PROJECT TITLE :

Unsupervised Multi-Discriminator Generative Adversarial Network for Lung Nodule Malignancy Classification

ABSTRACT:

In the detection of lung cancer, computer-aided diagnosis systems with deep learning frameworks have been utilized to identify benign and malignant pulmonary nodules. Large-scale labeled datasets are a well-known concept for training massive deep neural networks. In many medical picture fields, however, the wealth of labeled datasets is frequently missing. Deep learning models' generalization performance may suffer as a result of this aspect. For the classification of benign and malignant pulmonary nodules, we propose an unique multi-discriminator generative adversarial network model with an encoder in this research. We believe we are the first to use unsupervised learning to distinguish between benign and malignant lung nodules. To begin, we train a generative model with unlabeled benign lung nodule images using a multi-discriminator generative adversarial network. Then, using an encoder and the trained generative model, a mapping of benign pulmonary nodule images to the latent space is created. The GAN discriminator feature loss and image reconstruction loss are used to score benign and malignant lung nodules. On cancerous photos, the model produces high anomaly scores, while benign images provide low anomaly values. Experiments suggest that our method, when compared to existing supervised deep learning approaches, can generate better results with a smaller amount of unlabeled datasets.