Semantic segmentation of medical images has become increasingly popular for machine learning and deep learning researchers. Good semantic segmentation is important for automatic detection, localization, and classification of lesions. The application of Deep learning could save the radiologist a lot of time. Deep architectures, usually Convolutional Neural Networks (CNNs), are surpassing other approaches in terms of accuracy. The publicly available dataset LIDC- IDRI contains 1010 CT-chest scans of patients with varying degrees of lung cancer. This dataset is therefore often used for testing new segmentation algorithms to segment lung nodules.
Generative Adversarial Networks (GANs) are a relatively new technique for image generation. They have been used in generating magnetic resonance (MR) images from computed tomography (CT) images, but also to denoise CT images. Mostly, GANs are used to enlarge a data set by simulating more images. However, it would be interesting to see if a GAN would be able to produce an annotated area of interest from a full CT scan slice, or in other words, segmentation of lung nodules. When the annotations are made by a radiologist, the GAN would, ideally, be able produce annotations in the same manner.
The goal of the project is to create a automatic semantic segmentation method to annotate lung nodules from CT chest images using GANs and annotations made by the radiologist.
The project will therefore consist of several parts:
- Dataset preparation to make them applicable to deep learning
- Implementing a GAN-based method for nodule segmentation
- Validation of the method