Development of geometric, shape and semantic feature based similar case chest radiograph retrieval system

Abstract

Due to growth of medical image database. the interest of retrieving similar case image is getting high. Yet, the process has been slow because there are not enough open source and it is difficult justify the meaning of "similar". Thus, I made a system that gives variety of choices for doctors to decide what factors they are looking for. There are 3 features: geometric, shape, and semantic. Geometric features discover location, area, and ratio of a given ROI. The noise of images are reduced through normalization using a bounding box of a lung. Shape feature have good ability to find unique shape of the ROI mask. To make the retrieval faster, I made a distance predicting queue, which skips comparison time of adjacent element of the R-table. Semantic feature showed excellency at retrieving same class images but it was not explainable. Therefore, it went through resorting using a geometric or shape feature which is more explainable.