PREDCGAN: uma abordagem para geração de nódulos pulmonares sintéticos GAN usando pré-treinamento

Abstract

The diagnosis of lung cancer using computed tomography (CT) images is a complex and challenging job, even for experienced radiologists. To mitigate this difficulty, Computer Aided Diagnostic (CAD) systems are developed to assist the performance of specialists. One of the main areas used in CAD is Deep Learning (DL). However, because it is a difficult process to validate and depends on a well-trained professional, the construction of large databases of structured, well-validated and comprehensive medical images is one of the main challenges of DL applications in CAD in medical images. The generative adversarial networks (GAN) consist of a model to generate heterogeneous probabilistic data that presents itself as an adequate solution to supply the small number of pathological exams that make up the most diverse medical image banks. However, GAN training for pulmonary nodules does not use conditional information to generate synthetic images. Thus, in this work it is proposed to use knowledge transfer applied through GAN pre-training to result in more efficient synthetic pulmonary nodules for base increase. For this, the training of a textit Deep Convolutional GAN (DCGAN) was carried out in two moments, the first to generate segmented nodules and in a second moment a fine adjustment was made to generate nodules with the adjacent structures. Through the use of the images generated by our proposal as a base increase and a comparison with other techniques of base increase, we obtained an increase in AUC of 0.77 % in relation to the increase of classic base and when used our proposal in conjunction with increase of classic base we have an increase in AUC of 2.06 % in relation to the individual of the increase of classic base.