DosimetrEYE: AI Preclinical Dosimetry System

• DosimetrEYE is a generative AI-based system using a 3D conditional GAN to transform 2D SPECT/CT images into accurate 3D absorbed dose distributions.
• It integrates a hybrid anomaly detection layer combining statistical GMM and CLIP-PCA methods to ensure reliability by flagging out-of-distribution samples.
• Beyond preclinical nuclear medicine, DosimetrEYE serves as a conceptual framework for diverse dosimetry applications including ocular brachytherapy, active eye-lens monitoring, and real-time beam tracking.

DosimetrEYE most specifically denotes BIOEMTECH’s generative artificial intelligence model for preclinical nuclear medicine that estimates 3D absorbed radiation dose distributions from 2D SPECT/CT scans and is integrated into the eyes™ small-animal imaging systems (Binda et al., 11 Aug 2025). In the literature considered here, the same name is also used more broadly as a hypothetical label for several other dosimetry architectures, including ocular brachytherapy planning platforms, active eye-lens dosemeters, HDR brachytherapy source-tracking systems, and real-time beam-monitoring devices. This suggests that the term functions both as a concrete product designation and as a broader conceptual shorthand for real-time, model-based, or sensor-based dosimetry.

1. Scope and nomenclature

A common source of confusion is that not every appearance of the name refers to the same technical object. The most concrete and fully specified use is the BIOEMTECH system described in "Safeguarding Generative AI Applications in Preclinical Imaging through Hybrid Anomaly Detection" (Binda et al., 11 Aug 2025). There, DosimetrEYE is a 3D conditional GAN (Vox2Vox) for preclinical nuclear medicine.

By contrast, several other texts use the name as a hypothetical or prospective platform label. "Update of the CLRP eye plaque brachytherapy database for photon-emitting sources" describes how “DosimetrEYE” built on CLRP EPv2 can serve as a fully model-based, plaque-specific ocular brachytherapy planning and research platform (Safigholi et al., 2021). Other summaries similarly use the name for active personal dosimetry, eye-lens dosimetry, HDR brachytherapy tracking, proton beam monitoring, UHDR electron FLASH monitoring, and deformable scintillation dosimetry (Salvi et al., 2023, Beißer et al., 2023, Tho et al., 2018, Yap et al., 2021, Vanreusel et al., 2023, Cloutier et al., 2021). The technical details in those cases belong to the underlying systems rather than to a single deployed product.

2. BIOEMTECH DosimetrEYE in preclinical nuclear medicine

In its primary sense, DosimetrEYE is intended to replace a resource-intensive dosimetric workflow in small-animal SPECT/CT studies. In the conventional setting, accurate dose estimation usually requires full 3D SPECT/CT acquisitions at multiple time points and computationally expensive Monte Carlo (MC) simulations on 3D volumes. DosimetrEYE addresses this by learning a supervised mapping from planar 2D SPECT/CT images to simulated 3D absorbed dose maps, with the stated aim of enabling “non-invasive, real-time dosimetry in small-animal studies, facilitating efficient organ-level dose assessment” (Binda et al., 11 Aug 2025).

The model is built on a Vox2Vox architecture, described as a 3D conditional GAN for voxel-to-voxel translation. Conceptually, the generator learns

$$G:\text{(2D SPECT/CT)} \rightarrow \text{3D dose volume},$$

while a 3D CNN discriminator distinguishes real MC-derived dose maps from generated ones, conditioned on the corresponding 2D input. The paper does not reproduce the full Vox2Vox block structure, but it states that, given the reference design, the system likely uses 3D convolutions with downsampling and upsampling, skip connections, and a PatchGAN-like discriminator (Binda et al., 11 Aug 2025).

The objective is not given as a DosimetrEYE-specific loss, but the underlying Vox2Vox reference is summarized as using a standard conditional GAN objective with adversarial and reconstruction terms:

$$\mathcal{L}_{adv}(G,D)= \mathbb{E}_{x,y}[\log D(x,y)] + \mathbb{E}_{x}[\log(1-D(x,G(x)))]$$

and

$$\mathcal{L}_{L1}(G)=\mathbb{E}_{x,y}\big[\|y-G(x)\|_1\big],$$

with total generator loss

$$\mathcal{L}_G=\mathcal{L}_{adv}(G,D)+\lambda \mathcal{L}_{L1}(G).$$

The emphasis in the paper is on non-invasive, real-time estimation rather than on a customized training metric (Binda et al., 11 Aug 2025).

