Robust classification techniques for medical imaging

点击翻译按钮获取中文摘要

Machine learning techniques have had an enormous impact on the field of medical image analysis and have been successfully applied to a wide variety of classification, segmentation and registration problems. There are, however, significant challenges in taking a classification algorithm trained on one specific research dataset and translating it to routine clinical practice where images acquired at different centres in different patient populations may have very different characteristics. The problem is made even more difficult by the high cost of collecting good quality training images complete with expert annotations or ground truth diagnoses. The medical imaging community have typically addressed these issues by using knowledge about the physical characteristics of the images to normalize the input data, for example by using intensity rescaling or spatial interpolation before extracting features for classification. More recently domain adaptation or transfer learning techniques have been developed in computer science which aim to adjust classifiers learnt on one data set to cope with differences in the target dataset. These methods generally involve leveraging the information learnt from the labelled training data to classify the unseen target data where there are either few or no annotations available; examples include re-weighting of training cases based on their similarity to the test cases, deriving transformation functions that map features from one domain into another and semi-supervised clustering techniques. There has been very little research on semi-supervised learning and on domain adaptation in the medical imaging field to date. There are some significant obstacles to the widespread adoption of machine learning techniques in clinical practice; labelled data sets are typically orders of magnitude smaller than those used in machine vision but at the same time there is a much greater need for highly accurate results. Classifiers also have to deal with a wide variation in image quality and acquisition parameters. We propose to explore existing methods of semi-supervised learning and domain adaptation and also develop novel approaches to cope with the specific problems encountered in medical imaging; we will concentrate on two very different clinical applications namely breast MRI and digital pathology however the approaches we develop will be generalizable to other applications and will help to speed the translation of machine learning algorithms out of the research lab and into software packages used routinely to solve clinical problems.