Artículos
Assessment of methods for volumetric quantification of intracerebral hematomas in computerized tomography images
Evaluación de métodos para la cuantificación volumétrica de hematomas intracerebrales en imágenes de tomografía computarizada
Assessment of methods for volumetric quantification of intracerebral hematomas in computerized tomography images
Archivos Venezolanos de Farmacología y Terapéutica, vol. 37, no. 4, pp. 331-335, 2018
Sociedad Venezolana de Farmacología Clínica y Terapéutica
This work is licensed under Creative Commons Attribution-NonCommercial-NoDerivs 4.0 International.
Abstract: This work evaluates the performance of computational methods aimed at volume generation of five intracerebral hematomas (ICH), present in multi-layer computed tomography images. First one, a ground truth volume or reference volume (RV) is considered. This RV is obtained, by a neurosurgeon, using the manual planimetric method (MPM), which allows the generation of manual segmentations of space-occupying injuries which, in this case, are matched with the ICH. The MPM consists in: a) Manually drawing the contour that delimits the ICH, in each of the layers or cuts in which the ICH is present b) Calculate the partial area using the number of pixels contained in each contour c) Obtain the total area of the ICH by adding the partial areas d) Calculate the volume considering the total number of layers, the total area of the ICH and the thickness of each layer. In second place, the volumetry of the 5 ICH’s is obtained considering both the original version of the ABC/2 method and two of its variants, identified in this paper as ABC/3 method and 2ABC/3 method. The ABC methods allow for calculating hematoma volume under the geometric assumption that the ICH has an ellipsoidal shape. In these methods, A is the maximum diameter of the ICH, B is the length of the ICH measured, perpendicularly, with respect to the parameter A; while C represents the product of the thickness of the cut by the number of cuts in which the ICH is present. In third place, a smart automatic technique (SAT) is implemented to generate the three-dimensional segmentation of each ICH and from it the volume of the hematoma is calculated by multiplying the voxel dimensions by the number of voxels that make up the ICH. In the context of the present work, the expression SAT method will be used to refer to the new methodology that is proposed to calculate the volume of the ICH. The SAT consists of the pre-processing, segmentation and post-processing stages. During pre-processing, a thresholding algorithm and a bank of computational filters are used to address artifact and image noise problems. In segmentation, the growth of regions is applied to pre-processed images. Finally, a morphological dilation filter is used as a technique to perform the post-processing of the segmented images. In order to make judgments about the performance of the SAT, the Dice coefficient (Dc) is used to compare the dilated segmentations of the ICH with the ICH segmentations generated, manually, by a neurosurgeon. The combination of parameters linked to the highest Dc, allows to establish the optimal parameters of each of the computational algorithms that make up the SAT. Finally, the percentage relative error is calculated as a metric to evaluate the methodologies considered. The results show that the SAT method exhibits the best performance, generating an average percentage error of less than 3%.
Keywords: ABC Methods, Intelligent Automatic Technique, Segmentation, Intracerebral Hematoma Volumetry.
Resumen: Mediante este trabajo se evalúa el desempeño de algunos métodos orientados hacia la generación del volumen de cinco hematomas intracerebrales (ICH), presentes en imágenes de tomografía computarizada multicapa. Para ello, en primer lugar, se considera como volumen de referencia el obtenido por un neurocirujano usando el método manual planimétrico (MPM) el cual permite generar segmentaciones manuales de lesiones ocupantes de espacio, que en este caso se hacen coincidir con los ICH. El MPM consiste en: a) Trazar manualmente el contorno que delimita el ICH, en cada una de las capas o cortes en las que el ICH está presente. b) Calcular del área parcial usando el número de píxeles contenido en cada contorno. c) Obtener el área total del ICH mediante la suma de las áreas parciales. d) Calcular el volumen considerando el número total de capas, el área total del ICH y el espesor de cada capa. En segundo lugar, la volumetría de los 5 ICH es obtenida considerando tanto la versión original del método ABC/2 como dos de sus variantes, identificadas en este trabajo como método ABC/3 y método 2ABC/3. Los métodos ABC, permiten calcular el volumen del hematoma bajo la suposición de que el ICH tiene forma elipsoidal. En estos métodos, A es el diámetro máximo del ICH, B es la longitud del ICH medida, perpendicularmente, respecto al parámetro A; mientras que C representa el producto del grosor del corte por el número de cortes en los que está presente el ICH. En tercer lugar, se implementa una técnica automática inteligente (SAT) que genera la segmentación tridimensional de cada ICH y a partir de ella se calcula el volumen del hematoma multiplicando las dimensiones del vóxel por el número de vóxeles que conforman el ICH. En el contexto del presente trabajo, se utilizará la expresión método SAT para hacer referencia a la nueva metodología que se propone para calcular el volumen de los ICH. La SAT consta de las etapas de pre-procesamiento, segmentación y pos-procesamiento. Durante el pre-procesamiento, se emplea un algoritmo de umbralización y un banco de filtros computacionales para abordar los problemas de artefactos y ruido de las imágenes. En la segmentación, el crecimiento de regiones se aplica a las imágenes pre-procesadas. Finalmente, un filtro de dilatación morfológica se usa como técnica para realizar el pos-procesamiento de las imágenes segmentadas. Con el propósito de emitir juicios de valor acerca del desempeño de la SAT, se utiliza el coeficiente de Dice (Dc) para comparar las segmentaciones dilatadas del ICH con las segmentaciones del ICH generadas, manualmente, por un neurocirujano. La combinación de parámetros vinculada con el Dc más elevado, permite establecer los parámetros óptimos de cada una de los algoritmos computacionales que conforman la SAT. Finalmente, el error relativo porcentual es calculado como métrica para evaluar las metodologías consideradas. Los resultados muestran que el método SAT exhibe el mejor desempeño generando un error porcentual promedio inferior al 3%.
Palabras clave: Métodos ABC, Técnica automática inteligente, Segmentación, Volumetría de hematomas intracerebrales.
Introducción
In the clinical context, the use of digital brain neuroimaging allows for the diagnosis, approach and monitoring of diseases that affect the anatomy and/or physiology of the human brain. One of them, which is of special interest for the present work, is the pathology called intracerebral hematoma (ICH) that is characterized by the rupture of intracerebral blood vessels with extravasation of blood to the cerebral parenchyma1. The ICH forms a mass, usually oval, which can compress the adjacent brain tissue2. In addition, particularly, spontaneous non-aneurysmal ICHs are usually located in the ganglia at the base of the brain and are primarily due to inadequately controlled arterial hypertension1.
Additionally, it is important to note that digital brain neuroimages are accompanied by various imperfections such as noise3,4 and artifacts5. These imperfections become real challenges, when computational segmentation strategies are implemented oriented towards the generation of the morphology (normal or abnormal) of both the anatomical structures of the brain and of space-occupying lesions, such as, for example, hematomas6. Figure 1, generated based on multilayered computed tomography (MSCT) images, presents axial views of the 5 ICH that were considered in this work. In addition, it is observed the presence of the main problems typical of this type of image linked to noise (Poisson) and artifacts (Staircase and Partial Volume).
Figure 1.
On the other hand, the most relevant attribute or predictor of an ICH is its volume. The reason why this attribute is so important is that its numerical value defines, in a high percentage, both the prognosis of the patient and the behavior to follow to address this disease7. Due to this, some methodologies oriented towards the estimation of said predictor have been reported in the literature. Such methodologies are described below.
Yildiz et al.8, estimate the volume of 193 ICH present in brain MSCT images and use it as a fundamental predictor to establish the average number of deaths that occur in hospitalized patients suffering from this pathology. These researchers use the ABC/2 method. In this method, the axial view of that layer is used where the ICH exhibits its largest diameter which represents the parameter A. On the other hand, B is made to coincide with the diameter of the ICH perpendicular to the diameter; while C is the product of the thickness of the image by the number of cuts in which the ICH is present9.
Additionally, Rodríguez et al.10, study the inter-subject variability that occurs when estimating the volume of 40 ICH using the computer-assisted planimetric method. The time spent by the 5 specialists to segment the 40 databases exceeded 5 hours, which undoubtedly reflects the cumbersome nature of the manual method.
On the other hand, this article constitutes an extension of the work presented in reference6. The main contributions of the present work are:
a) Use an intelligent automatic technique (SAT) to calculate the volume of the ICH, present in 5 databases formed by three-dimensional brain images of MSCT. The aforementioned technique considers the stages of pre-processing, segmentation and post-processing. These stages are subjected to a validation process that uses the Dice coefficient to compare ICH segmentations obtained automatically and manually6.
b) Consider the percentage relative error (PrE) to perform a comparative study between the ABC methods and the SAT method, in such a way that their performance can be established when they obtain the volume of the ICH (Av). During the comparison, the one obtained by the planimetric manual method (MPM), applied by a neurosurgeon, is taken as the reference volume (Rv). The percentage relative error is calculated using the mathematical model given by equation 1.
Equation 1
MATERIALS AND METHODS
Description of the databases
The databases (DB) used were provided by the Central Hospital of San Cristóbal-Táchira-Venezuela, were acquired through the modality of MSCT and are constituted by three-dimensional images (3D), corresponding to the anatomical structures present in the head of 5 male patients. Their numerical characteristics are presented in table 1.
Table 1. General characteristics of the databases used in the present work.
| DB label | Voxels number | Voxel dimensions (mm3) | Age (years) |
| DB1 | 512x512x50 | 0.4883 x 0.4883 x 2.7483 | 65 |
| DB2 | 512x512x24 | 0.6211 x 0.6211 x 3.1139 | 69 |
| DB3 | 512x512x30 | 0.4453 x 0.4453 x 4.8300 | 17 |
| DB4 | 512x512x40 | 0.4766 x 0.4766 x 3.2985 | 75 |
| DB5 | 512x512x40 | 0.3964 x 0.3964 x 3.6785 | 54 |
As table 1 reveals, high variability in voxel size is observed in a group of 5 patients whose age ranges between 17 and 75 years. In a complementary way, manual segmentations are available, generated by a neurosurgeon, corresponding to the hematomas present in the DB considered. These segmentations represent the ground truth that will serve as a reference to validate the results linked with the segmentations.
Smart automatic technique (SAT) for ICH segmentation.
In figure 2, a schematic diagram synthesizes the computational algorithms that makes up the SAT. For a detailed description of the SAT, reference6 should be revised, since, as indicated above, this article is an extension of that reference.
Figure 2
Block diagram of the smart automatic technique (SAT) proposed in6.
On the other hand, it is necessary to point out that the Dice coefficient (Dc)3 is a metric used to compare segmentations of the same 2D or 3D image, obtained by different methodologies. In the medical context, usually, the Dc is considered to establish how similar spatially, manual segmentation (RD) and automatic segmentation (RP) are, at generating the morphology of any anatomical structure. Additionally, the Dc gets its maximum value when a perfect overlap between RD and RP is reached but it is minimal when RD and RP do not overlap at all. In addition, the values expected for the Dc are real numbers between 0 (minimum) and 1 (maximum). Equation 2 gives the mathematical model that defines the Dc.
Clinical utility of the volumes occupied by the hematomas
The main clinical utility of the characterization of hematomas by obtaining the volume lies in the decision making that is made to establish the behavior to be followed to address the presence of bruising in a patient. In this regard, patients whose lesions meet any of the following criteria9,11 must be taken to the surgery:
1- Lesion located in the anterior or middle cranial fossa with volume greater than 30 cm3.
2- Displacement of the midline (imaginary line between occipital eminence and crista gally) greater than one cm, from its original position.
3- Compression, displacement or occupation of specific brain areas (mass effect).
4- Lesion located in the posterior fossa (cerebellum, stem) with a volume between 10 cm3 and 15 cm3 depending on the clinical patient situation.
Quantification of hematoma considering the determination of its volume
Measuring the volume of bruises is important to define the final behavior before the process by which the patient passes. The volume and behavior of the lesion define parameters called surgical criteria, which are fundamental during treatment.
Obtaining volumes related to automatic segmentations
The proposed technique generates the automatic segmentation of the ICH present in each of the 5 databases described. From such segmentations, the volume of the hematoma, candidate to be characterized, is calculated by multiplying the voxel dimensions by the number of voxels that make up the automatically segmented ICH.
RESULTS
Quantitative Results
During the segmentation process, it was applied as a criterion that the optimal parameters of the algorithms that make up the SAT are those that produce the highest Dc. At the end of the tuning process, a maximum Dc of 0.8659 was obtained, which indicates a good correlation between the manual segmentations and those obtained by the SAT. Additionally, table 2 shows that the average value of the Dc obtained for the segmentation of the ICH, using the SAT method, is comparable to that reported in references12,13.
| Authors | Technique | Modality | Average Dc |
| Kamnitsas et al(2017)12 | Convoluting neural networks | MSCT | 0.8917 |
| Prakash et al(2012)13 | Regularized level sets | MSCT | 0.8432 |
| Vera et al. (Proposed technique in the current article) | SAT | MSCT | 0.8654 |
Qualitative Results
Figure 3, shows a 2-D view of both the original ICH and the processed versions after applying the SAT technique to one of the DB considered.
Figure 3
On the other hand, figure 4 shows an excellent three-dimensional representation of the segmented ICHs, corresponding to all the databases used in the present investigation.
Figure 4.
In figure 4, it can be seen that this type of hematoma does not have a defined shape and therefore, in general, it can be said that the geometric hypothesis considered by the ABC methods to estimate the ICH volumes is not always valid. In addition, in a study with 83 patients conducted by Huttner et al.14, only 44% had housings with ellipsoidal shape in the parenchymal tissue, that is, 66% of the patients, considered in this study, had ICH with a non-ellipsoidal shape.
On the other hand, table 3 shows the values for the volume calculated considering the automatic segmentation of the ICH, that is, using both the SAT method and the ABC methods.
Table 3. Values obtained for the volume occupied by each of the segmented hematomas.
| Volume (cm3) | |||||
| Database | MPM | SAT | ABC/2 | ABC/3 | 2ABC/3 |
| DB1 | 2.01 | 2.53 | 2.45 | 1.63 | 3.27 |
| DB2 | 39.77 | 40.52 | 42.84 | 28.56 | 57.12 |
| DB3 | 26.49 | 26.87 | 27.92 | 18.61 | 37.23 |
| DB4 | 6.91 | 7.56 | 8.07 | 5.38 | 10.76 |
| DB5 | 27.89 | 28.09 | 29.47 | 19.65 | 39.29 |
It can be inferred from the information, presented in table 3, that SAT, ABC/2 and 2ABC/3 methods overestimate the value of the volume; while the ABC/3 method underestimates it. According to Huttner et al.14, the ABC/3 method has not been validated clinically and, indeed, it can exhibit excellent behavior in cases in which the patient consumes anticoagulants or has undergone radio and/or chemotherapy. In addition, these authors assert, that in such cases the ABC/2 method presents a significant decrease in the accuracy of the estimation of the ICH volume.
On the other hand, table 4 presents the values corresponding to the relative percentage errors related to each of the methods considered.
Table 4. Values obtained for the percentage relative error related to each of the methods considered to obtain the volume of the 5 ICH present in the selected databases.
| Percentage Relative Error (%) | ||||
| SAT | ABC/2 | ABC/3 | 2ABC/3 | |
| DB1 | 4.98 | 21.89 | 18.74 | 62.52 |
| DB2 | 1.89 | 7.72 | 28.19 | 43.63 |
| DB3 | 1.43 | 5.40 | 29.73 | 40.53 |
| DB4 | 3.76 | 16.79 | 22.14 | 55.72 |
| DB5 | 0.72 | 5.67 | 29.56 | 40.89 |
| Average percentage relative Error (%) | 2.56 | 11.49 | 25.67 | 48.66 |
According to table 4, it can be stated that the SAT method generates the best average percentage relative error (PrE). In addition, the ABC methods exhibit the best performance in ABC/2, although in small volume hematomas it tends to produce higher errors. This may be a consequence of the fact that this method is based on the hypothesis that the ICH has an ellipsoidal shape and according to14, this is not always fulfilled (see, additionally, figure 4).
In this section, it is important to remember that the main surgical utility of the determination of the ICH volumes is that they define, in a high percentage, the behavior to be followed regarding the patient. In this sense, if only volume is considered, hematomas that exceed the threshold of 30 cm3 are susceptible to surgery. Following this criterion, and considering the results of the volume obtained by the MPM and the derivatives of the SAT method, which gave the lowest PrE, only the patient corresponding to the DB2 is a candidate to surgery.
CONCLUSIONS
In general, it can be said that the main characteristic of ABC methods is their simplicity and efficiency, although their performance, in many concrete situations, is not always the best option. In this sense, the fulfillment of the geometric hypothesis that an ICH has an ellipsoidal shape represents the main limitation of these methods, especially when it comes to patients who have ICH with no defined shape, relatively small and/or large volume. However, when the ICH complies with the aforementioned hypothesis, these methods have an acceptable performance and, in particular, the ABC/2 method has a good prestige since it has been clinically validated, while its variants do not yet have that condition. Additionally, in several investigations, it has been verified that these methods have as an additional disadvantage the property of being operator-dependent.
In the context of the present work, we have used an intelligent automatic technique (SAT) whose tuning allows the precise segmentation of the ICH, present in computed tomography images. This statement is based on the fact that the Dc obtained is comparable with that reported in the literature. The segmentations generated, automatically, by the SAT allow us to calculate the volume of each ICH in a precise and efficient manner. This volume is vital to address the hematoma that affects the health status of a patient and decide whether or not it is surgically treated.
Because the SAT method generated the lowest average percentage error, which did not exceed 3%, it can be affirmed that the performance of the SAT method exceeded the ABC methods considered. In part, this is due to the fact that the SAT does not assume any geometric consideration when it generates the volume of an intracerebral hematoma.
REFERENCES
1. Stippler M. Craniocerebral trauma. In: Daroff RB, Jankovic J, Mazziotta JC, Pomeroy SL, eds. Bradley's Neurology in Clinical Practice. 7th ed. Philadelphia, PA: Elsevier; 2016:chap 62.
2. Mezzadri J., Goland J., y Sokolvsky M. Introducción a la Neurocirugía. Capítulo: Patología vascular II. Ediciones Journal. Segunda edición. 2011.
3. Vera M. Segmentación de estructuras cardiacas en imágenes de tomografía computarizada multi-corte. Ph.D. dissertation, Universidad de los Andes, Mérida-Venezuela, 2014.
4. Maiera A, Wigstrm L, Hofmann H, Hornegger J, Zhu L, Strobel N, Fahrig R. Three-dimensional anisotropic adaptive filtering of projection data for noise reduction in cone beam CT. Medical Physics. 2011;38(11):5896–909.
5. Kroft L, De Roos A, Geleijns J. Artifacts in ECG–synchronized MDCT coronary angiography. American Journal of Roentgenology.2007;189(3):581–91.
6. Vera M., Huérfano Y., Contreras J., Vera M. I., Salazar W., Vargas S., Chacón J. y Rodríguez J. (2017). Detección de hemorragia intracraneal intraparenquimatosa, en imágenes de tomografía computarizada cerebral, usando una técnica computacional no lineal. Latinoamericana de Hipertensión. 12(5), 125-130.
7. LoPresti M., Bruce S., Camacho E., Kunchala S., Dubois B., Bruce E., Appelboom G., Connolly E. (2014). Hematoma volume as the major determinant of outcomes after intracerebral hemorrhage. J Neurol Sci. 2014 Oct 15;345(1-2):3-7.
8. Yildiz O., Arsava E., Akpinar E. y Topcuoglu M. Hematoma volume as a sole predictor of in-hospital mortality following spontaneous Intracerebral. Journal of Turkish Cerebrovascular Diseases 2011 17:2; 63-66.
9. Hu T., Yan L., Yan Peng., Wang X., Yue G. Assessment of the ABC/2 Method of Epidural Hematoma Volume Measurement as Compared to Computer-Assisted Planimetric Analysis. Biological Research for Nursing. 2016, 18(1) 5-11.
10. Rodriguez D., Boyko M., Subramaniam S., Klourfeld E., Jo P., Diederichs B., Kosior J., Dowlatshahi D., Aviv R., Molina C., Hill M. y Demchuk A. Magnitude of Hematoma Volume Measurement Error in Intracerebral Hemorrhage. Stroke. 2016;47:1124-1126.
11. Freeman, W., Barrett, K., Bestic, J.,Meschia, J., Broderick, D., Brott, T. Computer-assisted volumetric analysis compared with ABC/2 method for assessing warfarinrelated intracranial hemorrhage volumes. 2008, Neurocritical Care, 9, 307–312.
12. Kamnitsas K., Lediga C., Newcombeb V., Simpsonb J., Kaneb A., Menonb D., Rueckerta D., Glockera B. Efficient Multi-Scale 3D CNN with fully connected CRF for Accurate Brain Lesion Segmentation. Medical Image Analysis, Vol 23, pp.1603-1659, 2017.
13. Prakash K., Zhou S., Morgan T., Hanley D., Nowinski W. Segmentation and quantification of intra-ventricular/cerebral hemorrhage in CT scans by modified distance regularized level set evolution technique. Int J Comput Assist Radiol Surg. 2012; 7(5): 785-798.
14. Huttner H., Steiner T., Hartmann M., Köhrmann M., Juettler E., Mueller S, Wikner J., Meyding U., Schramm P., Schwab S. y Schellinger P. (2006). Comparison of ABC/2 Estimation Technique to Computer-Assisted Planimetric Analysis in Warfarin-Related Intracerebral Parenchymal Hemorrhage. Stroke. 2006;37:404-408.
Author notes
m.avera@unisimonbolivar.edu.co