The rapid growth and abundance of medicinal imaging technology is revolutionizing medicine. With medicinal imaging playing an increasingly prominent role in the identification and treatment of disease, the medical image analysis society has become preoccupied by the challenging problem of extracting, with the assistance of computers, clinically useful information concerning anatomic structures imaged through CT, MRI, PET, and other modalities. Although modern imaging devices provide excellent views of internal anatomy, the use of computers to analyze the embedded structures with precision and efficiency is limited. Accurate, repeatable, quantitative data must be efficiently extracted in order to support the spectrum of biomedical investigation and clinical activities from diagnosis, to radiotherapy, to surgery. So the main idea of Medical Image Recognition is the extraction of interest regions with high accuracy. This research report includes the study and analysis of existing algorithms for Region of Interest (ROI) extraction of lungs from CT images and implementation of an optimized algorithm. The overall process is divided into five stages and implementation of all these stages is discussed.
The field of digital image processing refers to processing digital images by means of a digital computer. A digital image is composed of a finite number of elements, each of which has a particular location and value. These elements are usually referred to as picture elements pixels. Generally speaking, digital image processing encompasses processes whose inputs and outputs are images and, in addition, encompasses processes that extract attributes from images, up to and including the recognition of individual objects.
The first computers powerful enough to carry out significant image processing tasks appeared in the early 1960s. Initially the work was done in using computer techniques to correct image deformation in the images sent by space probes. In the late 1960s, digital image processing techniques began to be used in medical imaging, remote Earth resources annotations, and astronomy. The innovation of Computerized Tomography (CT) scans in 1970s proved to be an important event in the application of image processing in medical diagnosis. From the 1960s until the present, the field of image processing has grown dynamically. Computer procedures are used to enhance the contrast or code the intensity levels into color for easier analysis of X-rays and other images used in industry, medicine, and the biological sciences. Image improvement and reinstatement procedures are used to process degraded images of unrecoverable objects or experimental results too costly to duplicate. Successful applications of image processing concepts can be found in astronomy, biology nuclear medicine, law enforcements, defense, and industrial applications.
The field of Computer Vision has the ultimate goal of using computers to imitate the human vision, including learning and being able to make inferences and take actions based on visual inputs. This area itself is a branch of artificial intelligence (AI) whose objective is to imitate human intelligence. The field of AI is in its earliest stages of immaturity in terms of development with progress having been much slower than originally anticipated.
1.2 Image Representation
An image may be defined as a two-dimensional function, f(x, y), where x and y are (plane) coordinates, and the amplitude of f at any pair of coordinates (x, y) is called the intensity or gray level of the image at that point. When x, y and the amplitude values off are all finite, discrete quantities, we call the image a digital image. Thus image is available as a two dimensional array for processing
Medical Image Recognition
The rapid growth and increase of medical imaging technologies is revolutionizing medicine. Medicinal imaging allows scientists and physicians to collect potentially life-saving information by peering noninvasively into the human body. With medicinal imaging playing an increasingly important role in the identification and treatment of disease, the medical image analysis community has become preoccupied by the demanding problem of extracting, with the assistance of computers, clinically useful information regarding anatomic structures imaged through CT, MRI, PET, and other modalities. Although modern imaging devices provide excellent views of internal anatomy, the use of computers to enumerate and examine the embedded structures with accuracy and effectiveness is restricted. In order to support the spectrum of biomedical investigation and medical activities from diagnosis, to radiotherapy, to surgery, accurate, repeatable, quantitative data must be efficiently extracted. So the main idea of Medical image recognition is the extraction of interest regions with high accuracy. Medical image recognition meets three main objectives:
Improve the quality of diagnosis
Increase therapy success by early detection of cancer
Avoid unnecessary biopsies.
1.4 Lung Cancer and CT Images
Lung cancer is known to be the form of cancer with the smallest survival rate after the diagnosis, by a gradual increase in the number of deaths every year. For this reason, the sooner it is detected, the greater the possibility of cure. Lung cancer is due to a special sort of structure known as nodule. The early detection of lung nodules is vital, either for close observation or biopsy to differentiate between benign or malignant nodules, or for timely therapy. The most common methods used to detect pulmonary nodules include chest X-ray and CT.
