A NOVEL DATA MINING ALGORITHM AND KNOWLEDGE-BASED DIAGNOSTIC RULES FOR MEDICAL THERMOGRAPH

Abstract

Diagnosis with medical infrared thermograph has long been recognized as lack of objective assessment and scientific analysis in clinical applications. In this study, a novel data mining algorithm with a parametric factor analysis protocol was utilized for generating knowledge-based diagnostic rules from infrared thermograph. These governing rules were developed through four-stage imaging analysis processes, i.e. geometric lofting standardization, abnormal region statistics, parametric factors analysis, and anatomical organs matching. In the first stage, Beier–Neely field morphing and linear affine transformation algorithms were used in geometric standardization for whole body contour and partial region, respectively. In addition, these two transformation methods were applied to solve the dilemma of thermal image deformation. In the second stage, pixels of abnormal body temperature were classified by a confident level. Cohesive and de-noises parameters were gathered to unify the pepper-like pixels into abnormal regions. In the third stage, 25 parametric factors were extracted from each abnormal region, and decision tree induction method was used to generate the knowledge-based diagnostic rules. In the final stage, the minimum box method was utilized for anatomical organ matching to identify the corresponding organ with abnormal temperature distribution. To verify the validity of the proposed knowledge-based diagnostic rules, a total of 71 and 131 female patients with and without breast cancer, respectively, were both analyzed in this study. Experimental results indicated that a total of 1750 abnormal regions (703 positive and 1047 negative) were detected. Based on decision tree induction method, 822 branches were broken down in the decision space. The more positive abnormal regions that were found in the terminal nodes of a specific branch, the higher the possibilities were for having cancer. Fourteen branches of the decision tree with terminal nodes were found to have more than four positive abnormal regions. Sixty one positive abnormal regions (61/703 = 8.6%) from 44 cancer patients (42/71 = 59.2%) can be found in the above-mentioned 14 branches. In conclusion, the test results reveal that the proposed data mining algorithm and knowledge-based diagnostic rules may be effective in clinical diagnosis of thermograph regardless of the type of disease.