Artificial Intelligence for Simplified Patient-Centered Dosimetry in Radiopharmaceutical Therapies

Basic Information

• Paper ID: 2510.12714
• Title: Artificial intelligence for simplified patient-centered dosimetry in radiopharmaceutical therapies
• Authors: Alejandro Lopez-Montes, Fereshteh Yousefirizi, Yizhou Chen, Yazdan Salimi, Robert Seifert, Ali Afshar-Oromieh, Carlos Uribe, Axel Rominger, Habib Zaidi, Arman Rahmim, Kuangyu Shi
• Classification: physics.med-ph cs.AI physics.app-ph
• Published Journal: PET Clinics 2025 (Accepted)
• Paper Link: https://arxiv.org/abs/2510.12714

Abstract

This paper presents a comprehensive review of artificial intelligence applications in radiopharmaceutical therapy (RPT) dosimetry, with emphasis on simplified patient-centered dosimetry approaches. With the rapid advancement of RPT, there is growing demand for personalized and patient-centric dosimetry. Artificial intelligence provides solutions to key limitations in current dosimetry calculations. The paper reviews major advances in AI-assisted simplified dosimetry and discusses future directions for AI in RPT dosimetry.

Research Background and Motivation

Problem Definition

1. Complexity of Traditional Dosimetry: Dosimetry calculations in radiopharmaceutical therapy involve complex time-integrated activity (TIA) estimation and Monte Carlo (MC) simulations, which are time-consuming and computationally expensive
2. Personalized Treatment Requirements: Different patients possess unique anatomical structures and radiation response characteristics, yet current RPT commonly employs fixed-dose regimens lacking individualization
3. Clinical Implementation Barriers: Traditional dosimetry methods require multi-timepoint SPECT/CT scans and complex calculations, limiting widespread clinical application of personalized dosimetry

Research Significance

• RPT has experienced significant growth over the past decade, with multiple radiopharmaceuticals including 177Lu-PSMA-617 and 177Lu-DOTATATE receiving FDA/EMA approval
• Emerging alpha-emitter therapeutics such as 225Ac-PSMA demonstrate higher energy deposition and stopping power
• Personalized dosimetry enables optimization of tumor dose while minimizing exposure to healthy organs

Limitations of Existing Methods

1. Heavy Imaging Burden: Multiple timepoint SPECT/CT scans required, increasing patient burden and hospital costs
2. Computational Complexity: While MC simulation is the gold standard, it is computationally time-consuming, with patient-specific simulations being even more complex
3. Simplified Assumptions: Tabulated S-values based on standard anthropomorphic phantoms ignore individual anatomical variations
4. Segmentation Challenges: Accuracy of lesion segmentation directly impacts dosimetry calculation quality

Core Contributions

1. Systematic Review: First comprehensive review of AI applications in simplifying RPT dosimetry
2. Classification Framework: Categorizes AI applications into four major classes: simplified imaging protocols, simplified dose calculation, lesion segmentation, and utilization of prior information
3. Technical Analysis: In-depth analysis of technical principles, advantages, and limitations of various AI methods
4. Future Perspectives: Proposes cutting-edge concepts such as digital twins for application in personalized RPT
5. Clinical Guidance: Provides practical guidelines for clinicians and researchers on AI-assisted dosimetry

Methodology Details

Task Definition

This review focuses on how AI simplifies four key aspects of RPT dosimetry:

• Input: Medical imaging (SPECT/CT, PET/CT), patient clinical information
• Output: Simplified dosimetry calculation workflow and more accurate dose estimation
• Objective: Achieve patient-friendly personalized dosimetry

Core Technical Framework

1. Simplified Imaging Protocols

Shortened SPECT Acquisition Time:

• Sparse-View SPECT: CNN synthesis of complete 120-view data from 30 projections
• Reduced Exposure Time: Deep learning recovery of low-count SPECT image quality
• Technical Implementation: Wikberg et al. proposed reduction from standard 120 views to 30 views, with CNN synthesis of 90 intermediate projections, achieving SSIM > 0.9

Single-Timepoint Dosimetry:

• Traditional Approach: Multiple timepoint scans required to fit time-activity curves (TAC)
• AI Methods:
  • Data-driven nonparametric generalized additive models
  • Machine learning prediction of effective half-life
  • Reduced sensitivity to timepoint selection

