Shape Decomposition Approach for Ultrasound Color Doppler Image Segmentation
ICPR '06 Proceedings of the 18th International Conference on Pattern Recognition - Volume 04
Improving classification accuracy using automatically extracted training data
Proceedings of the 15th ACM SIGKDD international conference on Knowledge discovery and data mining
Bayesian Classifiers Programmed in SQL
IEEE Transactions on Knowledge and Data Engineering
ICCNT '10 Proceedings of the 2010 Second International Conference on Computer and Network Technology
An analysis of Bayesian classifiers
AAAI'92 Proceedings of the tenth national conference on Artificial intelligence
IEEE Transactions on Information Technology in Biomedicine
Hi-index | 0.00 |
Generally patient data in healthcare environments exist in relational databases. Classification of echocardiographic images is an important data mining task that helps hospitals without transferring the data in any form. This paper proposes a novel method to accomplish this task using naïve Bayesian model via SQL. The proposed method has two phases. The first phase builds a knowledge base using many normal and abnormal subjects and the second phase uses this knowledge to categorize an unseen subject into appropriate class. The image features such as cardiac chamber dimensions (specifically Left Ventricle - LV), Ejection Fraction, Mitral valve (MV) orifice area, etc., are computed by first segmenting the image by employing advanced image processing techniques. For instance, to segment echo images we employ an efficient SQL based Fast K-Means algorithm combined with a greedy active contour algorithm for accurate boundary detection. Additional features such as textural, statistical, and histogram are computed and added to the classifier model by analyzing color Doppler echo images to strengthen the classifier accuracy. Our SQL based naïve Bayesian classifier model is built with 7 schema, simple yet efficient SQL queries and thus providing an accurate classification of patients as normal or abnormal. The model is trained with 112 patient data and we believe that the clinical decision is simplified and can happen on-the-fly. Experimental results presented in this paper show an increased accuracy of 87.48% against the other state-of-the art segmentation and classification methods reported.