Lesion detection of gastroscopic images based on cost-sensitive boosting

Gastroscopy is widely used for clinical examination of gastric cancer which is one of the most serious diseases. Computer-aided detection can help physicians to identify suspicious regions to reduce false negative diagnosis which costs too much more than false positive diagnosis. Three cost-sensitive boosting algorithms are compared in this paper in the task of lesion detection of gastroscopic images. The optimal cost structure is selected for each boosting algorithm. Threshold obtained adaptively from training set is adopted to get the final result of a novel sample instead of the sign function. Classification performance becomes better after adaptive threshold is used. Experimental results show that Cost-sensitive AdaBoost performs the best for lesion detection of gastroscopic images achieving a sensitivity of 77.34% with the threshold obtained on training set at a target detection rate of 80%. Lesion detection based on cost-sensitive AdaBoost can outline the lesion area more completely and accurately than AdaBoost method.