformaldehyde. The samples were then cryoprotected sequentially in 15% and 30% sucrose solutions, each for 24 h. Frozen sections were prepared and counterstained with DAPI, followed by imaging using a confocal laser scanning microscope.

2.5.5
In vivo biodistribution analysis
After intravenous administration of Charge Au@Pt NPs (200 μL, 2 mg/mL), mice were sacrificed at predetermined time points (1, 2, 3, 4, 5, 6, and 7 days). The heart, liver, spleen, lung and kidney tissues were harvested. In addition, urine and feces were collected. All samples were subjected to acid digestion, and the Au content was quantified by inductively coupled plasma mass spectrometry (ICP-MS).
It is noted that the subcutaneous tumor model was selected because orthotopic tumors are small, deeply located, and adjacent to the liver and abdominal organs. High-precision irradiation of such small orthotopic tumors in mice is limited by the beam collimation accuracy and positioning error of small-animal radiotherapy devices. In contrast, subcutaneous tumors are easily exposed with clear boundaries, allowing precise local irradiation and direct monitoring of tumor volume, making them suitable for quantitative assessment of radiosensitization effects.

2.6
Statistical analysis
For non-normally distributed data, differences between two groups were analyzed using the Mann–Whitney U test, and differences among multiple groups were analyzed using the Kruskal–Wallis H test. Normality was verified using the Shapiro–Wilk test. Pearson correlation analysis was further applied to explore relationships between dosimetric indices and geometric accuracy metrics of automatically segmented target volumes. All statistical analyses and figure generation were performed using Origin 2021 software, with
p < 0.05 considered statistically significant.