The operational pipeline is straightforward. First, 2D planar SPECT/CT scans are acquired with eyes™ systems. Second, each scan configuration is paired with a corresponding 3D absorbed dose distribution obtained by MC simulation from 3D SPECT/CT; these serve as supervised targets. Third, the network is trained to associate 2D projection patterns of radioactivity and anatomy with 3D dose distributions. At inference time, the scanner output is passed to DosimetrEYE, which returns a 3D voxel dose map aligned to the animal. Post-processing is only implied, but the text mentions organ-level dose integration and possible resampling or registration within the eyes™ software (Binda et al., 11 Aug 2025).

The intended impact is explicitly translational. DosimetrEYE is described as providing near real-time dosimetry directly from 2D acquisitions, reducing the need for repeated 3D scanning and MC runs, enabling higher-throughput and more standardized dosimetry, and contributing to a >80% reduction in animal sacrifice by extracting more information per animal and reducing repeated terminal procedures (Binda et al., 11 Aug 2025).

3. Hybrid anomaly detection and safeguarding

The most distinctive feature of the published DosimetrEYE system is not only generative dose prediction but its attached hybrid anomaly detection layer. Because the model is trained on a small, expensive dataset, the authors state that “it is imperative to reject out-of-distribution samples before analysis to reduce inaccurate predictions” (Binda et al., 11 Aug 2025).

The safeguard is implemented in Obz AI and combines two unsupervised detectors trained on in-distribution data only. The first branch uses first-order statistical features such as entropy, median, variance, and uniformity. These are concatenated into a feature vector and modeled with a Gaussian Mixture Model (GMM). For a new sample with feature vector $\mathbf{f}$, the framework computes its likelihood $p(\mathbf{f})$ and declares an outlier if that likelihood falls below an empirically chosen percentile threshold:

$$\text{declare outlier if } p(\mathbf{f}) < \text{GMM}^{-1}(\tau).$$

This branch is explicitly described as an interpretable feature-space density estimator (Binda et al., 11 Aug 2025).

The second branch uses CLIP image embeddings followed by PCA. For each embedding $\mathbf{x}_i$, the reconstruction loss after projection onto the top $r$ principal components is used as the anomaly score:

$s_i^{CLIP}=\|\mathbf{x}_i-\hat{\mathbf{x}}_i\|^2,$

with $$\mathcal{L}_{adv}(G,D)= \mathbb{E}_{x,y}[\log D(x,y)] + \mathbb{E}_{x}[\log(1-D(x,G(x)))]$$0. Samples with high reconstruction loss are treated as probable outliers:

$$\mathcal{L}_{adv}(G,D)= \mathbb{E}_{x,y}[\log D(x,y)] + \mathbb{E}_{x}[\log(1-D(x,G(x)))]$$1

The dimensionality $$\mathcal{L}_{adv}(G,D)= \mathbb{E}_{x,y}[\log D(x,y)] + \mathbb{E}_{x}[\log(1-D(x,G(x)))]$$2 is chosen by a parameter sweep so that the proportion of outliers detected in the test dataset matches most closely the proportion found in the training dataset (Binda et al., 11 Aug 2025).

The framework is called hybrid because it merges low-level classical statistics with high-level embedding-based structure. Integration policy is left configurable: a case may be flagged if either detector fires or only if both do. In deployment, flagged samples may be blocked from DosimetrEYE or routed for human review. The practical effect is real-time quality control, reduced manual oversight, and improved traceability through logged features and anomaly scores visualized in the Obz AI dashboard (Binda et al., 11 Aug 2025).

4. Evidence base, workflow, and limitations

The published study is explicit about its evidentiary scope. It is focused on safeguarding and system integration, not on an exhaustive dosimetric benchmark. It does not provide exact counts of mice, scans, or radiotracers; it does not report a table of predictive metrics such as MSE, RMSE, Dice, or SSIM for the generated dose maps; and it does not provide numerical AUROC or AUPRC for outlier detection (Binda et al., 11 Aug 2025).

What is documented is the deployment workflow. The imaging pipeline proceeds from acquisition of 2D SPECT/CT, to OD screening by the FOF-GMM and CLIP-PCA branches, to either immediate 3D dose generation for in-distribution samples or flagging and review for outliers. This operational framing is central to the paper’s claim that the OD layer enables DosimetrEYE to function “as an embedded product feature” rather than a research-only prototype (Binda et al., 11 Aug 2025).

The paper also frames the system in regulatory language. It states that, in industrial applications of generative AI for preclinical imaging and dosimetry, outlier detection is essential for product reliability, scalability, and regulatory readiness. Traceability, quality control, failure logging, and clear handling of unexpected inputs are presented as prerequisites for robust deployment, even though no specific framework such as CE marking or FDA clearance is named (Binda et al., 11 Aug 2025).

