Thyroid Tumor Classification Based On Mutli-mode Ultrasound Image

Thyroid tumor is a common and frequently-occurring disease and it is in the firstplace of head and neck cancer. It is also the fastest increase in the incidence of solidcancer. The developing medical imaging technology is an important measure ofdiagnosing whether the thyroid tumors is benign or malignant. Two-dimensionalgray-scale ultrasound and Doppler ultrasound are the basement to diagnose whether thethyroid tumors is benign or malignant. We can get a better diagnosis results bycombining new imaging diagnostic techniques like three-dimensional ultrasound, CEUS,elastosonography and so on. Most of the current computer-aided diagnosis of thyroiddeals with single-mode image information, that’s lack of fusion of multi-mode medicalimage information. This paper applies fusion of multiple classifiers to the field ofmedical images and combines different modes of image information to help doctors geta more accurate diagnostic results.Fusion of multiple classifiers became a hot research field of pattern recognitionsince the1990s. Fusion theory is based on differences in classification. Classifier camefrom different training data, different training algorithms can provide the classificationcomplementary information. Multiple classifier fusion system established by this hasmore comprehensive classified information and stronger generalization ability. Thispaper will propose a combination coefficients determined way for the combinationcoefficients problem in multiple classifier systems and the combination coefficientsdetermined way will apply to any fusion theory frame. This paper also proposes acomposite weighting for different categories based on the non-Bayesian fusionframework.1. The improvement of combination coefficients in multiple classifier systemsThis part mainly aims combination coefficients in multiple classifier systems andimproves after determination, so that it can be adapted to the different fusion frame. Asthe most common and effective classifier ensemble learning methods, classifiercombinations can be the final step of different fusion algorithms, it can be the finalfusion step of base classifier output. Previous combination coefficient’s determinationbased on Bayesian theory is limited Bayesian integration framework. So, in this paper,three kinds of metrics and the corresponding weighting algorithm is presented, they arebased on entropy, scattered values, mean variance ratio.2. A novel fusion strategy with multiple weight-based classifiers in a non-Bayesianprobabilistic frameworkThis part will improve a solution of optimal weights, this solution based onnon-Bayesian integration framework. Before improvement, this paper analyzes the difference in classifiers, and proposes the concept of complex weights based on thedifference. Multiple classifier fusion with different properties is an important factor forimproving the classification rates in general purpose classification approaches. Existingmethod usually evaluate the weight of individual classifiers by considering the globalprobability distribution of output with different categories. However, when theperformance of a classifier to the different classes is not uniform, the efficiency of thistype ensemble results may be affected negatively. In this paper, a novel fusion strategywith multiple weight-based classifiers is proposed. In proposed method, not only globalprobability distribution of classifier output with different features, but also probabilitydistributions with different classes are evaluated. Results demonstrate that the proposedscheme is effective and useful for improving the classification accuracy.The above algorithm is applied to multi-modal classification of tumors of thethyroid ultrasound images and the experiments certify that using multiple classifierfusion techniques can increase the accuracy of the classification of tumors.In addition,the proposed fusion strategy with multiple weight-based classifiers in a non-Bayesianprobabilistic framework is effective and useful for improving the classificationaccuracy.