br In Nurfauzi et al the proposed
In Nurfauzi et al. (2017), the proposed method is designed to reduce the computational time in Pu et al. (2008). The proposed method used in Pu et al. (2008) is improved by using local minima and maxima to select the pixel pairs in lung boundary that are sus-pected as nodule basin. The result has succeeded in accelerating the computational time 40 times faster than the previous one. However, the under segmentation value is increased. The decrease in performance of the covering nodule area in Nurfauzi et al. (2017) is due to the use of minima and maxima features extracted from the lung boundary. In this case, minima and maxima features are only able to detect the curvature in x and y directions.
Other limitation of the previous studies in Pu et al. (2008) and Nurfauzi et al. (2017) is due to the unavailability of lung fusion separation process. Separation of lung fusion itself is a challenging task. Without separation of lung fusion, the system may determine the mediastinum area as a nodule basin. Therefore, separation of lung fusion is necessary prior to lung boundary correction.
In short, there are two limitations of the aforementioned stud-ies. Firstly, there is no separation of lung fusion at the Concanamycin A stage. Secondly, the large computational time is required to correct the lung boundary. To overcome these limitations, multi threshold and extraction of corner feature are proposed to reduce pair con-nection in ABM. This approach is expected to detect curvature in all directions resulting in less error of covered nodule area as well as to reduce computational time.
2. Materials and methods
There are five stages proposed in this study, namely ground truth (GT) extraction, data preparation, tracheal extraction, separa-tion of lung fusion and lung boundary correction. Each stage is explained in detail as follow.
2.1. Ground truth (GT) extraction
Each 3D lung CT image in LIDC-IDRI has been evaluated by four radiologists to provide location and type of nodules. This informa-tion is called as ground truth (GT). Each GT of all patients is stored in the different folder in .xml format. The .xml to .png converter using Matlab has been offered by Lampert et al. (2016). The GT extraction results are stored in a folder named using the unique index number. The extraction process is shown in Fig. 2. An exam-ple of GT extraction result is shown in Fig. 3 (Nurfauzi et al., 2017).
2.2. Data preparation
The focus of this study is to segment lung area containing only juxta pleural and vascular nodules. There are three steps in data preparation. Firstly, find and select the main CT data. Secondly, CT data should be sorted. Finally, HU values are converted to CT values.
Load the ground truth
Convert .xml to
Store and named the
data with “UID”
Fig. 2. The flowchart of ground truth extraction.
Fig. 1. Category of lung tumors based on position: (a) well-circumscribed nodule;
Fig. 3. Presence of juxta pleural nodules in lung CT images in the LIDC-IDRI dataset annotated by 4 radiologists (Javaid et al., 2016).
Please cite this article as: R. Nurfauzi, H. A. Nugroho, I. Ardiyanto et al., Autocorrection of lung boundary on 3D CT lung cancer imagesq, Journal of King Saud University – Computer and Information Sciences, https://doi.org/10.1016/j.jksuci.2019.02.009
R. Nurfauzi et al. / Journal of King Saud University – Computer and Information Sciences xxx (xxxx) xxx 3
Based on the stored ‘‘Series Instance UID” in CT header, each patient file can be classified into two or even three of CT image groups. The groups include the main of 3D lung CT images, two coronal images and the second taken data of 3D lung CT images.
The juxta pleural and vascular with a specific nodule size can be found in GT images. The used nodule size in this study is more than 3 mm (Riccardi et al., 2011). Only data with the same UID as that of the GT file name will be selected. In this database, CT slices have not been sorted. ‘‘SliceLocation” or ‘NumberInstance’ stored in the CT header is used to sort slices in the 3D CT lung image.
CT images are stored in two image types, i.e. in CT scale and in HU scale. The relationship of both scales is formulated in Eq. (1) (Moosavi Tayebi et al., 2015).
HU ¼ ðPixelvalue SlopeÞ þ Intercept ð1Þ
Pixel value describes the intensity of CT image. Slope and Inter-cept values are obtained from ‘RescaleSlope’ and ‘RescaleIntercept’. These values are obtained from header of each CT slice. The flow-chart of these step is depicted in Fig. 4.
2.3. Lung extraction
The aim of this stage is to segment the lungs with attached organ in low intensity (trachea). There are two steps in this stage.