The limitations are equally important. Training data are small because MC-based 3D dosimetry is labor- and capital-intensive. The OD layer operates on inputs rather than on model-internal uncertainty, so Bayesian uncertainty estimation or ensembles are not part of the current design. No formal ablation isolates the contributions of FOF-GMM and CLIP-PCA. This suggests that the present publication establishes a deployment architecture and a safety narrative more strongly than it establishes a full quantitative dosimetric validation (Binda et al., 11 Aug 2025).

5. Ocular brachytherapy and eye-lens dosimetry

In a second strand of the literature, “DosimetrEYE” denotes a prospective ocular dosimetry ecosystem rather than the BIOEMTECH GAN. The clearest example is the CLRP EPv2 eye plaque dosimetry database, which is presented as a foundation for a fully model-based, plaque-specific ocular brachytherapy planning and research platform (Safigholi et al., 2021). EPv2 contains 17 plaque models, three radionuclides, and benchmarked 3D dose distributions on a $$\mathcal{L}_{adv}(G,D)= \mathbb{E}_{x,y}[\log D(x,y)] + \mathbb{E}_{x}[\log(1-D(x,G(x)))]$$3 grid with $$\mathcal{L}_{adv}(G,D)= \mathbb{E}_{x,y}[\log D(x,y)] + \mathbb{E}_{x}[\log(1-D(x,G(x)))]$$4 voxels. It distinguishes HOMO, HETERO, and HETsi scenarios, enabling both TG-43-like reference calculations and heterogeneous plaque-in-water modeling, with consistency to prior EGSnrc results to $$\mathcal{L}_{adv}(G,D)= \mathbb{E}_{x,y}[\log D(x,y)] + \mathbb{E}_{x}[\log(1-D(x,G(x)))]$$5 and to MCNP to $$\mathcal{L}_{adv}(G,D)= \mathbb{E}_{x,y}[\log D(x,y)] + \mathbb{E}_{x}[\log(1-D(x,G(x)))]$$6 (Safigholi et al., 2021). In this usage, DosimetrEYE is a model-based ocular brachytherapy planning concept.

The same ocular line extends to nanoparticle-enhanced plaque therapy. In the MCNP5 study of $$\mathcal{L}_{adv}(G,D)= \mathbb{E}_{x,y}[\log D(x,y)] + \mathbb{E}_{x}[\log(1-D(x,G(x)))]$$7 and $$\mathcal{L}_{adv}(G,D)= \mathbb{E}_{x,y}[\log D(x,y)] + \mathbb{E}_{x}[\log(1-D(x,G(x)))]$$8 plaques in the presence of 50 nm gold nanoparticles, a 16 mm COMS plaque with 13 seeds is used to show that tumor DEF rises with GNP concentration and is consistently higher for $$\mathcal{L}_{adv}(G,D)= \mathbb{E}_{x,y}[\log D(x,y)] + \mathbb{E}_{x}[\log(1-D(x,G(x)))]$$9 than for $$\mathcal{L}_{L1}(G)=\mathbb{E}_{x,y}\big[\|y-G(x)\|_1\big],$$0; at 30 mg/g in water, the tumor-apex DEF is about 4.91 for $$\mathcal{L}_{L1}(G)=\mathbb{E}_{x,y}\big[\|y-G(x)\|_1\big],$$1 and 3.66 for $$\mathcal{L}_{L1}(G)=\mathbb{E}_{x,y}\big[\|y-G(x)\|_1\big],$$2 (Asadi et al., 2015). The paper further argues that realistic eye modeling becomes increasingly important when GNPs are used, particularly for $$\mathcal{L}_{L1}(G)=\mathbb{E}_{x,y}\big[\|y-G(x)\|_1\big],$$3.

A separate eye-focused branch concerns active eye-lens dosimetry. "Active Personal Eye Lens Dosimetry with the Hybrid Pixelated Dosepix Detector" measures

$$\mathcal{L}_{L1}(G)=\mathbb{E}_{x,y}\big[\|y-G(x)\|_1\big],$$4

with a Dosepix-based prototype and reports normalized-response compliance with IEC 61526 for continuous reference fields between mean photon energies of 12.4 keV and 248 keV and for incidence angles between $$\mathcal{L}_{L1}(G)=\mathbb{E}_{x,y}\big[\|y-G(x)\|_1\big],$$5 and $$\mathcal{L}_{L1}(G)=\mathbb{E}_{x,y}\big[\|y-G(x)\|_1\big],$$6 (Beißer et al., 2023). In "Active Eye Lens Dosimetry With Dosepix: Influence of Measurement Position and Lead Glass Shielding," side-of-head measurements show no significant influence on the resulting $$\mathcal{L}_{L1}(G)=\mathbb{E}_{x,y}\big[\|y-G(x)\|_1\big],$$7 compared with measurements directly in front of the eye in the tested homogeneous fields, whereas lead glass pieces can fail to reproduce the behavior of real radiation safety glasses and can permit bypass paths at specific angles (Ullmann et al., 9 May 2025). In this ocular usage, DosimetrEYE denotes a head-worn, spectrometric, active eye-lens dosemeter concept.

