Classification of Candidate Pulmonary Nodules Segmented from CT Image by Using CNN
No Thumbnail Available
Date
2019-10
Authors
Journal Title
Journal ISSN
Volume Title
Publisher
Addis Ababa University
Abstract
Lung Cancer is a leading cause of human loss globally i.e. compared with other cancer related
deaths. The five-year relative survival rate of lung cancer is only 16%; however, early
recognition of nodules and proper treatment of this disease reduce the death rate due to lung
cancer up to 20%. To detect such nodules CT lung image analysis has been used by radiologists
all over the world. However, analysis of these images is a very challenging task for radiologists,
because the number of slices in one scan can be up to 600. Therefore, computer aided-detection
(CAD) systems are very important for a quicker and more precise assessment of the data. False
positive reduction is one of a vital element of computer aided diagnosis (CAD system), which
plays an important role in lung cancer diagnosis and early treatment. In this thesis we proposed
to design a framework for classification of candidate CT image slices by employing 3D
Convolutional Neural Network (CNNs) to reduce a significant number of false positive
candidates. 3D CNNs are favorable than 2D CNNs because 3D CNNs can encode richer spatial
information and extract more representative features via their structural architecture trained with
3D samples. The proposed design has been extensively validated by using the dataset obtained
from LUNA16 challenge providers.
The proposed approach mainly consists of three steps Pre-processing, Feature extraction
& Classification and Fusion. In the preprocessing phase we carefully examined our data set and
perform resampling to avoid image slice thickness variations because of different CT machines.
In addition to that we employed data augmentation to reduce class imbalance between ground
truth nodules and false positive candidates. Sizes of the nodules varied from 3mm up to 30mm,
so we extract four receptive fields to encompass all nodule types (i.e. small, medium and large
nodules), aiming to understand the effect of input patch sizes in the performance of the system.
We designed four 3D CNNs for the corresponding four patch sizes, each CNNs model contains
from 3 convolutional layers. We employed model fusion technique to acquire an improved result
by using the aggregate strengths of each model. The proposed framework has been tested by
the dataset provided by the LUNA16 Challenge and we achieved the competition
performance metric (CPM) score of 0.8541 with a highest sensitivity of 0.8706 with 1 false
positive per scan and 0.9275 at 8 false positives per scan.Generally, from the test results we noticed that increasing the input patch size increases the
overall CPM score because larger patch sizes can be able to encompass a large number of
nodules within the dataset and any reasonable fusion method can be able to boost the overall
classification performance.
Description
Keywords
Pulmonary Nodule Detection, Computed Tomography, Convolutional Neural Networks, Deep Learning, false positive reduction, computer-aided diagnosis