U-Net Lung Cancer Detection and Segmentation

Lung Cancer is the most deadly cancer. Early detection of the disease can improve survival rate. Automation of lung nodules detection aid radiologists in quickly and accurately diagnosing the disease. Developing computer aided diagnosis (CADx) systems for lung cancer is a challenging task. Several components make up CADx. One of the most significant components is lung segmentation, an essential prerequisite to efficiently detect and classify lung nodules. Lung segmentation is the process of segregating lungs fields from other tissues in the CT image. Conventional methods for lung segmentation either do not accurately segments normal and abnormal lungs or rely heavily on user generated features for the lungs. Deep learning has outperformed other methods in image processing tasks. Recently a new architecture has been proposed and implemented exclusively for medical images to solve this problem namely U-Net convolutional network. In this study u-net has been implemented on lungs dataset consisting of 267 CT images of lungs and their corresponding segmentation maps. The accuracy and loss achieved is.

Considering the computational constraints the results obtained are state of the art. Keywords – U-Net convolutional network; Lung Parenchyma; Segmentation methods; Thoracic CT Scans. Globally, lung cancer is the leading cause of cancer mortalities. There were 1.69 million deaths in 2015 due to lung cancer [1]. Early detection of lung cancer increases the survival rate, but is like searching for needle in the haystack. Abnormal small, round or oval shaped growth in the lung called lung nodules may be the first sign of lung cancer and detection of those is very exhaustive given the complex structure of lungs. Computed Tomography (CT) is an important diagnostic modality to detect lung nodules. The automation of detection and diagnosis of lung nodules benefits both the radiologists and patients. Accurate detection of lung nodule at an early stage leads to proper treatment and saves patients life. There are several computer aided diagnosis (CAD) systems developed over the years to assist in the diagnosis of lung cancer. CAD system components include lung segmentation, nodule detection and segmentation, false nodule reduction, nodule classification. Each component is complicated in its own way. The significance of segmentation of the lungs from chest CT scans has been elaborated in [2].

Lung segmentation is a prerequisite for the subsequent automated analysis of lung nodules since it allows for the estimation of lung volumes and detection and quantification of abnormalities within the lungs. In case of erroneous lung segmentation, findings might be missed or findings outside the lungs might be included in the analysis. The importance of accurate lung segmentation for the automated detection of nodules is illustrated in [5]. Their experiments have showed that accuracy of nodule detection has increased when lung segmentation has been applied. A naive lung segmentation algorithm was applied to 60 scan, 17% of nodules were not detected as a consequence of improper lung segmentation. Another lung segmentation algorithm improved the results and only 5% of nodules were not detected. The task of lung segmentation is challenging because of the complexity in the lung region and the existence of similar density structures, such as arteries, veins, bronchi and bronchioles, and the use of different scanning devices with different scanning protocols [3]. In general, the existing techniques for lung segmentation can be classified into different categories based on:

a) Intensity. Eg. Thresholding.

b) Region. Eg. Region Growing.

c) Shape. Eg. Sobel.

d) Edge. Eg. Wavelet Transform.

e) Machine Learning. Eg. SVM.

Automatic lung segmentation has always been a challenging task. Several algorithms have been proposed that address the problem of accurately extracting lung region from the CT images. Conventional methods mostly rely on the attenuation values of different areas on the CT. The fact that similar regions have same intensity helped the development of a popular intensity based segmenation method namely thresholding. Tremendous number of research papers has developed thresholding methods with many variations. Traditional algorithms are two-dimensional and process each axial section of the scan separately. 3D image processing of CT scans has been developed that take into account height, width and depth of an image. Most of the methods for lung segmentation begin by determining the lung fields using optimal gray-level thresholding and connected components or region growing method. The detected lung region contains trachea and bronchi which has to be removed from the image. When the lungs are joined in the anterior or posterior junctions, they have to be separated to obtain the left and right lung regions only. Post processing include applying of the morphological operations on the segmented image to smoothen the borders and fill the gaps.

The literature of lung segmentation include numerous papers which are described below: Hu et al (2001) were the first to publish a threshold-based lung segmentation method based on the method described above. Ukil and Reinhardt (2005) improved upon the method proposed by Hu et al (2001) by introducing a smoothing at the mediastinal area based on the airway tree to guarantee consistency among segmentations of different subjects and intra-subject over time. Sluimer et al (2005) and van Rikxoort et al (2009a) describe 3D threshold-based methods largely based on the method of Hu et al (2001). Sun et al (2006) presented a 3D method for the segmentation of the lungs from thick-slice CT images. First, a preprocessing was applied in which the signal-to-noise ratio was improved by applying an anisotropic filter, followed by a wavelet transform-based interpolation method to construct 3D volume data. In these 3D volume data, the lungs were obtained by region growing using gray-value, homogeneity and gradient magnitude as input. Cavities inside the resulting lung region were filled using morphological closing.