Ideas for Dealing with Reduced Datasets in Development of CADe Systems for Medical Uses

Abstract

This work presents a real time user friendly system to aid specialized professionals to analyze bone scans exams. In order to achieve this, some original ideas are applied. The first one is related to the use of each pixel of an exam as object of interest for classification. Another original idea is the use of operations that are normally applied in pre-processing as features for machine learning. With both, even using small dataset was possible to obtain enough amounts of entries to be used for training and testing. Initially, the feature vectors are composed by 64 features and one target attribute representing the classification result. The used bone scans set was composed of 42 images from 21 patients. At the end of the learning tasks a dataset of 2,512,386 records is computed. In order to reduce the cardinality of the vector of features, the Principal Component Analysis was employed leading to a new feature set with 25 components per object to be classified as with or without metastasis, the area under the Receiver Operator Characteristic curve achieved with this final set of features was 98%.