ESTRO 2021 Abstract Book

S1416

ESTRO 2021

inputs, was recently proposed for deep classification networks. The algorithm contributed to increased performance on classification in a variety of datasets, but so far has not been evaluated for image segmentation tasks. This paper aims to test whether the mixup algorithm can improve the generalization performance of deep segmentation networks. Materials and Methods We trained a standard U-net architecture to segment the prostate in 100 T2-weighted 3D magnetic resonance images from prostate cancer patients (50 for training, 50 for testing). The performance of repeated training sessions with and without mixup data augmentation was compared with the Dice similarity coefficient and mean surface distance from a reference segmentation made by an experienced radiologist. Results Our results suggest that mixup offers a statistically significant boost in performance compared to non-mixup training, leading to up to 2% increase in Dice and a 22% decrease in surface distance. PO-1690 Synthetic generation of pulmonary nodules using super resolution generative adversarial model Z. Wang 1 , L. Wee 2 , A. Dekker 3 , A. Traverso 1 1 Department of Radiation Oncology (Maastro), GROW School for Oncology, Maastricht University Medical Centre+, Maastricht, The Netherlands; 2 Department of Radiation Oncology (Maastro), GROW School for Oncology, Maastricht University Medical Centre+, Maastricht, The Netherlands; 3 Department of Radiation Oncology (Maastro), GROW School for Oncology, Maastricht University Medical Centre+, Maastricht, Maastricht, The Netherlands Purpose or Objective The early detection of pulmonary nodules is a key player to improve life expectancy of lung cancer patients. Among nodules, the detection of pulmonary ground-glass opacity (GGO) is the most challenging. GGOs are not obvious compared with other types of nodules and often can be overlooked. Since manual labelling of medical images is a time-consuming activity, there are difficulties in obtaining data for “edge cases. Most of the medical datasets are inadequate and imbalanced. For example, the GGO nodules are 6% in labeled data from the LIDC-IDRI dataset, making difficult to optimize computer aided diagnosis systems to recognize this type of nodules. The synthetic data by GANs can increase the number of “edge cases”, which can solve the imbalance problem. Though there are many 3D GANs to generate synthetic samples, most of them cut the whole CT into small patches. Nevertheless, the nodule can just appear in less than 10 slices, the traditional 3D input with a square shape will bring too much useless Spatial information. Furthermore, the small patches input will miss the information about the location of the nodule. Besides that, the original GANs with one discriminator can be easily led to the model collapse, which means two discriminators can increase the diversity of synthetic results. Materials and Methods We proposed a novel structure for generating GGO nodules through dual discriminator 3D-based SRGAN. In the network, we use the 3D Resblock as the backbone for the generator and discriminators. To avoid model collapse, two discriminators have different structures built by 2D convolutional layers and do not share parameters when training. For training and testing purposes we used independent subsets of the LIDC dataset. Differently from the small patches of the traditional 3D network, the input of our network is 512*512 with 3 slices. As loss function, we proposed a novel fusion loss that has the different weights of the SSIM loss, content loss, and Total Variation loss (TV loss) in both small patches of nodule images and the whole Conclusion The mixup algorithm can improve the performance of deep segmentation models.

images. Results

Figure 1 shows that the nodule can be created at an arbitrary position in the scan. Second, it can be easily recognized from the appearance that the synthetic nodule has the same type and position as the real nodule. In figure 2(d), the histogram shows that the synthetic nodule has a similar distribution as the real one at the voxel level.