Fiber-optic bronchoscopy is al so used, but has limited value for finding nodules other than those in the larger airways. Recently MRI is also used but it is not very successful as the patient has to hold his breath for 2-3 minutes. CT offers better contrast than chest X-ray between nodule and background with no overlapping structures, and several studies have shown that CT can detect smaller, earlier stage nodules with a higher sensitivity than chest X- ray. CT technology has undergone a major evolution by the introduction of multi-slice technology. With multi-slice CT, a full-lung, thin-slice (< 1 mm) scan can be done within a single breath-hold. It is hoped that, with the high-resolution CT data available from multi- slice CT scanners, cancerous nodules can be recognized while still small and in an early stage of lung cancer. Many researchers assume that this down-staging effect achieved by early detection of lung cancer will ultimately improve the survival rate [1]. Moreover, it is hoped that lung cancer screening of high-risk patient groups may significantly increase the rate of lung cancer cases which are diagnosed before the cancer has metastasized. Because of this dominance of CT images over X-rays in m using CT scan Images in the thesis.
1.5 Problem Statement
Analysis of existing algorithms for ROI extraction of lungs from CT images and implementation of optimized algorithm.
1.6 Input/Output of the System
The input to the application is the Computed Tomography (CT) images in the bitmap format and the output will be the extracted lungs. A sample input is shown
Fig 1.1 Sample Input Image
1.7 Motivation
With its great success in diagnosis of different diseases CT images are increasing rapidly and radiologists are facing problems in controlling this huge amount of data.
With multi-slice CT images there are hundreds of CT images generated for a single person. The radiologists have to interpret the whole data to get a reading. Humans are limited in their ability to detect and diagnose disease during image interpretation because of their nonsystematic search patterns and the presence of structural noise. The radiologists may miss important patterns because of several reasons like lack of time, fatigue or other human reasons. A complicated anatomy combined with perceptual problems that accompany the projection of a three-dimensional object (the patient) into two dimensions (the image plane), makes identification of lung nodules a burdensome task for radiologists. So the thesis focuses on techniques to address these challenges.
CHAPTER 2
REQUIREMENT DEFINITION DOCUMENT
2.1 Introduction
An application is to be developed for the extraction of region of interest in the chest CT images. The application uses certain features devised for this purpose and generates the output according to the user requirement.
2.2 Thesis Drivers
2.2.1 The Purpose of the Product
With the recent advance of computer technology, medical diagnosis can be benefited from computers which will assist doctors to analyze medical data and images in improving accuracy. It is particularly important for doctors to take full advantage of computer capability. The aim of the product is to make a system which will act as a secondary check for the radiologist in the detection of lung cancer. The purpose of the system will be to help the radiologist and achieve accuracy by reducing the mistakes in the interpretations.
2.2.2 The Scope
To study image processing techniques, lung extraction techniques, and to implement a few recently published research papers. Since this thesis is a research thesis, so there will be a lot of learning involved in the research area.
2.2.3 The Objective of the Thesis
The main objective of this thesis is to extract the Region of interest (ROI) from CT images of the lungs with high accuracy.
2.2.4 Users of the Product
There are two main users of this product:
Specialized radiologists
The radiologists are the primary users of this system. This system aims to provide help to the radiologist in the detection and classification of lung cancer nodules by extracting the Region of Interest . The purpose of the system will be to help the radiologist in achieving accuracy by reducing the mistakes in the interpretations.
Students of this area
The research and product will also facilitate the students who are doing research in
this area. This product will also help them in comparing the results of various techniques used for the cancerous nodules detection and classification.
2.3 Relevant Facts and Assumptions
The following facts and assumptions were taken during the creation of this product:
The input image must be a CT scan image of lung.
The thickness of CT images must remain constant. In industry CT images of 2mm and 5 mm are common.
The input image must be in some recognizable image format e.g. .bmp.
2.4 Functional Requirements
The following are the functional requirements of the product
1. The product must read an image of bitmap format
2. The input image must contain a CT scan lung image.
3. The product should be able to extract the lung part from the image.
4. The phase output should be displayed on the screen.
2.5 Non-Functional Requirements
2.5.1 Interface
Since, the user is a radiologist and not a computer expert so the application should have a simple and easy to use Graphical User Interface (GUI).
2.5.2 Performance
The application should give a reliable and efficient output as expected by the radiologist.
2.5.3 Maintainability
The design of the proposed system should be easy to maintain in the future. It should be properly modularized and documented.
2.5.4 Reliability
The main non-functional requirement of the proposed software system is that it should provide the results as close as possible. The report should help Radiologists in detection of lung cancer.
CHAPTER 3
LITERATURE REVIEW
3.1 Introduction
Region of interest from lungs images can be extracted using different techniques. The two main approaches are rule-based reasoning and pixel classification .A rule-based scheme is a sequence of steps, tests and rules. Most algorithms for the segmentation of lung fields fall in this category. Techniques employed are (local) thresholding, region growing, edge detection, morphological operations and fitting of geometrical models or functions. On the other hand, several attempts have been made to classify each pixel in the image into an anatomical class (usually lung or background, but in some cases more classes such as heart mediastinum, and diaphragm ). Classifiers are various types of neural networks, or Markov random field modeling, trained with a variety of (local) features including intensity, location, texture measures. It is also possible to use general knowledge-based segmentation methods, such as active shape models, or extensions of such methods for the segmentation of lung fields.
3.2 Rule based Reasoning
Rule-based systems offer their designer the freedom to express his knowledge regardingthe problem in any type of rule or processing imaginable. Furthermore, it naturally subdivides the problem in sub-problems. If a part of the system does not yield satisfactory performance, one can add pre- or post-processing steps to correct the problem. There are some obvious weaknesses as well. In practice, rule-based systems consist of a concatenation of steps, each containing several variables, thus leading to an overall scheme that often contains many ad hoc choices and a myriad of user adjustable parameters. It is usually impractical or impossible to make plausible, let alone prove, that the performance of the system is in some sense optimal. The natural way to improve the system is adding more rules, thus also adding complexity. The same system cannot be applied to a deferent task: rule-based systems do not generalize. On the other hand rule-based system is accurate and robust therefore most algorithms for the extraction of lung fields fall in this category. The techniques employed in this technique are:
Thresholding
Region Growing
Edge Detection
Morphological Operations
3.3 Pixel Classification
Segmentation can be treated as a pixel classification problem by calculating a feature vector for each pixel in the input image. Output is the anatomical class the pixel belongs to. Any classifier from statistical pattern recognition or neural network theory can be used to approximate this mapping. The classifier is trained with a large set of training samples (pixels from a large collection of training images). Although different types of classifiers will obviously lead to different results, the performance of these segmentation algorithms will depend mostly on the features of the input vector. In most implementations, pixel classifiers use local features only. It is important to realize the limitations of such systems. It seems impossible, even for human observers to always correctly classify a very small region as lung or non-lung without the global percept of the whole image. Classifiers are basically various types of Neural Networks trained with a variety of local features including:
Intensity
Location
Texture Measures
3.4 Analysis of Algorithms
A detailed analysis of the existing algorithms for the automatic segmentation of lung fields in standard lung CT images is reviewed on the basis of accuracy, sensitivity, specificity. The algorithms are based on different techniques: matching, pixel classification based on several combinations of features, a rule-based scheme that finds lung contours using a general framework for the detection of oriented edges in images. Each approach has been discussed and the performance of the nine systems on 50 different test images is compared and results available from the literature are summarized in the table 3.1. Best performance is obtained the rule-based segmentation algorithm which outperforms the pixel classification approach. The rule-based system is accurate and robust: the accuracy is above 95% for all images in the test set. The average accuracy of the scheme is 0.959 +0.054, which is close to the inter-observer variability of 0.984 + 0.005.
Table 3.1 Analysis of Different Algorithms
Method
Accuracy
Sensitivity
Specificity
Duryea's Rule-based method
Vittitoe's MRF
McNitt-Gray's 59 features
Tsujii's PC corr. Int., location
Mc Nitt-Gray's 8 features
Vittitoe's PC int. 3x3
Vittitoe's PC location
Duryea's classifying location
Vittitoe's fixed Thresholding
Duryea's all negative
0.959 ± 0.054
0.948 ± 0.016
0.932
0.923
0.918
0.893 ± 0.027
0.880 ± 0.035
0.879 ± 0.063
0.806 ± 0.071
0.751 ± 0.063
0.863 ± 0.11
0.907 ± 0.044
0.949
0.903
0.846 ± 0.057
0.820 ± 0.08
0.785 ± 0.10
0.960 ± 0.058
0 ± 0
0.987 ± 0.044
0.972 ± 0.020
0.922
0.930
0.925 ± 0.038
0.920 ± 0.044
0.934 ± 0.063
0.781 ± 0.112
1 ± 0
3.5 Optimized Solution
On the basis of analysis it was found that rule based algorithm is the optimized algorithm for ROI extraction of lungs from CT images. The basis of this algorithm involves finding a threshold in the density histogram of CT chest images. The approach involves following steps.
3.5.1 Thresholding.
Shapes like nodules, bones and vessels are brighter than other structures. It means that they have higher HU values. Thus thresholding is performed in order to extract the lung region. If A(x, y) is the input image seen in Figure 3.1, by applying the following rule Figure 3.2 was achieved. Here 1 represents white and 0 represents black.
IF I(x, y) < -500 HU THEN
A(x, y) = 1
ELSE
A(x, y) = 0
Fig 3.1 Input image
Fig 3.2 Thresholded Image
3.5.2 Region Growing and Edge Detection
As seen in Figure 3.2 the white lung region contains small black shapes representing nodules and vessels. To eliminate them we label the black shapes using connected component labeling (CCL) and analyze their sizes. When all black shapes in the lung region seen in Figure 3.2 were labeled, their sizes were analyzed using the following rule. If S(k) representing the size of kth black shape in pixels is smaller than 150 pixels, we assign 1 to the A(x, y) which is representing the pixels of kth shape of the image. Thus by eliminating the small black shapes in Figure 3.2, Figure 3.3 was achieved.
IF S(k)<150 pixels THEN A(x, y)=1
Fig 3.3 Initial Lung Mask
3.5.3 Morphological operations and lung region extraction
To extract the lung region in Figure 3.3, all white shapes were labeled using CCL and their morphologies were analyzed. As seen in Figure 3.3 the aspect ratio lung region is smaller than the other shapes. The morphologies of the shapes were analyzed using the following rule where AR(k) represents the aspect ratio of kth shape which is the ratio of height to width, W(k) represents the width of kth shape and A(x, y) represents the pixels of kth shape of the image. According to this rule the shapes, whose aspect ratios were less then 1.5 or
width were less then 50 pixels, were eliminated and thus the lung region was extracted.
IF AR(k)>1.5 OR W(k)<50 pixels THEN I(x, y)=0
At the end of the third step the lung region of Figure 3.1was segmented as Figure 3.4 and Fig 3.5.
Fig 3.4 Complete Lung Mask
Fig 3.5 Extracted Lungs
CHAPTER 4
IMPLEMENTATION
The Optimized Solution has been implemented using Matlab coding. The implementation of this approach involves five steps which are shown in a flow chart.
Fig 4.1 Flow Chart of the implementation Process
4.1 Image Acquisition
In image acquisition phase data was to be collected. So a database of 50 different CT images was obtained named LIDC Database. Each frame can separately be viewed in the Dicom viewer and exported in the bitmap format.
4.2 Image Preprocessing
The CT scan image of lung has many other structures apart from the lungs. This may include some bones, fat and/or bed etc. Moreover the background in the image is also undesirable. To proceed efficiently for nodule detection and its classification it is necessary to first remove these unwanted parts and extract only the area of interest which is lungs. So the purpose of this phase is to remove the unwanted parts and to extract the region of interests from the image. The input image and the desired output of this step are shown in figure.
Fig 4.2 Image Preprocessing
4.3 Gray Scale Image
The acquired CT image is converted into 16 bit grayscale image. After this step all the pixels value comes in the range of 0 to 216 -1. This helps in reducing the total number of distinct values. We have used IrfanView Software to convert the data images into the grayscale image. The actual image and its corresponding gray scale image are shown in figure 4.3 and figure 4.4 respectively.
Fig 4.3 Actual Image
Fig 4.4 Gray Scale Image