3
Results and discussion
3.1
Synthesis and characterization
Au@Pt nanoparticles (Au@Pt NPs) were synthesized via a seed‐mediated growth approach, as illustrated in Fig. 1A. The formation of Au@Pt NPs was jointly regulated by the Au/Pt molar ratio, reaction temperature, reductant concentration, and reaction time (Fig. S1 and Table S1). Systematic optimization showed that increasing the Au:Pt ratio to 3:1 yielded the most favorable balance between Pt deposition kinetics and growth dynamics, resulting in optimal dispersity and size uniformity (Fig. S1A). Ratios deviating from this value led to insufficient Pt incorporation or excessive particle shrinkage, underscoring the essential role of an appropriate Pt amount in obtaining structurally stable Au–Pt composite nanomaterials [15]. Temperature also markedly influenced Pt nucleation: low temperatures limited reduction, whereas excessively high temperatures induced rapid nucleation and aggregation (Fig. S1B). A reaction temperature of 80 °C provided an ideal balance between nucleation and growth, producing morphologically stable and interface‐uniform bimetallic structures [16]. In addition, an appropriate amount of ascorbic acid ensured continuous Pt deposition; insufficient reductant resulted in shell defects or particle coarsening [17]. The major growth stage occurred within approximately 30 min, after which the structure became stable. Based on these parameters, the optimal synthesis conditions were determined to be Au:Pt = 3:1, 80 °C, appropriate ascorbic acid concentration, and a 30-min reaction time, which produced uniformly dispersed and structurally homogeneous Au@Pt NPs—consistent with prior theoretical and experimental findings [18].
Based on previously reported strategy, ligands with distinct charge properties were introduced onto the surface of Au@Pt NPs [8]. Under acidic conditions, these ligands undergo protonation, which induces a redistribution of surface charges and alters the overall charge distribution; this process further enhances interparticle electrostatic interactions, thereby promoting nanoparticle aggregation. Consistently, the zeta potential exhibits a gradual and continuous transition from negative to positive values with decreasing pH, increasing from approximately −13 mV at pH 7.4 to near-neutral at pH 5.0, and further to about +30 mV at pH 3.0 (Fig. S2). This evolution indicates that the surface charge variation arises from progressive protonation of functional groups (e.g., –COO- to –COOH) and enhanced contribution of protonated amine groups, supporting a charge redistribution mechanism rather than a simple charge reversal. The morphology and structure of Au@Pt NPs and Charge Au@Pt NPs are shown in Fig. 1B–E. Transmission electron microscopy (TEM) images revealed a near‐spherical morphology with an average diameter of 21.30 ± 1.80 nm (Fig. 1B). Energy-dispersive X-ray spectroscopy (EDS) mapping confirmed the homogeneous distribution of Au and Pt within the particles (Fig. 1C). Surface charge modification slightly increased particle size without compromising dispersity (Fig. 1D). High‐resolution TEM revealed two sets of lattice fringes with spacings of 0.2604 nm and 0.2357 nm, corresponding to characteristic Au and Pt crystal planes, respectively (Fig. 1E). Their continuous and interwoven lattice arrangement within a single nanoparticle suggests the formation of tightly coupled Au–Pt heterointerfaces at the atomic scale. Such atomically interconnected interfaces facilitate rapid electron transfer between Au and Pt, thereby enhancing electron coupling and improving their synergistic catalytic behavior.
Structural analyses further corroborated these observations. X-ray diffraction (XRD) patterns (Fig. 1F) displayed typical face‐centered cubic reflections of noble metals, indicating well‐defined crystalline phases. X-ray photoelectron spectroscopy (XPS) spectra (Fig. S3) confirmed the incorporation of both Au and Pt, with characteristic Au 4f and Pt 4f peaks. Both elements exhibited their expected metallic oxidation states, and the stable binding energies supported the successful construction of the Au–Pt composite framework. Ultraviolet–visible spectroscopy(UV–Vis) spectra (Fig. 1G) showed that Charge Au@Pt NPs retained the characteristic Au surface plasmon resonance (SPR) at 520 nm, whereas the enhanced peak intensity suggested ligand‐induced modulation of the local electromagnetic environment, which may influence surface charge behavior and microaggregation dynamics [13].
In the H2O2 catalytic assays (Fig. 1H and I), Charge Au@Pt NPs significantly accelerated H2O2 decomposition and increased dissolved oxygen production, exhibiting a catalytic efficiency far exceeding the spontaneous decomposition of H2O2. These results indicate that surface charge engineering not only preserves the intrinsic catalytic advantages of the bimetallic structure, but also enhances enzyme‐like activity. This enhancement arises from interfacial electronic modulation and increased accessibility of active sites. Their strong oxygen‐generating catalytic performance provides a physicochemical basis for alleviating tumor hypoxia and improving radiosensitivity.
Collectively, both Au@Pt NPs and Charge Au@Pt NPs exhibited uniform structures, well‐defined crystallinity, and excellent catalytic properties. Notably, charge modification further enhanced optical responses and enzyme‐mimetic activity, highlighting their promising potential in tumor oxygen modulation and radiotherapy sensitization.

