Predicting the hypoxic volume of head and neck tumors from fluorodeoxyglucose positron emission tomography images
Wei Zhao1, Milan Grkovski1, Heiko Schoder2
1Department of Medical Physics, Memorial Sloan Kettering Cancer Center New York, 321 East 61st Street, NY 10065, USA.
Background And Purpose:
Tumor hypoxia is linked to lower local control rates and increased distant disease progression during head and neck (HN) radiotherapy. 18F-fluoromisonidazole (18F-FMISO) positron emission tomography (PET) imaging measured hypoxia can aid dose selection for HN patients, but its availability is limited. Hence, we tested the hypothesis that an artificial intelligence (AI) model could synthesize 18F-FMISO-like images from routinely acquired 18F-fluorodeoxyglucose (18F-FDG) PET images in order to predict primary tumor or metastatic lymph node hypoxic volumes.
Materials And Methods:
One hundred and thirty-four (training = 84, validation = 13, testing = 21, additional testing = 16) HN carcinoma patients, treated with chemoradiotherapy between 2011 and 2018 and scanned at treatment baseline with 18F-FDG PET/computed tomography (CT) and 18F-FMISO dynamic PET/CT, were analyzed. A pix2pix-architecture-based generative adversarial network was trained to yield 2D voxel-wise FMISO hypoxia images of target-to-blood ratios (TBRs) directly from the 18F-FDG PET/CT image slices. The hypoxic volume was defined consistent with clinical procedure as the malignant volume with TBR values above 1.2. The AI model hypoxia predictions were compared against scaled 18F-FDG PET values.
Results:
The AI model hypoxic volume predictions were well-correlated with 18F-FMISO hypoxic volumes on the held-out test subjects (Pearson correlation testing R = 0.96, additional testing R = 0.91, p < 0.001). Predictions from globally scaled 18F-FDG PET images also produced a significantly correlated but worse prediction.
Conclusion:
Voxel-wise prediction of hypoxia for HN cancers from a 2D deep learning model using FDG-PET images as inputs was shown to be feasible. Testing on larger institutional and multi-institutional cohorts is required to establish generalizability.