6. Real-time radiotherapy QA, in-vivo tracking, and beam monitoring

The broader literature also uses the name for real-time dosimetry and tracking systems in radiotherapy. One example is the EM-tracked plastic scintillation dosimeter for HDR brachytherapy: with the 5DOFthin Aurora sensor, the reported jitter error is $$\mathcal{L}_{L1}(G)=\mathbb{E}_{x,y}\big[\|y-G(x)\|_1\big],$$8 mm, reproducibility is $$\mathcal{L}_{L1}(G)=\mathbb{E}_{x,y}\big[\|y-G(x)\|_1\big],$$9 mm, average positional uncertainty is less than 0.2 mm in the optimal operating range, angular error is at most $$\mathcal{L}_G=\mathcal{L}_{adv}(G,D)+\lambda \mathcal{L}_{L1}(G).$$0, and the resulting positional contribution to dose uncertainty is less than 5% at 10 mm from a $$\mathcal{L}_G=\mathcal{L}_{adv}(G,D)+\lambda \mathcal{L}_{L1}(G).$$1 source (Tho et al., 2018). Closely related multi-point plastic scintillator systems, including IViST and the three-point mPSD platform, add real-time source triangulation, 100 kHz acquisition, dose deviations below 5%, positional accuracy within 1 mm up to 6 cm in phantom, and dwell-time verification with sub-second deviations (Rosales et al., 2020, Rosales et al., 2019).

Personal and area dosimetry provide another branch. PDOZ uses Geant4, an ICRU sphere, scintillators, SiPMs, and conversion curves from fluence to $$\mathcal{L}_G=\mathcal{L}_{adv}(G,D)+\lambda \mathcal{L}_{L1}(G).$$2 to build a personal electronic dosimeter for beta/electrons and gamma rays (Salvi et al., 2023). Dosepix in direct pulsed photon fields shows a normalized $$\mathcal{L}_G=\mathcal{L}_{adv}(G,D)+\lambda \mathcal{L}_{L1}(G).$$3 response close to 1.0 over pulse durations from 3.6 s to 2 ms and, for small pixels, remains within PTB-A 23.2 limits up to about 704 Sv/h in RQR8 beams (Haag et al., 2021). These results belong to distinct devices, but they show how the DosimetrEYE label is used to frame active, spectrometric, and real-time radiation protection instrumentation.

Beam monitoring papers extend the idea still further. A diamond detector with gated integrator electronics performs per-pulse LINAC monitoring with a 20 $$\mathcal{L}_G=\mathcal{L}_{adv}(G,D)+\lambda \mathcal{L}_{L1}(G).$$4s sampling period and measured dose per pulse values of 278 $$\mathcal{L}_G=\mathcal{L}_{adv}(G,D)+\lambda \mathcal{L}_{L1}(G).$$5Gy at 6 MV and 556 $$\mathcal{L}_G=\mathcal{L}_{adv}(G,D)+\lambda \mathcal{L}_{L1}(G).$$6Gy at 18 MV (Pettinato et al., 2021). Medipix3 in a 60 MeV ocular proton therapy beamline provides millisecond time resolution, count-rate linearity across the tested current range, and beam monitoring up to $$\mathcal{L}_G=\mathcal{L}_{adv}(G,D)+\lambda \mathcal{L}_{L1}(G).$$7 protons/s (Yap et al., 2021). ImageDosis uses a YAG:Ce scintillating coating and a high-speed camera for UHDR electron FLASH, showing linear response above 3.5 Gy up to at least 13 Gy, no dose-rate dependence up to 140 Gy/s, pulse-by-pulse discrimination up to 250 Hz, and about 3% missed pulses because some occurred during camera dead time (Vanreusel et al., 2023). A deformable scintillation dosimeter adds a different capability: simultaneous dose and DVF measurement with 0.3 mm tracking precision, DVF differences below 1.5 mm relative to Plastimatch, and dose within 1% of TPS and Hyperscint under fixed and deformed conditions (Cloutier et al., 2021).

Taken together, these systems do not constitute a single product, but they define a recognizable research pattern around the name DosimetrEYE: dosimetry coupled to model-based inference, spectrometric reconstruction, direct tracking, pulse-resolved acquisition, or deformation-aware measurement. The most established realization remains the BIOEMTECH preclinical GAN with hybrid anomaly detection (Binda et al., 11 Aug 2025); the broader usages indicate a convergent design vocabulary for high-fidelity, real-time, and deployment-oriented dosimetry across nuclear medicine, brachytherapy, radiation protection, proton therapy, and FLASH radiotherapy.