Classification and Size & Shape Recovery from Endscope Image for Supporting Medical Diagnosis

This paper proposes a method to detect and the classify the polyp from endoscope image, and futher proposes a method to recover the size and shape of polyp for the purpose of supporting medical diagnosis. Here, polyp detection is proposed using CNN (Convolutional Neural Network) which is learned for the general object recognition and additional learning is introduced using the Fine-Tuning which is fitted for the polyp detection. This Fine-Tuning enables the suitable detection in the endoscope environment which does not sufficient number of endoscope images. After the polyp detection, malignant or benign of polyp is judged to the proposed approach using three different kinds of endoscope images consisting of white light image, dyed image and narrow band image (NBI). Seven kinds of Support vector machine classifiers are learned for the CNN feature vectors obtained from each image input and the approach shows that voting processing of each output of SVM provides the higher classification ability. Shape recovery of polyp first introduces the analysis of blood vessel and tries to find the corresponding points between two frames. The paper shows the blood vessel detection is used as the key to the shape and size recovery of polyp. The approach is demonstrated via experiments using actual endoscope images