3.2
In vitro validation of the acid-responsive aggregation of Charge Au@Pt NPs
To verify the acid-responsive aggregation behavior of Charge Au@Pt NPs and assess their cellular uptake and retention properties, we conducted pH-dependent morphological, optical, and uptake-kinetic analyses. TEM imaging revealed that the nanoparticles were uniformly dispersed under physiological pH 7.4, whereas at acidic pH 5.0 mimicking the tumor microenvironment, the interparticle distance markedly decreased and the particles rapidly assembled into compact clusters (Fig. 2B), indicating a typical acid-triggered aggregation phenotype. This trend was further confirmed by dynamic light scattering, where the hydrodynamic diameter increased sharply from ∼31.31 ± 1.81 nm to nearly 1097.73 ± 77.98 nm, demonstrating pronounced aggregation under acidic conditions.
UV–Vis spectroscopy provided additional evidence for these structural transitions. Compared with the stable 520-nm SPR peak observed at pH 7.4, nanoparticles at pH 5.0 exhibited an obvious red-shift accompanied by reduced absorbance intensity (Fig. 2C and D). Such changes are consistent with enhanced interparticle electromagnetic coupling and SPR mode alterations induced by aggregation. Together, these results confirm that the rapid and reversible aggregation of Charge Au@Pt NPs in acidic environments is driven by pH-dependent surface charge switching and strengthened interfacial interactions. This behavior establishes a physicochemical basis for enhanced tumor retention and more efficient cellular internalization.
Following confirmation of their acid-responsive aggregation, we next evaluated the biocompatibility and intracellular behavior of the nanoparticles. CCK-8 assays showed no detectable cytotoxicity for either nanoparticle type, even at concentrations up to 80 μg mL−1 (Fig. 2E), indicating favorable biological tolerance. Notably, TEM observations of ultrathin cellular sections revealed clear differences in internalization patterns: unmodified Au@Pt NPs were sparsely distributed in the cytoplasm, whereas Charge Au@Pt NPs accumulated extensively and formed cluster-like aggregates within cells (Fig. 2F). This phenomenon suggests that surface charge modification substantially enhances particle–membrane interactions, thereby markedly increasing endocytic uptake, consistent with established mechanisms for charged nanomaterials [19].
Raman/confocal imaging further demonstrated distinct retention kinetics between the two nanoparticle types. The intracellular signal of Au@Pt NPs diminished rapidly to near-background levels within 24–48 h, whereas Charge Au@Pt NPs showed substantial accumulation as early as 4 h and maintained strong intracellular signals even at 48 h (Fig. 2G and H). This pronounced prolongation of intracellular residency aligns well with their acid-triggered aggregation behavior and corroborates the principle proposed by Souri et al., namely that pH-induced nanoparticle clustering can significantly extend intracellular retention [20].
Collectively, Charge Au@Pt NPs undergo controlled aggregation under acidic conditions, accompanied by corresponding morphological and optical changes that translate into enhanced cellular uptake and prolonged intracellular retention. This intrinsic acid-responsive aggregation–retention mechanism provides a robust physicochemical and biological foundation for their application as long-duration CT imaging probes.

3.3
In vivo CT contrast-enhanced imaging
After establishing the physicochemical properties and pH-responsive behavior of the nanoparticles, we further evaluated the in vivo imaging performance of Au@Pt NPs and Charge Au@Pt NPs and their contribution to tumor visualization. As shown in Fig. 3A, dynamic CT imaging revealed that, prior to injection, the tumor was indistinguishable from the liver parenchyma; a clear contrast emerged at 6 h post-injection, diminished by 24 h, and nearly disappeared by 48 h. Among 16 mice, 10 completed imaging at 48 h. Quantitative analysis of these 10 mice (Fig. 3B and Table S2) showed that tumor volume decreased from 0.33 ± 0.08 cm3 to 0.21 ± 0.05 cm3 within 24 h, while electron density declined from 1718.45 ± 232.44 HU to 916.73 ± 193.13 HU. Liver electron density peaked at 24 h before slightly decreasing, indicating rapid intratumoral clearance and a short effective imaging window for Au@Pt NPs.
Fig. 3C displays the CT imaging system and the 3D reconstruction after the administration of Charge Au@Pt NPs. The small HCC region exhibits high-density signals. Fig. 3D shows that Charge Au@Pt NPs generated a stable high-contrast signal from 6 h onward, which remained nearly unchanged up to 120 h, with a contraction of the enhancement volume observed at 144 h. Among 16 mice, 8 completed imaging at 144 h. As shown in Fig. 3E and Table S3, tumor volume in these 8 mice remained stable from 6 to 120 h but significantly decreased at 144 h (mean reduction 37.8%); electron density continuously increased and peaked at 144 h (1701.86 ± 184.58 HU → 2445.14 ± 161.01 HU), likely due to the formation of compact aggregates within the acidic microenvironment. Liver electron density remained stable throughout, suggesting rapid clearance from normal hepatic tissue.
Collectively, Charge Au@Pt NPs exhibited acid-triggered aggregation in the tumor microenvironment, markedly delaying in vivo clearance and thereby producing prolonged CT contrast enhancement. This behavior aligns with our previous observations for Charge AuNPs [21] and is consistent with the inherently strong X-ray absorption derived from the similar atomic numbers of Au and Pt. In contrast to the <24 h imaging window of Au@Pt NPs, Charge Au@Pt NPs extended the effective imaging duration beyond 120 h, outperforming previously reported pH-responsive NPs that showed approximately 48 h persistence in subcutaneous tumor models [22]. Such long-lasting enhancement arises from both the high X-ray attenuation of Au–Pt alloyed structures and the ligand-mediated pH-responsive regulation. This regulation induces aggregation or catalytic reactions in the TME, sustaining or further amplifying the signal [23]. In vivo experiments also confirmed that both nanoparticle types improved small HCC CT contrast while maintaining favorable biocompatibility.
In conventional clinical workflows, CT-based ART localization for liver cancer is often reliant on indirect liver contour matching, which provides limited accuracy. Metallic fiducial markers can improve precision but require invasive placement [24], whereas iodine-based contrast agents suffer from a very short imaging window and may pose toxicity risks if administered daily during radiotherapy [25]. In this study, Charge Au@Pt NPs required only a single injection to maintain stable tumor enhancement throughout a one-week (5-day) radiotherapy course, fulfilling the needs of ART target identification and registration and demonstrating strong clinical translation potential. Moreover, numerous previous studies have already validated nanoparticle distribution and imaging accuracy in hepatic tumors [26,27]; therefore, this work focused on imaging persistence and metabolic differences rather than repeating accuracy assessments.

3.4
Cross-validation and ablation study of the STS model
After confirming the long-duration and high-contrast CT enhancement of small HCC by Charge Au@Pt NPs, we constructed a deep learning–based automatic segmentation model, 3D STS-Net, tailored for nanoparticle-enhanced imaging. A total of 84 Charge Au@Pt NPs–enhanced 3D small HCC CT datasets were included for ablation studies and cross-validation. These were split at an 8:2 ratio into a training set (n = 68) and a test set (n = 17). The overall architecture of 3D STS-Net is illustrated in Fig. 4A.
The ablation results (Fig. 4C) showed that the complete 3D STS-Net produced segmentation masks most consistent with manual annotations. Removing both the SODAM and Cross-Attention Mechanism (Tran-CAM) modules markedly degraded model performance, reducing it to a conventional 3D U-Net, whereas removing only the Tran-CAM module resulted in intermediate performance. These findings confirm the necessity and complementary contributions of each module. Tenfold cross-validation (Fig. 4B, D and Table S4) further demonstrated the superiority of 3D STS-Net: the complete model achieved an accuracy of 0.86 ± 0.05, outperforming the version without Tran-CAM (0.84 ± 0.04) and the 3D U-Net (0.80 ± 0.08). Sensitivity, specificity, and AUC exhibited similar trends. Notably, the model lacking only Tran-CAM still surpassed the version without both modules, highlighting the critical contribution of Tran-CAM to overall performance. Collectively, preserving both SODAM and Tran-CAM significantly improves feature representation and lesion detection capability in 3D STS-Net.
Small lesion size and limited samples are major challenges in this task [28]. SODAM, which leverages 2.5D reconstruction and multi-sample augmentation, has been shown to enhance the learnability of small lesions and improve model generalization in prior studies [29,30], and was therefore used to strengthen local structural reconstruction and spatial contrast in STS-Net. Meanwhile, Transformers exhibit advantages in modeling low-contrast regions and small targets [31], and CAN enable selective mapping across multi-scale or heterogeneous features, enhancing feature complementarity and discriminative power [32]. Accordingly, we designed Tran-CAM to integrate the global semantic modeling capability of Transformers with the local feature extraction of 3D CNNs, using cross-scale attention to enhance representations at blurred boundaries and in low-contrast regions.
In summary, SODAM provides effective data-level sample augmentation and small-lesion enhancement, whereas Tran-CAM provides feature-level global–local semantic fusion. Their synergistic effect is the key determinant of the improved performance of 3D STS-Net in this task.

