# Artificial Intelligence and Quantitative Nuclear Medicine

## Introduction

Artificial intelligence (AI) and machine learning are transforming nuclear medicine through improved image reconstruction, automated quantification, lesion detection, outcome prediction, and workflow optimization. Simultaneously, advances in quantitative nuclear medicine enable precise measurement of tracer uptake, dosimetry, and treatment response. Together, these technologies are driving a shift toward precision nuclear medicine with reproducible, data-driven clinical decisions.

## Foundations of AI in Medical Imaging

### Machine Learning Concepts

Supervised learning trains algorithms on labeled data -- for example, images paired with known diagnoses. Unsupervised learning discovers patterns without labels, such as clustering similar scan findings. Deep learning uses neural networks with multiple layers and is the dominant approach for image analysis. Convolutional neural networks (CNNs) are specifically designed for image feature extraction and classification. Transfer learning allows pre-trained networks to be fine-tuned on nuclear medicine datasets, reducing the need for massive training collections.

### Key AI Tasks in Nuclear Medicine

The primary AI tasks span the nuclear medicine workflow: image reconstruction (noise reduction and artifact correction), segmentation (automated tumor and organ delineation), classification (disease detection and characterization), quantification (automated SUV, volume, and dosimetry calculations), and prediction (prognosis and treatment response forecasting).

![Overview of AI applications across the nuclear medicine workflow](images/ai-nuclear-medicine-workflow.png)

## AI in Image Reconstruction and Enhancement

### Deep Learning-Based Reconstruction

Deep learning denoising enables diagnostic-quality PET images from substantially reduced administered activity -- up to 50-75% dose reduction. Commercial implementations such as SubtlePET and GE TrueFidelity are already in clinical use. These algorithms maintain quantitative accuracy (preserving SUV values) while dramatically reducing image noise. The practical result is the potential to reduce scan time or administered activity without sacrificing image quality.

### Attenuation Correction

Deep learning generates pseudo-CT maps from MRI data for PET/MRI attenuation correction and can produce synthetic CT from non-attenuation-corrected PET data, enabling CT-free PET workflows. These approaches improve bone estimation in MR-based attenuation correction and reduce artifacts from metal implants and truncation.

### Motion Correction

AI-driven respiratory and cardiac motion compensation uses data-driven gating without requiring external devices. This improves lesion detectability and quantitative accuracy, particularly for thoracic and abdominal PET imaging.

## Automated Quantification

### SUV Measurement

The standardized uptake value (SUV) is the most widely used PET quantification metric. Different sampling strategies -- SUVmax, SUVmean, and SUVpeak -- offer varying degrees of reproducibility. AI-automated ROI placement reduces inter-observer variability that plagues manual measurements. Harmonization of SUV across different scanners and reconstruction methods remains an active challenge.

| Quantitative Metric | Definition | Clinical Application | AI Enhancement |
|---|---|---|---|
| SUVmax | Maximum voxel uptake value | Lesion characterization; response | Automated ROI placement |
| SUVpeak | Mean SUV in 1 cm3 sphere at hottest region | More reproducible than SUVmax | Automated sphere placement |
| MTV | Total volume above SUV threshold | Prognostic marker; tumor burden | Automated whole-body segmentation |
| TLG | MTV x SUVmean | Prognostic (stronger than SUVmax) | Automated calculation |
| Ki (metabolic rate) | Patlak or kinetic modeling | Gold-standard metabolism measure | Dynamic PET analysis |
| BSI (Bone Scan Index) | % skeleton with metastatic disease | Prognosis in prostate/breast cancer | FDA-cleared automated tool |

### Volumetric Metrics

Metabolic tumor volume (MTV) is the total volume of FDG-avid disease above a chosen threshold, and total lesion glycolysis (TLG) is MTV multiplied by SUVmean. AI enables automated whole-body tumor burden quantification that would be prohibitively time-consuming to perform manually. Importantly, MTV and TLG are stronger prognostic markers than SUVmax in many malignancies, including lymphoma and lung cancer.

### Bone Scan Quantification

The Bone Scan Index (BSI) quantifies the percentage of skeleton involved by metastatic disease. FDA-cleared applications such as EXINI bone calculate BSI automatically from conventional bone scintigraphy. BSI has demonstrated prognostic value in metastatic prostate and breast cancer and provides objective treatment response monitoring.

## AI for Lesion Detection and Classification

### PET/CT Applications

AI systems automate detection of FDG-avid lymph nodes and metastases, assist in detection and classification of prostate cancer lesions on PSMA PET, identify and quantify NET metastases on Ga-68 DOTATATE PET, and characterize lung nodules on the CT component of PET/CT.

