Lung Nodules Detection from Computed Tomography Scans Using Deep Belief Networks
No Thumbnail Available
Date
2018-10
Authors
Journal Title
Journal ISSN
Volume Title
Publisher
Addis Ababa University
Abstract
Lung cancer is the leading cause of cancer related deaths globally. Analyzing thousands of
computed tomography (CT) scans are an enormous burden for radiologists which results
for inefficient diagnosis. Hence, a need to read, detect, and provide an evaluations of CT
scans efficiently exist to assist radiologists by improving accuracy, time delay to diagnose,
human errors, and making bias for specific reasons. Many researches have been conducted
to detect lung nodules from CT scans. However, lung nodule detection focusing on the
lung than the CT image as a whole has not been conducted so far. Thus, in this research
work, lung nodules detection system, which segments lung and lesions from CT images to
reduce false positives and employing Deep Belief Network (DBN), is proposed to improve
nodule detection accuracy.
The study comprises three main phases namely: Image Processing, DBN training and
nodules classification. The process starts with DICOM to JPEG conversion. Median filter
and histogram equalization are applied for noise removal and contrast adjustment. We
designed lung segmentation algorithm from the concept of inverse and intersection
operation to separate lung object from the whole CT image. We applied adaptive
thresholding for segmenting detail elements of CT images. Lesions on the lung are
segmented. This thesis is conducted using datasets, publicly available on Lung Image
Database Consortium (LIDC) and Image Database Resource Initiative (IDRI). We
implemented our methods based on 201 DICOM files that consist of 36,520 samples from
which a total of 221,807 lesions are used to prepare feature vectors. This feature vector sets
are used as input to the DBN algorithm for training and model construction. A DBN with
5776-500-500-2000-2 architecture, learning rate of 0.1, 100 number of epochs, and
backpropagation as fining tuning algorithm is used in constructing the classifier.
The DBN classifier was validated with ten-fold cross validation technique. The proposed
classifier achieves a sensitivity of 81.143%, specificity of 92.47% and an accuracy of
91.247%. The classifier has also an Area under the ROC curve (AUC) value of 0.882 and
0.924 for malignant and benign cases respectively. Therefore, based on the result we found
that DBN model has the potential for lung nodule detection.
Description
Keywords
Computed Tomography, Ten-fold Cross Validation, DBN, DICOM, AUC, ROC curve