3.5
Geometric and dosimetric accuracy of automatic segmentation
Following model training and testing, we further evaluated the automatic segmentation performance of 3D STS-Net on the test set and applied the resulting target volumes in radiotherapy planning experiments to assess both geometric and dosimetric accuracy. Fig. 5A presents three models. As shown in Fig. 5B, tumor contours generated by 3D STS-Net exhibited the highest overlap with manual delineations. In the 17-case test set, 3D STS-Net applied to Charge Au@PtNPs-enhanced CT achieved the best performance across all geometric metrics (Fig. 5C and Table S5): DSC = 0.86 ± 0.02, UoI = 0.83 ± 0.02, 95% HD = 0.9 ± 0.3 mm, all significantly superior to 3D U-Net and 3D U-Net++ (p < 0.05), demonstrating enhanced robustness and boundary delineation capability.
CNN-based frameworks remain the mainstream in medical image segmentation, with U-Net and its extensions (e.g., 3D U-Net++, nnU-Net) performing robustly across various tumor segmentation tasks [33,34]. However, for small, low-contrast lesions such as non-iodine-enhanced small HCC, standard U-Net models still encounter performance limitations [11,12]. In this study, Charge Au@PtNPs markedly improved tumor–liver contrast, allowing even 3D U-Net++ to achieve favorable results (DSC = 0.83), approaching the performance reported for MR or iodine-enhanced CT in literature [35,36]. The proposed 3D STS-Net further achieved significant gains in geometric accuracy (DSC = 0.86), highlighting the synergistic advantage of nanocontrast enhancement and a task-specific segmentation network.
Fig. 5D, E illustrate radiotherapy plans based on automatic segmentation. As shown in Fig. 5F and Table S6, the V5000 of PTV-M generated by 3D STS-Net was 94.1 ± 0.98%, which, although statistically different from the manual PTV-A, remained within clinically acceptable limits. Other dosimetric parameters (D_mean, CI, HI) showed no significant differences, indicating that automatic segmentation can achieve dose coverage comparable to manual delineation. Previous studies have noted that geometric accuracy does not necessarily guarantee good dosimetric performance; careful verification remains essential. For example, Kawula et al. [37] reported in prostate radiotherapy that even with <2% deviation in 3 Gy coverage, manual review was required. In our study, 3D STS-Net maintained high dosimetric accuracy across the full dataset, with a maximum target volume coverage deviation of only 2.5%, surpassing prior reports.
To further investigate the relationship between geometric accuracy and dose delivery, Pearson correlation analysis was performed (Fig. 5G–Table S7), revealing a positive correlation and confirming that high-quality segmentation contributes to superior dosimetric precision, thereby reinforcing the clinical potential of 3D STS-Net in ART.
Because the model was trained on Charge Au@Pt NPs–enhanced CT images, its strong performance is partly associated with the distinctive imaging signature generated by these nanoparticles, reflecting the material–algorithm co-designed strategy demonstrated in this study. The model has not yet been validated on CT datasets enhanced by other nanoparticle formulations or conventional iodine-based contrast agents. However, since the network primarily learns high-contrast boundary and texture features from nanoparticle-enhanced images, it may retain certain transferability and could potentially be adapted to other contrast conditions through transfer learning or limited fine-tuning. In addition, the current model was trained and validated on a relatively small and homogeneous dataset derived from Charge Au@Pt NPs–enhanced mouse CT images acquired under specific imaging protocols and scanners. Future studies with larger and more diverse datasets will therefore be necessary to further evaluate and improve the generalizability of the model. Despite these limitations, the current model demonstrates promising potential for nanoparticle-enhanced ART applications.
In summary, the combination of Charge Au@Pt NPs–enhanced imaging with the 3D STS-Net significantly improves the automatic segmentation performance of small HCC, outperforming conventional models in both geometric and dosimetric accuracy. The enhanced signal remains stable for over five days, enabling automatic segmentation of tumor targets throughout a one-week radiotherapy course. This provides a reliable solution to the technical challenges of automatic target delineation in HCC ART, ultimately making HCC ART feasible and offering the potential to improve TGR from a physical dimension in radiotherapy.

