Lung nodule classification and creation of a patient similarity metric on the basis of deep learning algorithms
progresses. These advances offer great opportunities for the field of applied medical
image analysis. The goal of our study is to investigate the application of deep learning
algorithms to human lung screenings.
Deep learning is a machine learning technique which allows the representation of the
underlying data through a set of highly complex features. By visualizing / interpreting
those features we intend to augment the radiologists ability to detect lung nodules in an
early stage. This ultimately can result in an improvement of healing rates through early
medical treatment and thus in a reduced number of deaths from lung cancer.
However, the main drawback of deep learning algorithms is the large amount of image
data required for achieving accurate results. With the help of the NLST dataset, we plan
to create a deep Convolutional Neural Network (CNN) on lung images, in order to extract
features of high clinical relevance.
Aim #1: Localization of lung nodules and suspect regions.
An automated localization of lung nodules and suspect regions may result in a significant
workload reduction for the attending radiologist. Therefore, one of our main goals is to
train a Deep Learning algorithm that will detect and localize lung nodules with high
accuracy.
Aim #2: Probabilistic prediction of malignancy of a given nodule
Secondly, we intend to produce accurate, automatic probabilistic estimations of the
malignancy of a given nodule.
Aim #3: Extract complex features in order to generate patient / nodule similarity metric
On the basis of our localization and classification deep learning algorithms, we intend to
extract highly complex features to generate a patient/nodule similarity metric which will
ultimately support radiologists in their medical decision making.
Niklas Köhler, Hellsicht
Philipp Eulenberg, Hellsicht