PCB : A Predictive system for Classifying multimodal Brain tumor images in an image-guided medical diagnosis model
Lau Phooi Yee1, Ozawa Shinji2
1laupy@ozawa.ics.keio.ac.jp, Keio University; 2ozawa@ozawa.ics.keio.ac.jp, Keio University
The rapid advancement in computing technologies has been fueling researches covering a wide range of disciplines, spanning from life sciences, through physical sciences, information technology and engineering, to clinical medicine. We are moving towards a proactive diagnosis system, whereby computer can anticipate and provide treatment program automatically, and the study and experimenting towards this faculty in terms of resource management and application of this sector is necessary. The evolution of healthcare services from patient homogeneity to individual risk assessment and treatment selection, redefine the role of computerized decision support especially in terms of quality assurance and clinical training because patients is now looking at healthcare services from a new perspective. In our earlier work, we propose a framework for a computer-assisted medical image diagnosis or analysis of medical images incorporating computer-like terminal with high performance computing is to explore the next frontier in healthcare delivery for medical practitioners. This framework is focused on the major fundamental image processing components while trying to also establish an economic model for image-guised medical diagnosis system.
Over the past few years, we have seen an increasing amount of research in areas of image processing, indirectly contributing to the intelligent and automatic image classification system. Looking into the past literature, predictive image classification has not yet been broadly applied as a computer-assisted diagnosis tools in a clinical environment. We believe diagnosis research-oriented system having the capability to classify healthy and cancer patients shall become the next pressing issue.
Our proposed approach in this research includes the pre-processing block, employed to identify and preview input images or the subject according to age, gender and imaging modality. Later, it exploits the content-based, shape-based and texture-based technique, which is not constraint by imaging modality, to introduce some form of predictive diagnosis capabilities in the classification system block in our proposed framework. We studied the connection between disease and its tumor image properties in three different image perspectives: binary image, intensity image and selected-pixel intensity image. Binary and intensity image slice profiling are based on texture and shape-based classification technique while selected-pixel intensity image slice profiling is based on content-based classification technique. In the shape-based approach, input image are binarize using a certain threshold, then we extract the skeleton of that binarized image for profiling. In the texture-based approach, we analyze the pixel intensity value of input images using histogram and profiling methods. In the content-based approach, pixel values of problematic area are first determined because cancerous brain tissues often associate with higher pixel intensity value. In this study, we also looked at whether gender and age has played any role during input images slice profiling of both healthy and cancer.
The classification is implemented using MATLAB on an IBM ThinkPad Pentium III 733MHz computer with 384MB RAM. MATLAB is the core analysis software used in this paper. Another supporting software is the Visible Human 2.0 system. The experiments were conducted on two data sets: the Visible Human (labeled female) and the Whole Brain Atlas v. 1.0.
Experiment results reveals that a common trending was found during slice profiling of both binary and intensity images, without regards to age and gender. Other findings include the detection of a surge in the pixel intensity when the patient is having terminal diseases. Based on our results, selected-pixel intensity images image classification is more apparent than binary image or intensity image for cancerous tumor detection and diagnosis even at an early stage. . The results also show that tumor occurrences are best associate with: 1. Decrease in pixel count in binary images 2. Increase in image intensity in intensity images 3. High numbers of high intensity pixel in selected-pixel intensity images. In comparison, it is found that selected-pixel intensity image is more evident than binary or intensity image. Our future efforts will focus on introducing a user-friendly and practical system usable in clinics for medical specialists.