3.6
Radiotherapy Sensitization Effects
In previous studies, we demonstrated that the Au@Pt-based nanosystem achieves efficient tumor accumulation and CT contrast enhancement in vivo, providing a foundation for its potential radiosensitization effect. To further elucidate its biological function, we systematically evaluated the radiosensitizing performance and underlying mechanisms of Au@Pt NPs and Charge Au@Pt NPs both in vitro and in vivo.
CCK-8 assays (Fig. S4) showed that radiotherapy alone induced a typical dose- and time-dependent inhibition of HepG2 cell viability, with 4 Gy irradiation for 48 h reducing cell survival to 82.02% ± 10.30%. Treatment with either Au@Pt NPs or Charge Au@Pt NPs alone did not significantly affect cell viability (p > 0.05), indicating excellent biocompatibility and negligible cytotoxicity. Based on these results, subsequent experiments employed 4 Gy irradiation combined with 20 μg/mL nanoparticle treatment as the standard conditions.
Six-group comparative experiments further demonstrated that combination therapy significantly decreased cell viability (p < 0.05), with the Charge Au@Pt NPs + 4 Gy group exhibiting the strongest inhibitory effect (37.19% ± 5.47%) (Fig. 6A). Flow cytometry analysis (Fig. 6B) indicated that combination treatment effectively induced apoptosis, showing a transition from early to late apoptotic stages, accompanied by increased necrosis. Colony formation assays (Fig. 6C) revealed that Charge Au@Pt NPs +4 Gy reduced the colony formation rate to 4.69% ± 1.60% (p < 0.01), confirming significant long-term proliferation inhibition.
Moreover, TEM analysis of tumor tissue sections harvested 24 h after intravenous administration of Au@Pt NPs and Charge Au@Pt NPs (Fig. 6D) clearly revealed that the charged Au@Pt nanoparticles exhibited pronounced aggregation within the tumor region, forming dense clustered structures, whereas the unmodified Au@Pt nanoparticles remained largely dispersed. Confocal microscopy further confirmed the targeting capability of the delivery system, showing that Cy5-Charge Au@Pt NPs generated stronger fluorescence signals in the tumor region compared to Cy5-Au@Pt NPs (Fig. 6E). This distinct difference further supports the aggregation behavior of Charge Au@Pt NPs within the tumor microenvironment. To further evaluate their in vivo fate and biosafety, we systematically investigated the biodistribution and clearance behavior of Charge Au@Pt NPs. As shown in Fig. 6F, Au exhibited clear time-dependent excretion profiles. Specifically, the majority of nanoparticles were gradually eliminated via renal pathways, while a smaller fraction was excreted through the hepatobiliary route, indicating dual clearance mechanisms. Notably, the continuous increase in cumulative excretion over 7 days suggests that the nanoparticles are not permanently retained but undergo ongoing elimination in vivo. Biodistribution analysis further showed that nanoparticles predominantly accumulated in the liver at early time points, consistent with uptake by the reticuloendothelial system (RES), followed by a gradual decrease over time, indicating progressive clearance. Meanwhile, a measurable level of nanoparticle retention was consistently observed in tumor tissues throughout the study period, which may be partially attributed to the proposed pH-responsive aggregation behavior.
Animal studies corroborated the in vitro findings, with experimental design illustrated in Fig. 6G. In subcutaneous tumor models, body weight differences among groups were not significant (p > 0.05, Fig. 6J), indicating good overall tolerability. Tumor volume monitoring showed that combination therapy, particularly Charge Au@Pt NPs +12 Gy, achieved the most pronounced tumor inhibition (Fig. 6H and I), with the lowest final tumor weights (Fig. 6K). Analysis of organ indices (Fig. 6L) revealed no significant differences, suggesting no overt organ toxicity from the nanoparticles.
Collectively, these results indicate that Au@Pt NPs exhibit clear radiosensitization effects, primarily through two mechanisms: (i) the high Z properties of Au and Pt enhance photoelectric and Compton effects, increasing energy deposition and amplifying ROS generation [38]; (ii) the intrinsic enzyme-like activity of the nanoparticles catalyzes H2O2 decomposition to produce oxygen [39], as confirmed by the H2O2 catalytic experiments in Fig. 1H and I, alleviating tumor hypoxia. On this basis, Charge Au@Pt NPs demonstrated superior radiosensitization in both in vitro and in vivo settings, highlighting the importance of surface charge. Positive charge modification preserves the dual-metal catalytic and high-Z amplification effects while strengthening interactions with negatively charged cell membranes, significantly enhancing cellular uptake and tumor retention, thereby increasing effective accumulation [40]. Additionally, surface charge facilitates oxygen generation and amplifies radiation-induced ROS, triggering mitochondrial membrane potential depolarization, Ca2+ imbalance, and activation of MAPK/p53 stress pathways, resulting in stronger and more sustained cell death signals, consistent with previously reported mechanisms for positively charged metal nanoparticles [41].
This study elucidates a clear “dual-enhancement” mechanism: Au@Pt NPs achieve baseline radiosensitization through high-Z element amplification and catalytic oxygen generation, while positive surface charge modification further promotes cellular uptake and tumor retention, resulting in superior sensitization efficacy of Charge Au@Pt NPs. Supported by favorable biocompatibility, Charge Au@Pt NPs demonstrate potential as a multi-mechanism synergistic radiosensitization platform, offering the prospect of significantly enhancing the TGR of HCC radiotherapy at the biological dimension.
Moreover, blood routine tests and histological analysis of major organs (Fig. S5) showed no significant differences among groups. Organ structures were intact, with no observable tissue damage or pathological changes, indicating that both Au@Pt NPs and Charge Au@Pt NPs possess excellent in vivo biosafety.

