Automatic classification of Pulmonary Nodules, utilizing Convolutional Neural Networks and Generative Adversarial Networks
| Main Author: | Ioannis Dimitris Apostolopoulos |
|---|---|
| Format: | info software |
| Bahasa: | eng |
| Terbitan: |
, 2020
|
| Online Access: |
https://zenodo.org/record/3679321 |
Daftar Isi:
- Subject: Lung cancer is one of the highest causes of cancer-related death in both men and women. Therefore, various diagnostic methods for Solitary Pulmonary Nodules (SPNs) automatic characterization have been proposed to aid to the early detection, and, possibly, early remedy. Since the available image data often lacks the diversity and quantity to train robust classification frameworks, several data augmentation methods have been proposed. In this paper, data augmentation methods to aid to the accurate classification of a very small PET/CT dataset of 172 single SPNs are proposed. Materials and Methods: utilising Deep Convolutional Generative Adversarial Networks (DC-GAN), new fake SPNs are generated to create a large-scale dataset. For the training of DC-GAN, the publicly available dataset of LIDC-IDRI along with the PET/CT dataset of the study are joined. The performance of DC-GAN is evaluated by employing experts to distinguish between real and fake nodules. Besides, a hierarchical CNN, called as Late - Extractor VGG19 (L-VGG19) network, is developed to enhance the extracting information of VGG19, which is a pre-trained CNN commonly used for medical image classification tasks. L-VGG19 was firstly utilised to label the generated nodules, by training it on every dataset available. Secondly, L-VGG19 was trained on the publicly available datasets joined with the DC-GAN generated nodule images with the aim to enhance its classification performance. After the training, L-VGG19 predicted the malignancy rating of the 172 SPNs of PET/CT dataset, which remained completely hidden during training. Results: DC-GAN manages to generate realistic SPNs, as the experts achieved to distinguish only __% of the fake nodule images. L-VGG19 obtains the highest classification accuracy (83.13%) among other VGG16 and VGG19 modifications proposed by other works. The proposed training scheme and data augmentation method improves the classification accuracy by 7% compared to training without any augmentation.
- Contains Python Code (Keras, Numpy, Tensorflow, Imutils, Scikit)