Computerized Tomography for Industrial Applications and Image Processing in Radiology March 15-17, 1999, Berlin, Germany DGZfP Proceedings BB 67-CD Paper 21 New developments in industrial radiography at EDF B. Lavayssière, A. Bonin, S. Gautier, Chatou, France INTRODUCTION To ensure security on nuclear power plants, Electricité De France (EDF) uses radiographic inspection for systematic pipe control. The extreme temperature and pressure conditions to which pipes are submitted may generate structural defects that must be detected and characterized. The control procedure consists in introducing a gamma-ray source inside the inspected pipe and in imaging radiation onto radiographic films located on its periphery. Radiograph testing consists in interpreting the acquired radiographs and in building up a diagnosis on the basis of both the information they provide and the past experience. Though advances have been made in digital information processing, most of radiographs are still performed with conventional films, especially due to limitations in the spatial resolution of the digital radiographic systems. As each point of a radiograph contains a projection of what the radiation has encountered in passing from the source through the material, the internal structures encountered are thus superimposed on each other in the acquired image, which complicates interpretation. So testers are facing the delicate task of establishing a complete diagnosis including the detection, the localization, the characterization and the accurate positioning of incoming faults within the thickness of the part being tested. Since 1998 EDF has been involved in the evaluation of the performance of RADView LIBERTY film digitizer and of software tools developed by its R&D Division, such as : - ENTRAIGUES software which performs 2D information processing, - SIROCCO-3D information processing software, which enables 3D reconstruction of the volume surrounding any faults from radiographic projections acquired at various angles (1, 2, 3), - MODERATO simulation software used for radiographic data simulation (4). RADView workstations and software provide a simple, easy-to-use PC-based solution for managing, displaying and analyzing radiographic images. The system includes a high-resolution monitor, image printer, flatbed document scanner, workstations, optical disk storage device and image digitizer (5). The overall quality of the digitization protocol has been evaluated. Firstly, by interpretation of the images obtained, testers can assess whether the digital radiograms can be compared to conventional ones. DGZfP Proceedings BB 67-CD 115 Paper 21
Secondly, to evaluate the image quality of the processing system, classical parameters as spatial resolution, contrast resolution, dynamic range, local and global distortion have been estimated. Specific guidelines for EDF film digitization procedures have been written, based on the digitization of a standard film (6). To determine the detail perceptibility of the imaging system, i.e. performance of digitization, an object with preliminary known geometry and with different levels of object contrast has been used (steel specimen composed of various electro-eroded slots). The obtained perceptibility depends on both the width and the depth of the slots. Conventional exposures of such an object, under controlled conditions, have been compared to digitized images in order to decide whether the expected information is in fact visible on the image. Herewith the information capability of the system has been determined (fig. 1). Conventional Interprétation des résultats Interprétation des résultats Interprétation des radiograms radiogrammes Without processing With processing 3 5% 1 23% 2 3 2% 1% 3 5% 1 44% 1. No defects 2. Defects 3. Errors 2 51% 2 72% 1 97% On the left, conventional interpretation from radiograms (classical double loading procedure) In the middle, single image interpretation without the use of any filters or contrast enhancement On the right, single image interpretation (all tools provided on the RADView system can be used) Figure 1 : Results of defects detection So single image digitization is not sufficient to detect very thin defects. That is why EDF uses «digital double loading», which means : - digitization of the two radiograms coming from the same exposure (double loading procedure), one after the other, - registration of the images, - combination of the images. The obtained image has increased signal to noise ratio and a very good quality that helps for interpretation. IMAGE PROCESSING ENTRAIGUES image processing tools have been developed by EDF Research Division, in order to provide a non-destructive testing software for identifying defects and characterizing their evolution. But image data suffer from very poor quality, exhibiting such characteristics as : - systematic signal drift with arbitrary orientation, - high level of noise, - important variability in defects size and shape, - blurred effects along defects edges. Because of the poor quality of the digitized images (i.e. trend with any orientation, high noise ratio) and of the defect characteristics (i.e. variable number, shape and size, edge fuzziness), image processing and segmentation problem are very difficult to solve. Since defects correspond to area with local mean lower than background mean, the drift tends to prevent any image interpretation and, therefore, must be removed. A preprocessing scheme has been proposed for drift removal and edge enhancement DGZfP Proceedings BB 67-CD 116 Paper 21
(fig. 2). Figure 2 : Level flattening The main difficulty after drift removal is due to the characteristics of noise. This noise seems to be distributed roughly according to a Gaussian law with a very high standard deviation. Indeed, noise deviation is of the same order as the defects characteristics. The high level of noise, added to large variability in defects size and shape, explain the systematic failure of standard filtering approaches. So noise filtering is described within a theoretical framework based on a multiscale analysis of the image structure (fig. ). Figure 3 : Noise reduction Image analysis and interpretation require extracting such primitives as edges and/or regions. Applied to noisy images, segmentation proves to be an ill-posed problem, since noise tends to generate spurious edges. Finally, ENTRAIGUES uses an approach based on a multiscale analysis allowing to achieve simultaneously defect-enhancement and noise-filtering and leading to accurate defects segmentation using mathematical morphology. The developed method has been tested on a set of radiographs including various defect patterns, leading to impressive segmentation results.(fig. 4). Figure 4 : Quantitative measurements (automatic segmentation, recognition of individual defects and image analysis) 3D RECONSTRUCTION 3D reconstruction uses several radiographic projections acquired from different angles in order to reconstitute volumes surrounding any faults (fig. 5). 3D technique uses a specific test protocol, in order to guarantee the desired degree of precision. DGZfP Proceedings BB 67-CD 117 Paper 21
x y z S S S Object to be examined 3D Reconstruction 3D Display Fault Diagnosis Support R 1 R 2 R 3 Digitization R i Radiographs containing fault indications S Gammagraphic source On-site acquisition Figure 5 : Principle of 3D acquisition Off-line expert assessment The main objective is to estimate the flaw orientation of the studied specimen. This information makes sense to experts, with regard to the deterioration rate of the flaw. This technique has been evaluated in metal parts radiograph testing with a considerable thickness (60 mm and over). For this reason, the offset shots angle was limited to around 20, given that greater angle would have given radiographs which, being too dense, could not be exploited. In practice, 5 to 7 shots are needed, to achieve accuracy to within 2 mm. So a specific test protocol is used in order to guarantee the desired degree of precision. This situation is similar to the well-known ill-posed problems in mathematics that are solvable only by introducing a priori information on the solution. Because information is not entirely contained in available data, this 'a priori' information can be gathered from external knowledge and thus incorporated into the problem solving procedure. Sources of a priori information can be various : data constraints (e.g. image values are positive), results from a finite elements software, hints about noise nature (e.g. noise is Gaussian or Poisson-like), knowledge about materials or degradation types (e.g. cracks are likely to be aligned). The second step is to convert prior information into an appropriate function in order to deal with both raw data and prior information the same way. It can use deterministic (e.g. smooth functions) or stochastic (e.g. Gaussian distributions) tools. Some classical algorithms allow us to obtain solutions even with missing projections. Application of these methods leads us to a discretisation of the reconstructed object, that is described by small volume-elements (voxels) within which the attenuation function takes a constant value. Such a system can be written as : where p = Af + n p represents the available projections, A the projection matrix containing the intersection between rays and voxels, n the noise on data due to radiographic artifacts, f the unknown values of the radiographed object. A solution can be achieved by a classical iterative A.R.T. - type method (Algebraic Reconstruction Technique) which evaluates iteratively the unknown image function representing the local attenuation properties of the object. Its principle is to estimate the image function in a voxel for the (k+1) th step from the estimate at the k th step, by an update-correction procedure including the error between the correct projection value DGZfP Proceedings BB 67-CD 118 Paper 21
and the projection calculated at the same point from the estimated projection after k steps. One major advantage of this approach is to allow the introduction of general constraints. An a priori constraint (positivity) can be introduced because all densities of the reconstructed object are non-negative. Furthermore we introduced a support constraint. This is done by delimiting on the 2D radiographs geometric zones around the detected default and by calculating the 3D volumic intersection of defined cone-projections. This allows us to restrict severely the 3D space we have to reconstruct. The first step of this algorithm provides us a rough reconstructed tridimensional zone containing the flaw. Then the main idea of our approach is to apply a restoration to this 3D reconstructed zone containing the defect in order to reduce the reconstruction noise and thus to obtain a better estimate of the flaw. We introduced some geometrical and physical knowledge about the default and the experimental conditions by using a Bayesian method based on a modelisation by a Markov Random Field (MRF). The 3D restored zone provides some information about the default : this knowledge becomes a support constraint that is injected again as a reconstruction constraint. Since the real process of image acquisition introduces noise of various origins, experiments were later performed on real radiographs of perfectly identified steel parts with electro-eroded faults (fig. 6), and later with shrinkage, and were conducted under the same conditions as those applied to piping. Radiograph Reconstruction Figure 6 : 3D reconstruction of electro-eroded notch Complementary assessments (such as destructive examination and ultrasound) carried out on these parts pointed out the good match in fault localization with 3D reconstruction from radiographic images, as long as prior information remains realistic, and confirmed the quality of the results obtained. The quality of the results obtained with these methods based on applied mathematics proves that they are useful in the context of non-destructive testing, and more particularly in industrial radiography. MODERATO MODELLING MODERATO is a radiographic simulation program that simulates a radiographic process from the source to the detector according to a microscopic physical model. Following EDF's requirements sources are Iridium and Cobalt ones with energy ranges from 0.3 to 1.3 MeV. Radiographed object and flaw geometry are described by CAD models. The physical process has been extended to take into account different media simultaneously. Detectors consist in films and Pb screens. DGZfP Proceedings BB 67-CD 119 Paper 21
Figure 7 : Simulated images with various CAD descriptions The physical model for the photons takes into account Compton and Rayleigh scattering and photoelectric absorption in the material. Up to now pair effect is neglected because it has little incidence in our energy range. The model is a little more complex in the detector because electrons are also taken into account. Electrons come from Compton scattering because after a Compton interaction an electron is emitted absorbing part of photon s energy. They do not interfere in the object but are responsible for the image creation. Photons and electrons behaviors are computed in the films and screens. Photons encounter the same interactions as in the object and electrons encounter ionization, Bremsstrahlung and Coulomb scattering. Bremsstrahlung can be neglected though because energy loss are very small compared to ionization in this energy range. Monte-Carlo simulation implies that we compute photons and electrons behaviors one after the other with all the physical interactions and the objects, flaws, screens and films intersections. So Monte-Carlo simulation is a very time-consuming process. That is the reason why EDF s radiographic simulation MODERATO program, integrates optimization techniques to decrease execution time. Optimization in Monte-Carlo simulations is a major issue and has lead to numerous publications. Optimization is achieved by modifying the physical process and by correcting the result with statistical weights. Photoabsorbed photons do not contribute to the image so it is possible not to take into account this interaction. Each photon s weight decreases according to the probability distribution of the neglected interaction. Photons are killed when their weights become to low. This creates a bias that can be corrected by Russian Roulette technique. This classical technique requires to determine the value of at least two parameters: the weight s threshold and the Russian Roulette parameter. The efficiency of the optimization depends on these two empirical parameters. That is why we developed a new complementary technique that does not require any empirical parameter. We decided to study all the empirical probability distributions of the simulated radiographic process. We noticed that the statistical laws were different if we considered only the photons that exit the object. So we proposed to pre-simulate to estimate these empirical laws and then to simulate the actual image with a physical process that follows the new laws. We correct the results with the appropriate statistical weights. This allows to obtain the same image with a more efficient process so in less time. DGZfP Proceedings BB 67-CD 120 Paper 21
1 5 9 Computerized Tomography and Image Processing, 1999 14 12 10 Number of photons 8 6 4 Optimized Monte- Carlo simulation (10 M photons) "Classical" Monte- Carlo simulation (20 M photons) 2 0 13 17 21 25 29 33 37 41 45 49 53 57 61 65 69 73 77 81 85 89 93 97 Cross-section pixel Figure 8 : Simulated notch -Radiant images profiles The images are qualitatively the same. The analysis of image profiles shows that it is difficult to distinguish between a classical simulation and a new optimized simulation. CONCLUSION From this paper, it is clear that results from test pieces trials or other relevant experimental data provide valuable evidence on inspection capability for use in a technical assessment. But another method for generating evidence is to use the predictions of mathematical models. It is cheaper to run a model than to conduct test piece trials. It is also quicker, because no additional delays are induced by specimen manufacturing. So modeling could be used before any other tests, for inspection capability, but also for assessment of 2D or 3D processing tools. REFERENCES 1. C. Klifa, «Reconstruction 3D d'objets à partir d'un nombre très limité de projections. Application à la radiographie industrielle» PhD Thesis, Télécom Paris, 1991 2. B. Lavayssière, «Reconstruction en 3D : un nouveau défi en radiographie industrielle», 6th European Conference on Non Destructive Testing, Nice, 1994 3. C. Klifa, K. Sauer, «Bayesian Tomographic Reconstruction of 3D Objects from Limited Numbers of Radiographs» Proc. IEEE 1992 Nuclear Science Symposium & Medical Imaging Conf., Orlando FL 4. B. Lavayssière, A. Bonin, S. Gautier «3D reconstruction of faults in industrial radiography» Proc. International Conference Computer Methods and Inverse problems in nondestructive testing and diagnostics, Minsk, 1998 5. «RADView workstation software package : user manual», LIBERTY 6. «Digital radiographic image storage and retrieval», NDE Center, EPRI TR-104626, 1995 DGZfP Proceedings BB 67-CD 121 Paper 21