Autor(es): Silva, F ; Pereira, T ; Frade, J ; Mendes, J ; Freitas, C ; Hespanhol, V ; Costa, JL ; Cunha, A ; Oliveira, HP
Data: 2020
Identificador Persistente: https://hdl.handle.net/10216/145280
Origem: Repositório Aberto da Universidade do Porto
Assunto(s): Autoencoder; Lung cancer; Malignancy assessment; Transfer learning
Lung cancer late diagnosis has a large impact on the mortality rate numbers, leading to a very low five-year survival rate of 5%. This issue emphasises the importance of developing systems to support a diagnostic at earlier stages. Clinicians use Computed Tomography (CT) scans to assess the nodules and the likelihood of malignancy. Automatic solutions can help to make a faster and more accurate diagnosis, which is crucial for the early detection of lung cancer. Convolutional neural networks (CNN) based approaches have shown to provide a reliable feature extraction ability to detect the malignancy risk associated with pulmonary nodules. This type of approach requires a massive amount of data to model training, which usually represents a limitation in the biomedical field due to medical data privacy and security issues. Transfer learning (TL) methods have been widely explored in medical imaging applications, offering a solution to overcome problems related to the lack of training data publicly available. For the clinical annotations experts with a deep understanding of the complex physiological phenomena represented in the data are required, which represents a huge investment. In this direction, this work explored a TL method based on unsupervised learning achieved when training a Convolutional Autoencoder (CAE) using images in the same domain. For this, lung nodules from the Lung Image Database Consortium and Image Database Resource Initiative (LIDC-IDRI) were extracted and used to train a CAE. Then, the encoder part was transferred, and the malignancy risk was assessed in a binary classification—benign and malignant lung nodules, achieving an Area Under the Curve (AUC) value of 0.936. To evaluate the reliability of this TL approach, the same architecture was trained from scratch and achieved an AUC value of 0.928. The results reported in this comparison suggested that the feature learning achieved when reconstructing the input with an encoder-decoder based architecture can be considered an useful knowledge that might allow overcoming labelling constraints.
