Retrospective Fat Suppression for Lung Radiotherapy Planning with CycleGANs

Abstract

<h3>Purpose/Objective(s)</h3> MR-based radiotherapy planning and MR-guided radiotherapy are gaining traction. In current practice, several MR sequences are used for planning, including T2 weighted nonfat-suppressed (NFS) images that show fat between organs at risk and fat-suppressed (FS) images that optimize tumor conspicuity. To spare time and avoid potential motion between sequences, retrospective fat suppression with AI has been explored. Rowland et al. found that models learned to introduce artifacts present in training data. In this study we investigate whether training models on a smaller dataset without artifacts produces better results than training on a large dataset including artifacts. <h3>Materials/Methods</h3> 21 lung cancer patients were each scanned twice on a 1.5T Aera. At each session, 2D TSE, navigator triggered, T2-weighted images (TE/TR=102/5200ms, 30 × 3.5mm slices with matrix 320 × 320, 1.25 × 1.25mm pixels, TA ∼6 minutes) were acquired both with and without fat suppression. Ten NFS/FS slice pairs were randomly selected as the test dataset and remaining images were used for training. Three CycleGANs with identical structures were trained as follows. Full model was trained on the full dataset (910 image pairs) for 400 epochs. For the partial model images with fat suppression artifacts were removed (resulting in 500 pairs) and the model was trained for 720 epochs. Transfer model was trained on full dataset for 200 epochs, then partial dataset for 360 epochs. This selection results in equal training time for all models. Fat and tumor intensities are similar, thus fat suppression can cause undesirable reduction in tumor signal. To assess changes in both areas, tumor and subcutaneous fat were delineated on test images and mean intensity change compared to NFS image in delineated areas was calculated. Also, an MR radiographer and a medical physicist scored quality and fidelity of the tumor region in 10 sets of generated FS images against a corresponding image with prospective FS on a 5-point Likert scale. <h3>Results</h3> Table 1 shows results of qualitative and quantitative analyses. The partial and transfer models were better than full network at suppressing fat without introducing artifacts. Within tumors partial model provided superior sharpness, while transfer model better maintained intensity level and fidelity. <h3>Conclusion</h3> Despite differences being small, transfer learning approach produced the result closest to the real FS, indicating that both quality and quantity of training data are important.