Predicting F-18 FMISO PET Hypoxia Measurements From F-18 FDG PET Scan Using a Generative Adversarial Network
W Zhao*, M Grkovski, N Lee, H Schoder, J Humm, H Veeraraghavan, J Deasy, Memorial Sloan Kettering Cancer Center, New York, NY
Presentations
PO-GePV-I-82 (Sunday, 7/10/2022) [Eastern Time (GMT-4)]
ePoster Forums
Purpose: Tumor hypoxia is an import factor contributing to poor prognosis and radiotherapy treatment response, thus could be used as an input for personalizing radiotherapy dose. However, ¹⁸F-FMISO hypoxia scans are much less available than ¹⁸F-FDG PET scans in clinics. We therefore were motivated to use AI/machine learning methods to predict pixel-wise hypoxia measurements derived from ¹⁸F-FMISO called tumor-to-blood ratio (TBR) by using a conditional generative adversarial network (GAN) with commonly available ¹⁸F-FDG PET scans.
Methods: Ninety-seven head and neck cancer patients treated with chemoradiotherapy were included. Each patient received pre-therapy ¹⁸F-FDG PET and dynamic ¹⁸F-FMISO PET scans, all of which were rigidly registered. Image patches (32 x 32) enclosing treatment targets contained in PET slices (75 primary tumors and 101 lymph nodes) were analyzed. Dynamic ¹⁸F-FMISO was used to derive k1 and TBR images representing perfusion and hypoxia, respectively. Patient-wise splitting was done to provide 70% training, 10% validation and 20% testing patches. Two pix2pix GAN networks were constructed to predict TBR using: model A) FDG and k1, and model B) FDG as inputs. During the training, the best model was selected based on the smallest mean absolute error (MAE) between predicted and real TBR for validation targets. The generalization performance of this selected model was evaluated with testing targets, using normalized MAE (=MAE/mean of real target TBR) and Spearman correlation of predicted and real TBRᵐᵃˣ.
Results: The normalized MAE for testing targets was 15.7±7.8% for model A and 16.2±8.2% for model B. Moderate correlation of TBRᵐᵃˣ was obtained (A: 0.596, B: 0.539). Using clinically applied TBRᵐᵃˣ>1.2 as hypoxia criterion, model A and B resulted in test AUC of 0.937 and 0.916, respectively.
Conclusion: The trained pix2pix models produced promising test AUC (>0.91) for hypoxia classification, demonstrating its potential to predict ¹⁸F-FMISO derived TBR from ¹⁸F-FDG scans.
Funding Support, Disclosures, and Conflict of Interest: Nancy Lee is on the following advisory boards: Merck Merck EMD Elsie Mirati
Keywords
Hypoxia, FDG PET, Image Analysis
Taxonomy
IM- PET : Machine learning, computer vision
Contact Email