Abstract
699
Objectives: For glioblastoma patients, in vivo hypoxia imaging provides useful clinical information to evaluate tumor malignancy and to predict patients’ prognosis. However, hypoxia imaging tracers for PET, such as 18F-fluoromisonidazole (FMISO), are available in limited institutions; therefore, it would be greatly beneficial if hypoxia was predictable using common imaging modalities. CNN is one of the deep learning techniques for image recognition. Unlike conventional machine learning techniques, CNN trains itself using existing data without requirement of human-made feature values. Therefore, CNN has potential to discover unknown patterns of MRI and FDG PET that are associated with tumor hypoxia. Because localizing hypoxia is important for surgical resection and radiation therapy planning, we analyzed images in a region-based manner instead of a patient-base manner. CNN requires huge number of samples for training, and a region-based analysis helps increase the sample number. Thus, we aimed to apply CNN to predict tumor hypoxia from common imaging modalities, i.e., magnetic resonance imaging (MRI) and 18F-fluorodeoxyglucose (FDG) PET.
Methods: In this retrospective study, we reviewed a total of 32 glioblastoma patients (age, 63.5±14.8 years) who underwent FMISO and FDG PET, and MRI before surgical intervention. A total of 7 image series FMISO PET, FDG PET, non-enhanced T1-weighted images (T1WI), gadolinium-enhanced (Gd) T1WI, T2WI, fluid-attenuated inversion recovery (FLAIR), and diffusion weighted image (DWI) were used for analysis after SPM12 coregisterated all the images to individual T1WI. Every image was normalized by averaged value of contralateral frontal and parietal cortices. Polygonal regions of interest (ROIs) were drawn on all the slices that contained high intensity lesion on FLAIR image. Each ROI was cropped into 30x30 matrix square regions. When >25% pixels in the square region have higher FMISO signal than the threshold (tumor-to-normal ratio > 1.3), the square region was labeled as hypoxia positive; otherwise labeled as hypoxia negative. AlexNet, a well-established CNN implementation, was employed as an automated hypoxia probability classifier. CNN was trained and validated using the randomly selected 4/5 square regions while the data of the remaining 1/5 square regions were used for test purpose. The process was repeated 5 times to calculate the mean and SD of accuracy.
Results: A total of 3,453 square regions were provided per each modality after ROI cropping, among which 993 square regions were hypoxia positive and 2,460 were hypoxia negative. When GdT1W1 was given to AlexNet, the accuracy was up to 86.4±0.3% (The table shows sensitivity and specificity). Similarly, the accuracy for FDG PET and DWI was 81.6±0.8% and 76.2±0.3%, respectively. The accuracy for T1WI, T2WI, and FLAIR was inferior to that for DWI. Thus, to clarify whether multi-modality information improves the performance, we merged GdT1WI, DWI, and FDG to generate single image. The final products, which were 3-channel RGB colored images with each channel representing GdT1WI, DWI, and FDG signal, respectively, were given to AlexNet. The accuracy was further improved to 88.4±1.1%.
Conclusion: AlexNet and GdT1WI successfully predicted regional hypoxia, possibly because tumor hypoxia may be associated with abnormal blood vessels and blood brain barriers. In addition to that structural reason, tumor viability (represented by FDG) and tumor cell density (represented by DWI) also might have a partial role to induce hypoxia. Thus, CNN with MRI and FDG PET predicted hypoxia on glioblastoma with high accuracy. Research Support: