Neural Pca Network for Lung Outline Reconstruction in Vq Scan Images

This research focuses on design of a software-based image analysis system towards facilitating automated diagnosis of Pulmonary Embolism using ventilation-perfusion scans and correlated chest x-rays. This proposed system takes the digitized ventilation-perfusion scan images of lungs as input, identify a template according to the size and shape of the lungs and thereby approximate and reconstruct the outline of the lung. The proposed lung outline reconstruction system was designed to facilitate the PIOPED–compliant diagnosis procedure. The system was trained with actual patient lung images, where lung images were compiled to represent the shape and size variation of the population in general. Both adult male and female lung images were utilized. A neural principal component analysis network was used and tested with actual patient lung images obtained through the Medical College of Ohio patient repository, which represented five probability classes as they exist in the PIOPED criteria. Testing results, which were obtained through MATLAB simulation, indicate that neural PCA algorithm trained with generalized Hebbian learning performed well although it demonstrated performance degradation for high probability Pulmonary Embolism cases. The work as presented proved in concept that it is feasible to construct the outline of the lungs and is open to enhancement in numerous ways. INTRODUCTION Image processing and recognition are extensively used in medicine for the purpose of automating identification of various organs in diagnositic images and diagnosing the diseases related to the same. Lungs being a fundamental part of the human respiratory system, a number of researchers have developed algorithms to identify the human lung and the defects related to them in a variety of imaging modalities. One of the many research projects being carried out is for the diagnosis of Pulmonary Embolism [Gabor F.V., 1994], [Fisher R. E., 1996], [Armato S., 1997]. Pulmonary Embolism is a disease of the lung where the arteries to various anatomical regions of the lung are occluded by emboli, which originate from venous thrombosis, interfering with the normal gas exchange. Anatomically, lungs are divided into two regions: the right lung and the left lung. The right lung is further divided into the upper, middle and lower lobe bronchi; and the left is divided into only the upper and lower lobe bronchi. The lobes are further divided into segments. In Ventilation-Perfusion (VQ) scanning, the ventilation scans are taken before the perfusion scans. In ventilation imaging, 10 to 20 mCi of Xenon 133 is administered by employing a number of commercially available delivery and rebreathing units. The scan is taken while the patient holds his breath for 15 seconds. Perfusion scans are taken by injecting particulate radiopharmaceuticals during quiet respiration when the patient is in the supine position. The perfusion scans are obtained on a large-field-of-view or a standard-field-of-view camera with a diverging collimator. Images are obtained in the anterior, posterior and both lateral positions. These VQ scans are employed by the radiologists to determine the possibility of the patient suffering from Pulmonary Embolism. The radiologists look for possible defects in the perfusion scans and corresponding mismatch in the ventilation scans to determine the possibility of Pulmonary Embolism. The following is the mental procedure followed by nuclear radiologists to extract feature data from VQ scans and correlating chest x-rays to aid in diagnosis of Pulmonary Embolism: • First construct a mental image of the outline of the lungs. • Mentally superimpose the segmented anatomy over the reconstructed outline of the lungs. • Mentally locate and list the number and size of the defects in each segment. • Finally, match the observed defects with the PIOPED criteria to conclude a diagnosis. Manual interpretation of the ventilation-perfusion scans and the correlated chest xrays may not be very precise, since these images are of relatively low resolution and can at times be fuzzy. Such interpretations also depend on the skill and experience level of the reader, in this case a nuclear radiologist. Therefore, there may be significant variations in the diagnosis from one reader to another. Automation of the scan-reading procedure that mimics the radiologists' approach to extract data from the VQ scans has the potential to eliminate inter-observer variability. Even though a considerable amount of research has been carried out in the field of lung imaging, it is still not at a stage to directly benefit the automation of the diagnosis of Pulmonary Embolism (PE). The automated outline reconstruction procedure should yield a smooth curve without eroded surfaces, which can result in false estimation of the probability of Pulmonary Embolism to the desired degree. The lung image also needs to be partitioned into anatomic segments so that the location of a defect in a certain segment in the lung can be pinpointed. Determination of the segment, where the defect is present, is necessary to evaluate the probability of Pulmonary Embolism according to the PIOPED criteria [PIOPED 1990]. A defect in any of the lung segments in the perfusion scan is known as a segmental perfusion defect. A large segmental perfusion defect is one where more than 75% of the lung is affected with PE. A moderate segmental perfusion defect results when 25% to 75% of a lung segment is damaged and, when less than 25% of a lung segment is damaged, it is known as a small segmental perfusion defect. Reported studies for lung outline reconstruction using VQ scan images are limited in number and mostly appeared during the second half of the last decade. Nevertheless, some of these studies report significant progress towards achieving the goal of a lung outline reconstruction system. Gabor et. al. [1994], in their study, made use of an expert system to diagnose Pulmonary Embolism. The analysis performed by the expert system was on the basis of ventilation-perfusion scans only and excluded correlating chest xrays. Analyzing the abnormalities in the images required standardizing the images with respect to shape and size: creating a template of lung anatomy for all projections facilitated the standardization and the image from each projection was stretched onto the appropriate anatomic template. A pixel-to-pixel comparison was made between the stretched patient image and the composite normal file to identify the VQ defects within the images. Any pixel with a standard deviation value below 2.2 was considered as abnormal, and to label any area as abnormal required five or more contiguous pixels from that area. The stretching of the images caused mis-registration of the defect into different pulmonary segments on some views. This resulted in the same defect being counted twice and falsely increased the probability of Pulmonary Embolism. Research carried out by Armato et. al. [1997] presented a distinct method for reconstructing the outline of the lung in the ventilation and perfusion scans with the help of digital chest radiographs. They superimposed ventilation and perfusion images on the chest radiographs to determine the outline of the lung. This was done by identifying lung contours on all the images using an iterative global level thresholding scheme. Once the lung contours had been identified, scaling factors were obtained to appropriately match the dimensions of images from the two modalities. Alignment was achieved by means of vertical translation based on the apex locations and horizontal translations based on the mediastinum locations on the images. The result was a set of three superimposed images. Since this method makes use of the chest x-ray, the lung outline cannot be reconstructed from all views (right lateral, left lateral, right lateral oblique, etc.). Hasegawa et. al. [1998] made use of a shift invariant neural network to reconstruct the lung outline from the digital chest radiographs. The method of lung segmentation consisted of three stages, preprocessing for the background adjustment, the convolution network to extract the rib cage, and the post processing to smooth the boundaries as well as to reduce noise. The input images were miniaturized and then processed by the histogram equalization algorithm, which adjusts the gray tone distribution of the input image. The images were then processed by the trained neural network to enhance the boundaries of the lung fields. In post processing, the lung boundaries were extracted and the boundaries were smoothed using a boundary smoothing technique. This technique calculates the tangent change at each point on the boundary. The change of tangents at eroded areas is large whereas the change is small in smooth areas. Thus detecting a large change in the absolute value of the tangents can identify eroded areas. This method of lung segmentation leaves many eroded areas on the boundary of the lung image after reconstruction, which can be misinterpreted as a defect. Also, this method cannot be extended to views other than anterior or posterior since chest x-rays required for those views are not ordinarily available. Although above referenced algorithms appear to be promising to address the problem of lung outline reconstruction, we tend to believe there is merit in exploring an alternate algorithm for this problem. Hence, we propose principal component analysis methods since these algorithms naturally lend themselves to application in image reconstruction. There is vast and stable literature on principal component analysis. Specifics and particulars of this problem makes it possible to identify neural principal component analysis networks as potentially good paradigms to address the problem of outline reconstruction for VQ scan images of lungs. The ultimate goal for a completely automated image recognition system is to emulate the mental process a radiologist follows while analyzing the VQ scans and correlating chest x-ray to diagnose PE. The required elements of such an image recognition system consists of a subsystem to reconstruct outline of VQ images of lungs possibly distorted by defects due to PE, a subsystem to superimpose the segmented anatomy of lungs on VQ scans of dysfunctional lungs, a subsystem to evaluate the complete list of VQ defects, their sizes and location, and finally, another subsystem to correlate defects with chest x-ray images of the same lungs. The completely automated image recognition system would accept the set of ventilation-perfusion scan images and correlating chest x-rays as input, and generate diagnostic data as required by the