Robust Segmentation of Various Anatomies in 3D Ultrasound Using Hough Forests and Learned Data Representations
3D ultrasound segmentation is a challenging task due to image artefacts, low signal-to-noise ratio and lack of contrast at anatomical boundaries. Current solutions usually rely on complex, anatomy-specific regularization methods to improve segmentation ac
- PDF / 2,582,936 Bytes
- 8 Pages / 439.363 x 666.131 pts Page_size
- 81 Downloads / 213 Views
Computer Aided Medical Procedures, Technische Universit¨ at M¨ unchen, Germany Computer Aided Medical Procedures, Johns Hopkins University, Baltimore, USA Department of Neurology, Klinikum Grosshadern, Ludwig-Maximilians-Universit¨ at M¨ unchen, Germany
Abstract. 3D ultrasound segmentation is a challenging task due to image artefacts, low signal-to-noise ratio and lack of contrast at anatomical boundaries. Current solutions usually rely on complex, anatomy-specific regularization methods to improve segmentation accuracy. In this work, we propose a highly adaptive learning-based method for fully automatic segmentation of ultrasound volumes. During training, anatomy-specific features are obtained through a sparse auto-encoder. The extracted features are employed in a Hough Forest based framework to retrieve the position of the target anatomy and its segmentation contour. The resulting method is fully automatic, i.e. it does not require any human interaction, and can robustly and automatically adapt to different anatomies yet enforcing appearance and shape constraints. We demonstrate the performance of the method for three different applications: segmentation of midbrain, left ventricle of the heart and prostate.
1
Introduction and Related Work
Manual segmentation of ultrasound volumes is tedious, time consuming and subjective. In the attempt to produce results that are invariant to the presence of noise, drop-out regions and poorly distinguishable boundaries, current computeraided approaches either use complex cost functions, often regularized by statistical prior models, or require extensive user interaction. Many optimization-based methods utilize cost functions based on local gradients, texture, region intensities or speckle statistics [10]. Methods employing shape and appearance models often require a difficult and time-consuming training stage where the annotated data must be carefully aligned to establish correspondence across shapes in order to ensure the correctness of the extracted statistics. Learning approaches have been successfully proposed to solve localization and segmentation tasks both in computer vision [7,12] and medical image analysis [6]. Handcrafted features which exhibit robustness towards the presence of noise and artefacts have been often c Springer International Publishing Switzerland 2015 N. Navab et al. (Eds.): MICCAI 2015, Part II, LNCS 9350, pp. 111–118, 2015. DOI: 10.1007/978-3-319-24571-3_14
112
F. Milletari et al.
employed to deliver automatic segmentations [8]. Recent work [3] in the machine learning community focused on approaches leveraging single [5] or multi-layer [9] auto-encoders to discover features from large amount of data. In particular, sparse auto-encoders with a single-layer have been proven to learn more discriminative features compared to multi-layer ones [5] when a sufficiently large number of hidden units is chosen. In the medical community, recent approaches [4] have employed deep neural networks to solve segmentation tasks, despite their computational burden due
Data Loading...
