BRAIN IMAGES ANALYSIS FOR SDH AND EDH TRAUMA CLASSIFICATION

Abstract

Imaging is a potential way to preserve accurate information of a sense at a particular time for repetitive observation and analysis in future. Clinical brain imaging captures the information about actual condition of brain and presents it in pictorial form for observation. Brain hemorrhage is a common disorder with high death toll. As per Indian Head Injury Foundation report, every year 100000 lives have come to an end due to brain hemorrhage. An effective medical support system is, thus, in high demand to gain clinical control over hemorrhage by reducing the death rate and the pain of survivors with disabilities. In developing countries like India, Computed Tomography (CT) imaging is the majorly used scanning modality because of its cost and speed. Though different mathematical models are reportedly established for automatic or semi-automatic computation for hemorrhage detection and classification using CT images, better works are still in demand for commercialization. A commercially available Computer Aided Diagnosis (CAD), if used in different hospitals, will create homogeneous reports which will be easy to exchange and compare. This research work is focused to support medical system by designing a CAD for hemorrhage detection and classification of Epidural Hemorrhage (EDH) and Subdural Hemorrhage (SDH). To make the CAD useful and acceptable, our objectives are as follows: (1) Developing a simple but highly accurate system. (2) Making the system affordable i.e. the implementation cost must be low. (3) Designing fully automatic system, so that it can be used for (a) Mass screening during emergency, (b) In absence of practitioners for initial diagnosis, (c) To get lower subjectivity and no tiredness. (4) Sensitivity of the system must be high enough to make the decisions of machine acceptable. (5) Designed CAD must offer reliable and dependable performance. Outline of the entire research work is presented as follows. The analysis of images involves segmentation, feature extraction and classification. Segmentation is done in two steps. In the first step the region of interest (ROI) which is the brain area encapsulated within skull is extracted from head scan images. In the second step, hemorrhage is segmented from diseased affected ROIs. As different hemorrhage has different shape, using the shape features hemorrhages are classified. Results of each step are discussed and compared with recently reported potential research outcomes. Image analysis depends on features which are extracted from descriptors of the target image. Commonly used image features are color, shape and texture features. For easy and quick information extraction, a single threshold based binary descriptor is proposed. From this descriptor, three major ii binary image features called Information Packing Factor (IPF), Compactness (C) and Porousness (P), and two auxiliary features Scatterness (S) and total pore area (w) are calculated to handle multi-slice brain image dataset. These features describe the availability and spared of foreground information of an image and are useful for the purpose of brain CT image indexing. These are used in different steps of the research to improve CAD performance. To narrow down hemorrhage search area, the brain region is segmented from head scan image. In the pre-processing stage dataset cleaning and arranging in anatomical order are done. Dataset cleaning is done by removing images having no significant brain information. Pore count and IPF feature have efficiently handled this requirement. Using stereo matching the cleaned dataset is arranged in expected anatomical order. Image with maximum brain area in the pre-processed dataset is selected as master image to create mask for brain segmentation. Two masks are created by automatic seed point finding and region growing method. One covers only the intracranial area of master image and another includes its skull area too. The larger mask is used as global mask for the dataset, but the smaller one propagates as adaptive mask. First, it segments the brains of the adjacent images of master image; then it is redefined by adjacent area search and used to segment the next adjacent images. This method efficiently segments the brain image dataset with 98.17% accuracy and 100% sensitivity. The extracted brain images are considered as input for hemorrhage segmentation process. To locate the threshold intensity for the hemorrhage, brain image histogram and expected histogram are considered. The expected histogram is the intensity distribution which is calculated from actual histogram history. Threshold intensity of an image is found by locating the crossover point, addressed as ‘upset point’ in this thesis, between actual histogram and expected histogram. The threshold intensity search is kept limited in the region beyond the maximum intensity of brain histogram. Normal dataset are removed to reduce load on system. Segmented images of a dataset are fused linearly to locate highest potential area which is then converted into a mask for hemorrhage segmentation. Proposed method has reported average accuracy as 93.19%, average sensitivity as 93.47% and dice coefficient as 92.05% which are much higher than other popular methods. The potential of this method is its speed, accuracy and sensitivity. The higher dice co-efficient of end result has proved that the upset point separates hemorrhage from brain matter without much loss. Segmented hemorrhages are classified in two steps using decision tree classifier. Initially the target classes are separated from other hemorrhages. Target class is then classified into EDH and SDH. For better performance of classifier secondary shape features are calculated from primary features and, to reduce input load on classifier, feature selection is done using their separability index. Target classes are separated from other hemorrhage types with 100% accuracy. EDH and SDH dataset are also classified with 100% accuracy. Each image of a dataset is classified. The class which contains more images is considered as the class of that dataset. The important observations of the described classification technique are listed below. iii (1) Target hemorrhage classes stay in the immediate vicinity of brain boundary. This single feature is strong enough to separate them from other classes (2) Secondary shape features have more potential than primary shape features for classification because of their higher dependency on object shape. (3) IPF and compactness have demonstrated noticeable strength in classification of EDH and SDH. The proposed CAD has significantly high accuracy in segmentation and classification. No false negative result is reported in any step of the entire process. This feature has made the system highly sensitive and dependable for commercial use. The proposed CAD can handle any size of dataset for hemorrhage inspection and classification, even when the scanned images are not in the expected anatomical order. The research target has achieved successfully and the journey has offered some useful bi-products which can also be used for other brain disease detection. The significant bi-products are the master image selection technique, arranging CT images in anatomical order and upset point finding.