JoVE - Journal of Visualized Experiments
JoVE Visualize
Contact Us

Predicting the hypoxic volume of head and neck tumors from fluorodeoxyglucose positron emission tomography images using artificial intelligence

  • 0Department of Medical Physics, Memorial Sloan Kettering Cancer Center New York, 321 East 61st Street, NY 10065, USA.
Physics and Imaging in Radiation Oncology +

|

June 30, 2025

|

June 30, 2025

View abstract on PubMed

Summary

This summary is machine-generated.

Artificial intelligence can create 18F-FMISO-like PET images from 18F-FDG PET scans to predict tumor hypoxia in head and neck cancer patients. This AI model shows strong correlation with actual hypoxia measurements, improving treatment planning.

Area Of Science

  • Radiology and Nuclear Medicine
  • Artificial Intelligence in Medical Imaging
  • Oncology

Background

  • Tumor hypoxia is a key factor in head and neck (HN) cancer treatment resistance.
  • 18F-fluoromisonidazole (18F-FMISO) PET imaging quantifies hypoxia but has limited availability.
  • Predicting hypoxic volumes is crucial for optimizing radiotherapy dose selection.

Purpose Of The Study

  • To develop and validate an artificial intelligence (AI) model for synthesizing 18F-FMISO-like PET images from routine 18F-fluorodeoxyglucose (18F-FDG) PET scans.
  • To predict primary tumor and metastatic lymph node hypoxic volumes in HN cancer patients using AI-generated images.
  • To assess the feasibility of using AI to overcome the limitations of 18F-FMISO availability for hypoxia assessment.

Main Methods

  • A generative adversarial network (pix2pix architecture) was trained on 134 HN carcinoma patients' baseline 18F-FDG PET/CT and 18F-FMISO PET/CT scans.
  • The AI model generated 2D voxel-wise 18F-FMISO-like hypoxia images from 18F-FDG PET/CT slices.
  • Hypoxic volume was defined as malignant tissue with a target-to-blood ratio (TBR) > 1.2, and AI predictions were compared against 18F-FMISO measurements.

Main Results

  • The AI model's hypoxic volume predictions showed strong correlation with 18F-FMISO measurements in held-out test subjects (Pearson R = 0.96 and R = 0.91).
  • AI predictions demonstrated significantly better correlation compared to predictions derived from globally scaled 18F-FDG PET values.
  • The study confirmed the feasibility of voxel-wise hypoxia prediction using a 2D deep learning model with 18F-FDG PET as input.

Conclusions

  • AI-driven synthesis of 18F-FMISO-like PET images from 18F-FDG PET is a feasible approach for predicting tumor hypoxia in head and neck cancers.
  • This AI method holds promise for improving radiotherapy planning by providing accessible hypoxia information.
  • Further validation on larger, multi-institutional cohorts is necessary to establish the generalizability of the AI model.

Keywords: