Although large-scale datasets are required for learning accurate Computer-Assisted Diagnosis, most available medical datasets are small or fragmented. This paper [1] proposed 3D Multi-Conditional GAN (MCGAN) to generate realistic and diverse nodules placed naturally on lung CT images to boost sensitivity in 3D object detection. Generally, GAN has two networks: a generator which generates realistic images, and a discriminator which determines whether the image is real or not, and these networks are made to compete with each other. In order to MCGAN generate more realistic and diverse nodules, authors used two discriminators: the context discriminator and the nodule discriminator. The context discriminator learns to classify real vs synthetic nodule/surrounding pairs, and the nodule discriminator learns to classify real vs synthetic nodules with size/attenuation conditions (Figure 1). Visual Turning Test results by two radiologists showed that it was difficult for even expert radiologists to classify real and MCGAN-generated nodules (Table 2). In addition, nodule detection CNN achieved higher sensitivity by adding these data to learning (Figure 2). In the future, these data are expected to not only help learning more accurate detection models, but also train medical students.
(Accuracy denotes the physicians’ successful classification ratio between the real/synthetic images.)
DOI: https://doi.org/10.1109/3DV.2019.00085
CAUTION：This is Fujifilm Global Website. Fujifilm makes no representation that products on this website are commercially available in all countries. Approved uses of products vary by country and region. Specifications and appearance of products are subject to change without notice.