A refined SVM applied in medical image annotation

• Authors: Bo Qiu
• Venue: CLEF'06 Proceedings of the 7th international conference on Cross-Language Evaluation Forum: evaluation of multilingual and multi-modal information retrieval
• Year: 2006

For the medical image annotation task of ImageCLEF2006, we developed a refined SVM method. This method includes two stages, in which accordingly the coarse and fine classifications are performed. At the coarse stage, the common Support Vector Machines (SVM) method is used with the corresponding image feature, down-scaled low resolution pixel map (32×32). At the refined stage, the results from the first stage are refined following the steps: 1) select those images near to the borders of SVM classifiers, which are to be re-classified and selected by a predefined threshold of SVM similarity value; 2) form a new training dataset, to eliminate the influence of the great volume unbalance among classes; 3) use three classification methods and image features: 20×50 low resolution pixel maps feed into SVMs; SIFT features feed into Euclidean distance classifiers; 16×16 low resolution pixel maps feed into PCA classifiers. 4) combine the results from 3). At last the results from the two stages are combined to form the final classification result. Our experimental results showed that with the two-stage method a significant improvement of the classification rate has been achieved.