4
Conclusion
In this study, we successfully developed Charge Au@Pt NPs with acid-responsive aggregation properties. In an orthotopic small HCC model, this probe achieved sustained high-contrast CT enhancement for up to 120 h, overcoming the limited imaging window (∼48 h) of conventional nanoprobes [22] and providing a long-lasting, high-contrast imaging foundation for adaptive radiotherapy of small hepatocellular carcinoma. Building on nanoparticle-enhanced imaging, we further established a Transformer-CNN-based 3D STS-Net AI model, which enabled high-precision three-dimensional automatic segmentation of HCC in CT images in small-animal experiments. This approach successfully addressed the technical bottleneck of automatic target delineation in CT-guided HCC ART, demonstrating the feasibility of the “nanoparticle-enhanced imaging + AI segmentation” strategy for online ART and offering a novel solution and developmental framework for HCC adaptive radiotherapy. In terms of biological radiosensitization, both in vitro cell studies and subcutaneous tumor models preliminarily validated the radiosensitizing effects of Charge Au@Pt NPs.
Collectively, this work introduces, for the first time, a synergistic strategy combining “physical adaptation + biological sensitization” to enhance tumor growth response in HCC radiotherapy, providing an integrated theoretical framework and practical direction for precision HCC radiotherapy.
Nevertheless, this study has several limitations. Adaptive radiotherapy (ART) generally involves two levels: (i) image-guided target re-delineation and plan adaptation, and (ii) real-time dose or biological effect modulation during treatment. The present work focuses on the first level by improving nanoparticle-enhanced CT imaging and AI-based automatic segmentation of HCC targets, while the second level remains beyond the scope of this study.
In addition, imaging experiments were conducted in orthotopic liver tumor models, whereas radiosensitization studies used subcutaneous models, as precise irradiation of small orthotopic tumors in small animals remains technically challenging. Future studies will integrate nanoparticle-enhanced imaging, AI-driven segmentation, and image-guided radiotherapy in orthotopic models to advance toward a fully integrated ART closed-loop system.

CRediT authorship contribution statement
Ji-Gang Piao: Data curation, Conceptualization. Liting Chen: Data curation. Weiyi Cheng: Methodology. Lijuan Shen: Data curation, Validation. Wenfan Deng: Validation. Yuan Yang: Methodology. Yicun Li: Investigation. Wenhao Lin: Writing – review & editing. Jianjun Lai: Writing – original draft.

Ethics approval and consent to participate
All animal experiments were conducted in accordance with the ARRIVE guidelines and approved by the Animal Ethics Committee of Zhejiang Chinese Medical University (Approval No. IACUC-20240611-20). All experimental procedures complied with national guidelines for the care and use of laboratory animals and the NIH Guide for the Care and Use of Laboratory Animals.

Funding
This study was supported by the Zhejiang Provincial Medical and Health Science and Technology Program (2025KY003) and the Zhejiang Provincial Education Science and Technology Program (Y202454960).

Declaration of competing interest
The authors declare that there is no conflict of interest regarding the publication of this article.