### SPECT Applications

In myocardial perfusion SPECT, AI performs automated segmentation, quantification, and risk prediction. It calculates ejection fraction from gated SPECT, predicts obstructive coronary artery disease from perfusion data, and can assist with thyroid scan interpretation.

### Performance Metrics

AI systems must be evaluated by sensitivity, specificity, accuracy, and area under the ROC curve. Comparison against expert reader performance establishes non-inferiority or superiority. External validation on independent datasets is essential because single-center performance may not generalize. Clinical utility assessment should go beyond pure diagnostic accuracy to consider impact on management decisions.

![AI-automated whole-body tumor segmentation on FDG PET/CT showing metabolic tumor volume quantification](images/ai-tumor-segmentation-pet.png)

## Dosimetry and Theranostics

### AI-Enhanced Dosimetry

Automated organ and tumor segmentation forms the basis of absorbed dose calculations. Voxel-level dosimetry from serial post-therapy SPECT/CT or PET/CT enables precise measurement of dose delivery. AI can predict treatment response based on pre-therapy dosimetric estimates and guide personalized activity prescription, moving beyond empiric fixed-dose regimens.

### Theranostic Applications

Radiomics feature extraction from pre-therapy PET can predict PRRT response. Digital twin models simulate various treatment scenarios. AI-guided treatment planning optimizes Lu-177 DOTATATE and Lu-177 PSMA therapy delivery. Toxicity prediction (nephrotoxicity, hematotoxicity) integrates dosimetric and clinical data.

## Radiomics and Radiogenomics

### Radiomics

Radiomics extracts high-dimensional quantitative features from medical images -- including texture, shape, intensity, and wavelet features -- that go beyond what the human eye can assess. Radiomics signatures have been shown to correlate with prognosis, molecular subtypes, and treatment response. Standardization through IBSI (Image Biomarker Standardisation Initiative) guidelines is essential for reproducibility across institutions.

### Radiogenomics

Radiogenomics correlates imaging features with genomic and molecular data, offering the potential for non-invasive prediction of tumor molecular profiles from PET/CT features. This could guide targeted therapy without tissue biopsy, though the field remains in early stages and clinical translation is limited.

## Implementation Challenges

Successful clinical deployment of AI in nuclear medicine requires addressing several challenges. Large, diverse, well-annotated datasets are needed for training. Models trained on single-center data may not generalize (the domain shift problem). FDA clearance is required for clinical decision support tools. The "black box" concern demands explainable, interpretable AI outputs for clinician trust. Integration into PACS, reporting systems, and clinical pathways remains technically complex. Training data must represent diverse patient populations to avoid algorithmic bias.

## Ethical and Regulatory Considerations

AI should function as a decision support tool rather than an autonomous decision maker. Physicians retain responsibility for final interpretation and clinical decisions. Patient privacy and data security in AI training pipelines must be rigorously maintained. Continuous monitoring for model drift and performance degradation is necessary after deployment. Transparent reporting of AI tool limitations to clinical users is essential for safe implementation.

![Schematic of radiomics feature extraction pipeline from PET/CT images](images/radiomics-pipeline-pet.png)

## Clinical Pearls

Deep learning-based PET reconstruction can maintain diagnostic image quality with 50-75% dose reduction, offering significant benefit for pediatric patients and those requiring serial imaging. Metabolic tumor volume and total lesion glycolysis, now quantifiable through AI-automated segmentation, are stronger prognostic markers than SUVmax in many malignancies. AI-enhanced voxel-level dosimetry enables personalized activity prescription for theranostic applications, moving beyond empiric fixed-dose regimens. Radiomics and AI models require rigorous external validation and standardization before clinical implementation, as single-center performance may not generalize across diverse populations.

## References

1. Bradshaw TJ, et al. "Nuclear Medicine and Artificial Intelligence: Best Practices for Algorithm Development." *J Nucl Med*. 2022;63(4):500-510.
2. Zaharchuk G, et al. "Deep Learning in Neuroradiology and Nuclear Medicine." *Radiology*. 2018;288(2):318-328.
3. Uribe CF, et al. "Machine Learning in Nuclear Medicine: Part 1 — Introduction." *J Nucl Med*. 2019;60(4):451-458.
4. Zwanenburg A, et al. "The Image Biomarker Standardisation Initiative: Standardised Quantitative Radiomics for High-Throughput Image-Based Phenotyping." *Radiology*. 2020;295(2):328-338.