Shapes, textures and colours contain information in image reading. Machine reading gets its’ functionality from transformation of textures to features, segmentation and classification. Result is interpretation of image content. Many medical diagnoses rely on information content from images of surfaces and volumes of human body. Quality of image supported diagnosis relies on of imaging technology, approach, machine; image interpretation relies on experts’ expertise and on machine that is supporting him. Advanced medical environments encourage machine supported decision making process. Lung cancer is frequent among cancers. It results in high mortality, especially when it appears as secondary cancer. Much effort is spent in techniques and procedures for early diagnose. Auto Fluorescent Bronchoscopy (AFB) is used for early detection in bronchial tubes. Progress in computer science, embedded systems, programming for real-time applications gave means to integrate machine support in the diagnostic process. Physician – expert makes decisions and stands behind her diagnose, and supporting tools add to quality of decision making process. Tools make experts’ work less frustrating. Medical doctors accept machine support of their work well. Image processing is used in manufacturing processes and in logistics for years. Objects are well defined in these branches of human activity. Image processing in medical diagnostics is more complex. Examined structures can vary in size, shape, colour and texture. Tolerated margin for error in feature detection is about nil. We increase AFB sensitivity and specifity with image processing, machine learning of differences among benign lesions, precanserosis and cancer of mucus membrane in lower bronchial tubes. Images were taken at the Department for lung disease and allergy at Univerzitetni kliniˇcni center in Ljubljana. LIFE R(Xillix Technology, Vancouver, BC) apparatus was used. Images were classified into 3 groups i.e. normal, suspicious for precancerosis and cancer. Each image is supported with biopsy and histopathological diagnose. Development of system for autonomous detection of pathological changes with AFB is presented. Procedure for improvement in tissue classification is developed. We researched implementation means and use of feature detection and machine learning for autonomous interpretation of AFB images. The goal was to select most reasonable components and to produce optimal system for machine detection of bronchial cancer in its most early stage. Images, histopathology data, medical diagnoses, consulting with physicians, feature detection algorithms and machine learning procedures define start of
the project of machine diagnose from the AFB image. Suspicious areas in AFB image have to be precisely detected. To improve area definition we introduced filtering as image pre-processing. Canny edge detection method was applied. It is unsusceptible to small local areas and it is computationally efficient. Errors in area definition are noticed in real-time and corrective input is given. We detected suspicious precancerosis and malignant areas with data from Ground Truth (GT) norm and our new procedure. We compared 2 methods for feature extraction. 1st method extracts features from Grey Level Co-occurence Matrix (GLCM); 2nd method extract features from image histogram and semi Normal distribution of number of pixels versus their colour intensity. To distinguish between suspicious only and malignant areas we first carefully selected small number of information rich features. Then we applied naive Bayes, k-NN and Support Vector Method (SVM) classifiers. SVM produced best results. Two thirds of samples were used for machine learning, one third for testing. Quality of binary classification was assessed by different measures – precision, sensitivity, specifity and Area Under Curve (AUC). Both methods are aimed at high sensitivity accompanied by high specifity. The goal is to minimize False Positive (FP) and False Negative (FN) results, which leads to smaller amount of biopsies. Second method is computationally more efficient and it takes less computer resources (memory). Wilcoxon signed-rank test was used to evaluate ranking efficiency of both methods. Null hypothesis is that both methods produce about same results. p value was calculated for selected confidence level (_ = 0.05). Testing of statistical hypothesis ranked efficiency of both methods. Significant value is below confidence level which confirms efficiency of the new method. Results are ready for implementation in embedded system for real-time machine supported AFB diagnostics, which improves accuracy and specifity. Consequently less biopsies is performed at preserved quality of diagnostic work.
|
