Multilevel transfer learning Frameworks for classification and Annotation with limited medical Datasets

Abstract

In recent years, there is a significant development in the healthcare newlineindustry due to the digital technologies that could help to transform newlineunsustainable healthcare systems into sustainable ones. Imaging technology that newlineplays a leading role in healthcare industry today shapes the evolution of the field newlineto attain its place of prominence. Medical image classification is one of the most newlineimportant research areas in the image recognition field. Computer Aided newlineDetection (CAD) systems have been extensively used as a fundamental tool in newlinethe medical image classification field, due to their improved performance in such newlinedetection and diagnosis tasks. Such systems are able to analyze medical images newlineand identify suspicious areas, which are relevant to the radiologist findings. newlineWhen solving the problems of medical imaging, the efficacy of the CAD newlinetechniques is confined by limited data availability. Even the machine learning newlineand deep learning based techniques, which are popular today, are not able to newlinedeliver good performance when the size of datasets is limited. newlineFor example, Digital Breast Tomosynthesis (DBT), a new imaging newlinemodality, which is widely used for breast screening nowadays, is the most newlineeffective method for detection of early breast cancer. However, the collection newlineof huge amounts of the DBT images are complex, as they are not publicly newlineavailable and therefore, developing classification systems based on these newlineimages becomes challenging. Similarly, when we consider rare diseases, newlineclassification of such rare diseases accurately is challenging and considered as newlinea bottleneck in medical image diagnosis newline