Computer Assisted Cytological Medical Image Analysis

点击翻译按钮获取中文摘要

The proposed research leverages advances in Pattern Recognition, ImageProcessing, Computational Geometry and Machine Learning to develop novelsolutions to address problems in the area of Medical Imaging. Specifically, Iam furthering my research on Computer Aided Medical Diagnosis for cytologicalimagery of breast cancer fine needle biopsies, and of thin blood smears formalaria detection. For these types of cytological images, my long termobjective is to develop robust, accessible, rapid, and accurate MedicalDiagnosis tools to provide an automatic independent second opinion, to helpavoid errors/oversights, to Pathologists, Clinicians and Medical Specialists,or act as an expert system to aid the non-specialist medical practitioner.The process of computer-assisted medical diagnosis or malignancy grading canbe described in two main stages: feature extraction from medical imagery andclassification of the imagery with respect to specific medical issues. Oneobjective of the proposed work is the determination of improved feature setsfor specific types of cytological images. Many of the frequently usedfeatures used for medical image classification problems require accurateimage segmentation prior to determining these features. Therefore, I amstudying approaches to improve the accuracy and speed of the segmentation ofcells in cytological images by extending my previous work ranging from theuse of the Hough transform to using active contours models, to incorporateeffective use of texture, per-pixel classifiers, and shape modeling. We arealso extending this research to the accurate segmentation of new medicalimaging technologies such as very large full slide virtual images, 3D images,and images created with extended focal imaging (EFI).Another objective of the proposed research is to improve the accuracy of themalignancy classification in cytological images of breast cancer fine needlebiopsies. Specifically, we study the problems of malignancy diagnosis todetermine whether the slide is malignant or benign, and of malignancy gradingwhere we use the Bloom-Richardson malignancy grading. To accomplish this,with optimized feature sets, we will develop systems of classifiers thatimprove sensitivity rates, particularly with minimal rates of falsenegatives. Since the classifiers have to work in a clinical setting, we willadapt the classifiers to do automatic parameter tuning. The third objectiveof the proposed research is to improve the detection of malaria parasites incytological images of blood smear slides. Building on our previous completeblood count framework, we will develop a computer-assisted malaria detectionsystem to determine the presence of malaria and return accurate infectioncounts per RBC counts in thick blood smears; and to differentiate between thefour species of the malarial parasite in thin blood smears. We will developsystems of classifiers with improved sensitivity rates.My work in Health Informatics contributes to the larger trend in medicalimagery to develop computer assisted medical diagnosis tools to complementand facilitate the work of Clinicians and Pathologists. All developments aredesigned to use commonly available microscopy and computing resources. Forexample, in the rural areas of malaria stricken countries there arefrequently no specialists to do microscopic screening of blood slides formalaria. Through the automatic analysis of these slides, the computerassisted expert systems proposed in my research would help fill the knowledgegap between a specialist with the necessary expertise and the practitioneractually examining cytological slides, thus leading to better patient care.