A proposed framework: Group-based image analysis to enhance accuracy of image classification for tumor diagnostic

Abstract

Accurate diagnostic of tumor is crucial to reduce unnecessary number of biopsies and surgeries. Thereby, an enhancement of classification technique is required to accommodate multiple images (from multi-view) automated diagnostic. Moreover, it will be beneficial for radiologist during diagnostic procedures. Studies underlying Multi-Instance (MI) problem were reviewed, and it is found that there exist few studies discuses on collective approach by combining multi-instances for bag-level decision. However, there is none focuses on purely bag level decision which has been main focus of this study. In conventional approach, an issue occurred when an instance in a bag give negative label even it may contain a very small portion to be a positive label. This decision will be improved if represent corresponding to the complete image from collective information from all instances. A preliminary experiment was conducted using conventional techniques. It proved that single level decision acquired ‘not good’ performance need to be improved. Thus, a new framework using group-based image analysis strategy is proposed. This framework is aimed for extend conventional classification algorithms to meet the image analysis needs and improvising the accuracy of tumor diagnostic.