2. Simplified Dose Calculation

AI as Monte Carlo Simulation Alternative:

Traditional Method: TIA → MC Simulation → Dose Distribution
AI Method: TIA + CT → CNN/UNET → Dose Distribution

Key Technologies:

• UNET Architecture: Combines CT anatomical information and TIA maps to predict dose distribution
• Multi-Kernel Approach: Uses AI to correct boundary errors in nuclear convolution methods
• Density-Aware Processing: Integrates CT density information into neural networks to improve prediction accuracy

3. Lesion Segmentation

Automated Total Metabolic Tumor Volume (TMTV) Segmentation:

• TMTV-Net: Fully automated metabolic tumor volume segmentation for lymphoma PET/CT
• Transfer Learning: Adaptation from PET pre-trained models to SPECT applications
• Multi-Center Validation: Generalization capability across different institutions and radiopharmaceuticals

SPECT-Specific Challenges:

• Blurred boundary handling: Employs fuzzy C-means clustering
• Lesion appearance variability adaptation
• Cross-radiopharmaceutical generalization capability

4. Utilization of Prior Information

Pre-Treatment to Post-Treatment Prediction:

• Treatment Pairing: Leverages 68Ga-DOTATOC PET to predict 177Lu-DOTATATE dosimetry
• Machine Learning Integration: Combines radiomics and dosimetry features
• Voxel-Level Prediction: Enables characterization of intra-organ heterogeneity

Digital Twin Technology:

• Personalized Computational Models: Integrates baseline imaging, clinical data, and prior dosimetry information
• PBPK Modeling: Physiology-driven pharmacokinetic models
• MBAM Optimization: Manifold boundary approximation method for model simplification

Experimental Setup

Dataset Characteristics

• Multi-Modal Imaging: SPECT/CT, PET/CT, planar scintigraphy
• Multiple Radiopharmaceuticals: 177Lu-PSMA, 177Lu-DOTATATE, 90Y-SIRT, etc.
• Multi-Center Data: Validation of method generalization capability
• Clinical Parameters: Patient age, weight, renal function, etc.

Evaluation Metrics

• Image Quality: SSIM (Structural Similarity Index) > 0.9
• Dose Prediction Accuracy: Mean Absolute Error (MAE) < 1.6 Gy
• Segmentation Accuracy: Dice coefficient, sensitivity, specificity
• Time Efficiency: Reduction in computation time compared to traditional methods

Technical Implementation

• Deep Learning Frameworks: CNN, UNET, ResNet, Transformer
• Optimization Strategies: Adam optimizer, learning rate scheduling
• Data Augmentation: Rotation, scaling, noise addition
• Cross-Validation: K-fold cross-validation ensures result reliability

Experimental Results

Main Results

1. Imaging Protocol Simplification Effects

• Sparse-View SPECT: Reduction from 120 to 30 views with image quality SSIM > 0.9
• Fast SPECT: Reduction from 20-minute standard protocol to 3 minutes while maintaining diagnostic quality
• Single-Timepoint Dosimetry: Significant MAE reduction compared to multi-timepoint methods, particularly for day-0 imaging

2. Dose Calculation Accuracy

• AI vs MC: CNN methods show high consistency with Monte Carlo simulation with significantly reduced computation time
• Organ Dose Prediction: Liver, spleen, and kidney dose prediction MAE < 1.6 Gy
• Tumor Dose Prediction: R² = 0.64, MAE = 0.73 Gy/GBq

3. Segmentation Performance

• TMTV Segmentation: High consistency with expert annotations, Dice coefficient > 0.85
• Multi-Organ Segmentation: Fully automated segmentation of 104 anatomical structures
• Cross-Modality Generalization: PET-trained models successfully adapted to SPECT applications

Ablation Studies

• Importance of Density Information: Addition of CT density information significantly improves dose prediction accuracy
• Multi-Timepoint vs Single-Timepoint: Single-timepoint methods achieve multi-timepoint accuracy under specific conditions
• Network Architecture Comparison: UNET demonstrates superior performance in medical image segmentation tasks

Case Studies

177Lu-PSMA Treatment Case:

• Pre-treatment 68Ga-PSMA PET successfully predicts post-treatment organ doses
• AI-segmented lesion volume significantly correlates with treatment response
• Personalized dose adjustment improves treatment efficacy

Related Work

Traditional Dosimetry Methods

• MIRD Scheme: Standard methodology established by the Medical Internal Radiation Dose Committee
• Monte Carlo Simulation: Software packages including GATE and GEANT4
• Simplified Methods: Hänscheid method, Madsen method, and other single-timepoint estimation approaches

AI Applications in Medical Imaging

• Image Reconstruction: Deep learning acceleration of CT/MRI reconstruction
• Image Segmentation: Universal segmentation frameworks such as nnU-Net
• Image Synthesis: GAN applications in medical image generation

Personalized Medicine

• Precision Medicine: Genomics-based personalized treatment
• Digital Twins: Applications in cardiovascular and oncology domains
• PBPK Modeling: Personalized pharmacokinetic prediction

Conclusions and Discussion

Main Conclusions

1. AI Significantly Simplifies Dosimetry Workflow: Reduces imaging frequency, shortens scan time, and accelerates computational processes
2. Maintains or Improves Accuracy: AI methods maintain dosimetry calculation accuracy while simplifying procedures
3. Clinical Practicality: Patient-friendly approaches enhance clinical feasibility of personalized dosimetry
4. Technical Maturity: Most AI technologies possess technical foundation for clinical translation

Limitations

1. Data Dependency: AI models require large annotated datasets; multi-center data sharing remains challenging
2. Generalization Capability: Cross-institutional, cross-device, and cross-radiopharmaceutical generalization requires further improvement
3. Regulatory Challenges: Complex regulatory approval processes for AI medical devices
4. Clinical Validation: Large-scale prospective clinical trial validation remains insufficient

Future Directions

1. Digital Twin Development: Personalized treatment models integrating multi-modal data
2. Federated Learning: Multi-center collaboration while preserving privacy
3. Real-Time Dosimetry: Dynamic dose monitoring and adjustment during treatment
4. Multi-Modal Fusion: Comprehensive prediction models combining imaging, genomics, and proteomics

In-Depth Evaluation

Strengths

1. Comprehensiveness: Systematically covers AI applications across all aspects of RPT dosimetry
2. Practicality: Emphasizes clinically feasible simplified methods
3. Forward-Looking: Proposes cutting-edge concepts such as digital twins
4. Technical Depth: Detailed analysis of principles and implementation of various AI techniques
5. Clinical Orientation: Consistently targets patient-friendly and clinically practical objectives

Limitations

1. Lack of Quantitative Comparison: Direct performance comparison between different AI methods is absent
2. Limited Clinical Validation: Most methods remain in research phase with insufficient clinical validation data
3. Cost-Benefit Analysis: Lacks economic evaluation of AI method implementation
4. Standardization Issues: Insufficient discussion of standardization and quality control measures for AI dosimetry methods

Impact

1. Academic Value: Provides important reference for AI applications in nuclear medicine
2. Clinical Guidance: Offers guidelines for clinicians in selecting AI-assisted dosimetry methods
3. Industry Promotion: Facilitates development of AI medical devices in nuclear medicine
4. Policy Implications: Provides scientific evidence for regulatory agencies in policy formulation

Applicable Scenarios

1. Large Medical Centers: Tertiary hospitals with AI technology implementation capability
2. Research Institutions: Nuclear medicine and medical physics research teams
3. Medical Device Enterprises: Development and validation of AI medical products
4. Regulatory Agencies: Approval and regulation of AI medical devices

References

This paper cites 159 relevant references, covering:

• Standard dosimetry methods from the MIRD Committee
• Deep learning applications in medical imaging
• Clinical research in radiopharmaceutical therapy
• Latest advances in AI for nuclear medicine
• Theoretical foundations of digital twins and PBPK modeling

Overall Assessment: This is a high-quality review paper that systematically summarizes the current state and future prospects of AI applications in radiopharmaceutical therapy dosimetry. The paper features clear structure, in-depth technical analysis, and significant value in promoting AI applications in nuclear medicine. Further strengthening of quantitative method comparison and clinical validation research is recommended.