PIOPED criteria for the interpretation/classification subsystem. The input image set includes ventilation-perfusion scan images of the lungs (anterior perfusion, posterior perfusion, right lateral perfusion, left lateral perfusion, right posterior oblique perfusion, left posterior oblique perfusion, equilibrium ventilation) and the correlating chest x-ray. The scope of this work will be limited to lung outline reconstruction through template recall for anterior perfusion scans. Functionality of the block, which reconstructs lung outline through template recall, involves storing a number of lung templates (anterior perfusion scans), and when presented with the image of a partly dysfunctional lung, recalling the closest matching template, thereby reconstructing the outline of the lung image through a best-fit approximation approach. We will employ a linear principal component analysis (PCA) neural network that will be trained with the generalized Hebbian learning rule. This network is a deceptively simple and easy to implement yet computationally powerful enough to solve the lung outline reconstruction problem. The algorithm will be tested using cases created using the PIOPED criteria. The recalled images or reconstructed outlines will be validated by a nuclear radiologist with the Medical College of Ohio (MCO). SIMULATION STUDY A simulation study was performed to test the performance of neural PCA algorithm on a set of actual human lung images, which were obtained from the Medical College of Ohio (MCO) patient repository. The software implementation of neural PCA algorithm was realized in MATLAB version 5.3.0. The healthy lung templates employed in the simulation, where three images are for males and eight images for females, are perfusion images in the anterior view and were chosen to model the variation and therefore to represent a good initial approximation to the general population as was also confirmed by a radiologist at the MCO. All the lung templates are stored in the bitmap format (.bmp) with a resolution of 128 120 × pixels. The pixel values of the lung image template range from 0 to 255: the images are gray level with 8 bit resolution. Pixel values are quantized to binary values –1.0 or +1.0 with a threshold value of 190 prior to processing them through the neural PCA algorithm. Test cases are generated in accordance with the PIOPED criteria: test cases vary according to the number of defects and the size and location of the defects in the lungs. Compliance of test cases with actual medical cases is also validated by a nuclear medicine radiologist. These images are created by adding segmental or sub-segmental defects in accordance with the PIOPED criteria to the original template lung images: all five probability cases are included in the test image set. The procedure for creating the test cases is as follows: a healthy lung image from the existing set of eleven templates is picked at random. The segmented perfusion lung image in the anterior view is mentally superimposed on the template. Next, the number and size of defects consistent with the PIOPED criteria for a particular probability class is created. Design of the neural network followed the consideration that the first principal component might be sufficient to extract the needed information. Since only one principal component was considered for each lung template image, the neural PCA network was instantiated with one output node [Oja, 1982; Oja and Karhunen, 1985]. The lung template images when converted into a data vector have 15360 elements: The lung template images consist of 128 120× binary pixels, which result in 15360 binary pixels when represented by a one-dimensional array. Thus, the number of input nodes for the neural PCA network was determined as 15360. The neural PCA network was trained with the generalized Hebbian algorithm. Elements of the weight vector were initialized to real numbers in the range [0, 1] using a uniformly random distribution using the MATLAB function Rand() available in version 5.3.0. The neural PCA network topology is as shown in Figure 1. Training the neural network for principal component analysis involves computing the principal components of each of the eleven lung templates. One template is presented to the network at a time and the network is allowed to stabilize; i.e. the elements of the weight vector stop changing following repeated adaptations through the learning rule. The network is guaranteed to stabilize if the learning rate chosen is not very high. The weight vector is updated using Equation 1, where the synaptic weights of this network are denoted by the term i w ,with 15360 ,..., 2 , 1 = i . The term ) (k w i represents the value of i element of weight vector at time instant k and is updated by the following formula: ( ) [ ] , 15360 ,..., 2 , 1 for ) ( ) ( ) ( ) ( ) ( ) ( = − = ∆ i k y k w k y k x k y k w i i i η (1) where ) (k wi ∆ is the change applied to the synaptic weight ) (k wi at time instant k, and η is the learning-rate parameter which is taken as 0.001 for this network. Once the network stabilizes i.e., + ∈ ∋ ≥ = ∀ + ∆ = ∆ Z m m k i k w k w i i and 15360 ,..., 2 , 1 ), 1 ( ) ( , the output of the network computed as shown in Equation 2 is the first principal component of the input lung template image vector and the weight vector is its corresponding eigenvector: