Detecção e Classificação de Acidente Vascular Cerebral em Imagens de Tomografia Computadorizada por Extração de Características com Base na Análise Adaptativa da Vizinhança e Estimação de Parzen

Abstract

Stroke is a severe global health problem with high mortality rates. This disease is the second leading cause of death globally, with approximately 5.7 million per year, characterizing about 10% of all deaths worldwide. Also, stroke is still a significant cause of morbidity, hospitalizations, and acquired disabilities. However, the chances of recovery are high when the diagnosis is early, and the appropriate treatment is administered. Among the imaging tests that assist in diagnosing and treating stroke, Computed Tomography (CT) is the most appropriate technique since it allows to verify the extent and severity of the detected cases. Since it is an emergency with difficult identification and early measurement, several Computer Vision studies have developed new techniques and systems capable of automatically detecting and classifying a stroke. These computer-aided diagnoses use the images obtained by CT scans to extract information invisible to the human eye. In this sense, this work focuses on characterizing the region of interest, proposing two new methods for extracting features based on the analysis of the patterns of radiological densities of brain tissue. These new descriptors are improvements to the Analysis of Brain Tissue Densities (ABTD) method and aim to increase precision and decrease subjectivity in the radiological density ranges. The first method, called Adaptive Analysis of Brain Tissue Densities, has as the main contribution an adaptive analysis of the pixel’s neighborhood to determine which Hounsfield unit (HU) range the pixel belongs to. Thus, Adaptive ABTD considers the local characteristics of each pixel and not only the HU value of a pixel as in the original implementation of ABTD. This adaptive analysis in each pixel of the image makes the feature maps created by the proposed method smoother, showing regions of the brain tissue more clearly. The second method, Parzen Analysis of Brain Tissue Densities (PABTD), is a descriptor that performs a probabilistic analysis on the radiological attenuation levels of each pixel of the image. The PABTD estimates a probability density function for each pixel using the Parzen-Rosenblatt window technique. The insertion of this probability estimation step gives the method greater precision and reliability in determining the actual pixels of brain tissue. These methods are specific approaches for extracting relevant features in CT images that assist in identifying and classifying stroke types. For the performance evaluation of the proposed descriptors, their results were compared with extractors already consolidated in the literature, such as grey-level co-occurrence matrix (GLCM), local binary standards (LBP), Statistical Moments, and the original ABTD method. The extractors were applied to a database of CT images of the skull containing three classes of stroke: normal, ischemic, or hemorrhagic. Then,

the resulting feature vectors were used as input for the Bayesian, K-NN, MLP, SVM, and OPF classifiers, which generated scores to represent the probability of the samples belonging to each of the three classes. The proposed methods showed promising results, surpassing the other extractors in all evaluated statistical metrics. Both ways achieved the best results when combined with the SVM classifier with kernel RBF, reaching an average accuracy rate above 98%, F1 score above 97%, and negative and positive predictive value above 96%. Based on the results obtained, we conclude that the Adaptive methods ABTD and PABTD are practical algorithms for extracting features of CT images of the brain and superior to the other state-of-the-art descriptors. Thus, we conclude that these methods can integrate systems that assist medical diagnosis for detecting and classifying strokes.