Guided Random Forests for Identification of Key Fetal Anatomy and Image Categorization in Ultrasound Scans
In this paper, we propose a novel machine learning based method to categorize unlabeled fetal ultrasound images. The proposed method guides the learning of a Random Forests classifier to extract features from regions inside the images where meaningful str
- PDF / 443,759 Bytes
- 8 Pages / 439.363 x 666.131 pts Page_size
- 49 Downloads / 190 Views
Institute of Biomedical Engineering, Engineering Science, University of Oxford 2 Nuffield Department of Obstetrics & Gynaecology University of Oxford
Abstract. In this paper, we propose a novel machine learning based method to categorize unlabeled fetal ultrasound images. The proposed method guides the learning of a Random Forests classifier to extract features from regions inside the images where meaningful structures exist. The new method utilizes a translation and orientation invariant feature which captures the appearance of a region at multiple spatial resolutions. Evaluated on a large real world clinical dataset (~30K images from a hospital database), our method showed very promising categorization accuracy (accuracytop1 is 75% while accuracytop2 is 91%). Keywords: Random Forests, Ultrasound, Image Categorization, Classification, Normalized Cross Correlation.
1
Introduction
Image categorization is a well-known open-research problem in computer vision for which many solutions have been proposed [1-3]. In medical image applications, image categorization is relatively under-investigated but nevertheless important. The volume of digital images acquired in the healthcare sector for screening, diagnosis or therapy is very large and increasing steadily. Ultrasound based fetal anomaly screening is usually performed at 18 to 22 weeks of gestation. Several images of fetal structures are acquired following a standardized protocol. This screening scan aims to determine whether the fetus is developing normally by assessing several ultrasound images of different fetal structures. The number of different structures that are imaged and acquired in a complete scan varies, for example, the UK NHS Fetal Anomaly Screening Programme (FASP) recommends 21 views to be assessed and at least 9 images to be stored. The number of individual women undergoing a scan is often of the order of several thousands per department per annum. Most clinical departments save these scans to an archive system without any labeling of these images. This means it is not possible to recall images of body parts conveniently for later review or measurement nor to compare scans of the same fetus over time, or conduct automatic measurement post-acquisition. Manual categorization is of course theoretically possible. However, it is expensive as it requires a good level of expertise alongside being tedious and time consuming. In this paper, we propose a method to automatically categorize fetal ultrasound images from anomaly scans. The new method is built on a machine learning classifier (Random © Springer International Publishing Switzerland 2015 N. Navab et al. (Eds.): MICCAI 2015, Part III, LNCS 9351, pp. 687–694, 2015. DOI: 10.1007/978-3-319-24574-4_82
688
M. Yaqub et al.
Forests RF) in which we propose novel ideas to guide the classifier to focus on regions of interest (ROI). Although there are a number of methods which have been proposed to address different medical image categorization problems [4-6], very little research has been done in fetal ult
Data Loading...
