A common parasitosis in South Brazil and other tropical and subtropical environments is the neurocysiticercosis (NC). It has been shown that 26% to 35% of all patients with partial epilepsy have Intracranial Cysticercus as an etiology, thus being this a major cause of epilepsy in development countries.
Before the use of the computeed tomography (CT) scans was widespread, only 20% to 30% of patients with partial epilepsy were suspected of having an intracranial organic etiology. The tomographic investigation on these patient groups demonstrated the existence of a high rate of secondary epilepsy, which can be up to 55% – 65% [2,3].
Cysticercus cellulosae is the larval form of the Taenia solium, one of the most common intestinal parasitic diseases. In fact, it seems to be the main epilepsy aetiology and is the most common parasitic disease of the central nervous system in the development countries [4,5]. The diagnosis can be performed trough clinical history and epidemic discoveries and ratified with complementary exams, such as the analysis of the brain- and spinal liquor and neuroimagery. Reliable diagnosis confirmation is considered to be only possible through biopsy [7,8,9].
With the recent progress of the neuroimagery, CT has become the most reliable method for the diagnosis of NC [10,11]. It has also the advantage of informing about the activity or not of disease [10,11]. The practice of neuroradiological diagnosis of calcified lesions or cystic findings in the cerebral parenchyma through CT has shown that neuroradiological findings cannot, in certain situations, correspond exactly to the anatomopathological result of necropsy and there are some differences between tomographic NC diagnosis and anatomopathological findings, as demonstrated in experiments using pigs showing NC [10,11,12]. Those differences may have occurred due to difficulties for the medical staff to discriminate NC calcifications appearing in contiguous slices as one or two lesions, as well as due to overlooked small calcifications.
In the scope of this research, we developed a computer vision method for the detection of NC-related lesions. It was implemented as a graphic software tool to assist radiologists in performing better, more reliable and simpler diagnosis. The system counts and measures the calcifications related to NC in CT scans, reducing errors regarding visual inspection. The system increases diagnosis quality through:
providing a correlation of findings in different slices of the same volume, automatically relating findings belonging to the same lesion, reducing lesion counting errors during visual inspection
providing better quantitative data through automated measurement of cysticercuses volumes
providing data about the position of lesions (in CT coordinates).
Objectives
The research work described in this paper was performed in the scope of the Cyclops Project [13], which is a Brazilian-German Cooperation Project aimed on the development of intelligent software tools for the support of radiological diagnosis. The aim of this particular work was to develop a method for the automated detection of NC-suspect areas in CT scans found in common hospitals. The requirements for this method were stated together with radiologists and neurologists participating on the Project. The main target was that this method should be usefull to implement a tool to facilitate the identification, counting and measurement of calcifications related to NC in CT scans, mainly reducing errors and providing better quantitative data. There were 4 general objectives that were taken into account during development of this method:
It should enable the reliable detection of NC-suspect areas, slice-by-slice, highlighting them in different colours on the scan images. Another, non-NC related calcifications, should not be highlited. This discrimination should be performed automatically.
Since the association of findings that appear in more than one slice to the same lesion seems to be one of the most important error sources in counting of lesions, the method should enable a spatial analysis of the whole CT volume after classification of NC-suspect areas, associating automatically findings in different slices that could pertain to the same lesion.
The method should enable the handling of both, modern helicoidal scans and also data generated in development countries´ hospitals, taken with old CT scans with slice thicknesses from up to 1 cm.
It should be possible to provide data about the localisation of NC-suspect areas, thus providing information for further studies relating NC and epilepsy.
Besides implementation and tests of this tool as a software by itself, we also performed a first evaluation of the performance of the system through the a) analysis of the agreement among 3 specialists in quantifying the number of calcifications and b) the comparison of these with the results provided by our tool.
Methods
The data used in this work were CT scans showing single and multiple NC calcifications from eighteen patients. All data belonged to case studies from the case collection of the University Hospital of the University of Santa Catarina. Patients were originally examined at the outpatient Multidisciplinary Clinic for Epilepsy at the Brazilian National Health Service in Florianópolis. All scans were taken on a Toshiba CT scan at the Florianópolis Public Hospital. All patients were showing partial epilepsy symptoms and had NC as presumed aetiology. The images were processed through the system in 4 steps: (a) image segmentation, slice by slice, using a specific region-growing technique, (b) the segmentation results were classified through neural networks [14] for NC determination and (c) the segments classified as NC from different slices considered to represent the same finding were automatically associated, and (d) the resulting data were used to perform a 3D reconstruction aiming to show NCs spreading through more than one slice and to measure the volume of each NC finding. The results can be viewed slice by slice or as 3D-reconstruction.
Different segmentation algorithms were tested and the best results were obtained with the algorithm of Mumford & Shah [15]. Results obtained were considered adequate for the purpose of this work (see figure 2). A segmentation based on an elastic energy model, such as the Mumford & Shah method has strong advantages over more widespread watershed methods in segmenting NC-suspect brain areas, mainly because the elastic boundary of regions leads to a better cohesion of NC areas, which tend to be round.
These segments were submitted to a pre-classification based on simple criteria like mean greyvalue of segments, were only strong NC candidates were left. After this pre-classification, segments were submitted to the classification through a Backpropagation Neural Network [14]. Experiments have shown that even a simple network with a hidden layer of 10 neurons is capable of a reliable classification performance.
For the first tests described here, the neural networks were trained with feedback information from different neurological staff, who provided a by-hand classification of segments of different patients. To train the networks, the resulting segments had to be described adequately in terms of parameters. The data used for describing NC candidate segments for the neural network classification were pattern vectors consisting of segment parameters such as median greyvalue, variation coefficient of greyvalues, gravity center of the segment and segment area (in CT scan coordinates), greyvalue standard deviation and some special developed shape parameters. For the development of these description criteria we based on earlier work performed also in the field of classification of regions of radiological images [16] [17]. After finishing the classification, the system marks NC findings, replacing automatically all areas in the original slices classified as NC through specially coloured areas, each with a different colour.
Ater finishing the NC classification on all slices, the system starts performing a 3D reconstruction of each finding, searching through the slices for NC-classified areas belonging to the same finding (see figure 3). On the 2D-visualisation window, finding colours are changed, thus leaving areas in different slices belonging to the same finding coloured with the same colour.
As a final step, the system performs a 3D reconstruction of the patient’s skull, encephalic mass and findings boundaries, generating a 3D representation of the patient’s head and the localisation of NC findings. Volumes of each NC finding are also calculated. This 3D reconstruction is generated using the VRML [18] virtual reality representation language, being able to be viewed on any virtual reality browser or renderer, avoiding the use of special hardware or software.
For a first validation of the system accuracy, discrepancies among classification and NC counting results from 3 different specialists and system performance where compared. For the performance analysis, the Anova multiple-variance test was used.
Results
For the 18 patients os our first test set the system performed correctly for NC detection and classification. All areas containing NC-suspected regions of the all patients were segmented and classified correctly. The trained neural network was able to classify correctly 100% of all images used in this first tests, even given the low-quality data provided by the CT scan used. Calcifications due to other aetiologies were also correctly classified as non-NC through the system.
The three-dimensional reconstruction of the affected areas was also performed correctly in all cases examined to date. Presently the system shows a three-dimensional reconstruction of the patient’s head with the NC highlighted.
Examining classification accuracy, we found the following mean and standard errors among specialists: specialist 1 (2,5 +/- 1,23), specialist 2 (3,16 +/- 1,41), specialist 3 (3,22 +/- 1,32) and the software (3,3 +/- 11,17), without statistic significance (p > 0.9).
Figura 1
The software tool, which is being refined for better usability now, runs on all plattforms, both UNIX and MS Windows, and can communicate over a network directly with DICOM image databases and CT scans that are conform to the DICOM standards.
Conclusions and Future Work
Our results showed that the system performed slightly better in comparison with specialist results. Some discrepancy happened among the specialists when there were more than 2 lesions in the CT scans. Results were discussed in a feedback session among all specialists and software results were accepted as best results due to the better correlation of different findings in contiguous slices to the same lesion that was performed by the software. Comparison of results obtained by the specialists also has confirmed that discrepancies between manual diagnoses occur mainly where one lesion appears in more than one slice.
The next steps in the development of this approach are: a) more accurate validation of the performance of the system with a larger data set and b) development of a deformable digital brain atlas based on the Thalairach Atlas for the automated localisation of findings. One field which to date lacks on reliable statistical data is the correlation of the position of NC lesions and the symptoms and form of epilepsy shown by the patients. Implementing this atlas, we will be able to perform an automated localisation of each finding, automatically determining which brain areas are affected and be able to collect more accurate data on the correlation of localisation of lesions and epilepsy. First implementation studies on this atlas basing on a deformable anisotropic octree structure are already being performed.
References
Narata AP et al. “Neurocysticercosis: a tomographic diagnosis in neurological patients.” Arq Neuropsiquiatr 1998; 56:245-9.
Machado LR, Nobrega JP, Barros NG, Livramento JA, Bacheschi LA, Spina-Franca A. Computed tomography in neurocysticercosis: a 10-year long evolution analysis of 100 patients with an appraisal of a new classification. Arq Neuropsiquiatr 1990; 48:414-8.
Carpio A., Escobar A. Hauser W.A. Cysticercosis and Epilepsy: a Critical Review. Epilepsia 1998; 39:1025-42.
Trevisol-Bittencourt P.C., da Silva N.C., Figueredo R. Prevalence of Neurocyticosis among epileptic in-pacients in the west os Santa Catarina-southen Brazil. Arq Neuropsiquiatr, 1998; 56:53-8.
Svetlana A , Yela DA, Gomes AE. Edema Cerebral crônico na Neurocisticercose Arq Neuropsiquiatr1998; 56:369-76.
Agapejev S. Epidemiology of neurocysticercosis in Brazil. Rev Ins Med Trop S Paulo 1996; 38:207-16.
Spina-França A. Síndrome liquórica da neurocisticercose. Arq Neuropsiquiatr 1961; 19:307-14.
Bouilliant-Linet E, Brugieres P, Coubes P, Gaston A, Laporte P, Marsault C. Diagnostic value of x-ray computed tomography. Apropos of 117 cases[Cerebral cysticercosis]. J Radiol 1988; 69:405-12.
Shraberg D, Weisberg L, de Urrutia JR, LaCorte WS. Cysticercosis cerebri: evolution of central nervous system involvement as visualized by computed tomography. Comput Tomogr 1980; 4:261-6.
Aluja AS, Gonzalez D, Rodriguez Carbajal J, Flisser A. Histological description of tomographic images of Taenia solium cysticerci in pig brains. Clin Imaging 1989; 13:292-8.
Gonzalez D, Rodriguez-Carbajal J, Aluja A, Flisser A. Cerebral cysticercosis in pigs studied by computed tomography and necropsy. Vet Parasitol 1987; 26:55-69.
Berman JD, Beaver PC, Cheever AW, Quindlen EA. Cysticercus of 60-milliliter volume in human brain. Am J Trop Med Hyg 1981; 30:616-9.
v.Wangenheim, A.; Barreto, J. M.; Richter, M. M.; Krechel, D.: Cyclops – Expert System Shell for the Development of Applications in the Area of Medical Image Analysis, in: Jähnichen; Lucena (Eds.): Proceedings of the 4th German-Brazilian Workshop on Information Technology. Edited by: DLR – German International Bureau for the Federal Ministry of Education, Science, Research and Technology and CNPq – The National Council for Scientific and Technological Development, Porto Alegre/Berlin, 1997.
Rummelhart, D.E., Hinton, D.E., Williams, R.J.: Learning Representations by Back-Propagationg Errors. Nature 323, pp. 533-536, 1986.
Munford D., Shah J.; Optimal Approximations by Piecewise Smooth Functions and Associated Variational Problems, Comm. Pure Appl. Math., 1989.
Krechel, D., Hess, F., Comes, F., v.Wangenheim, A., Blasinger, K.: Mammalyzer II: A Decision Support System for Early Detecion of Breast Cancer in Contrast Enhanced MRI, in Bildverarbeitung fuer die Medizin 1998, Aachen, Germany, Springer Informatik Aktuell, 1998.
Krechel D., v.Wangenheim, A.; Automatic Registration of MRI Head Volumes in “Automatische Analyse von Tomographie-Daten”, Institut für Medizinische Statistik und Epidemiologie der TU München, April, 1996.
Marrin, C., Campbell, B.: VRML 2. Sams.Net Publishing, Indianapolis,USA, 1997.
