Leveraging Mid-Level Semantic Boundary Cues for Automated Lymph Node Detection
Histograms of oriented gradients (HOG) are widely employed image descriptors in modern computer-aided diagnosis systems. Built upon a set of local, robust statistics of low-level image gradients, HOG features are usually computed on raw intensity images.
- PDF / 1,582,905 Bytes
- 9 Pages / 439.363 x 666.131 pts Page_size
- 80 Downloads / 162 Views
Abstract. Histograms of oriented gradients (HOG) are widely employed image descriptors in modern computer-aided diagnosis systems. Built upon a set of local, robust statistics of low-level image gradients, HOG features are usually computed on raw intensity images. In this paper, we explore a learned image transformation scheme for producing higher-level inputs to HOG. Leveraging semantic object boundary cues, our methods compute data-driven image feature maps via a supervised boundary detector. Compared with the raw image map, boundary cues offer mid-level, more object-specific visual responses that can be suited for subsequent HOG encoding. We validate integrations of several image transformation maps with an application of computer-aided detection of lymph nodes on thoracoabdominal CT images. Our experiments demonstrate that semantic boundary cues based HOG descriptors complement and enrich the raw intensity alone. We observe an overall system with substantially improved results (∼ 78% versus 60% recall at 3 FP/volume for two target regions). The proposed system also moderately outperforms the state-of-the-art deep convolutional neural network (CNN) system in the mediastinum region, without relying on data augmentation and requiring significantly fewer training samples.
1
Introduction
Quantitative assessment of lymph nodes (LNs) is routine in the daily radiological workflow. When measuring greater than 10 mm in short-axis diameter on an axial computed tomography (CT) slice, LNs are generally considered clinically relevant or actionable [13], indicative of diseases such as lung cancer, lymphoma, or inflammation. Manual detection of enlarged LNs, critical to determining disease progression and treatment response, is a time-consuming and error-prone process. Thus, there has been active research in recent years to develop accurate computer-aided lymph node detection (CADe) systems. A challenging object class for recognition, LNs exhibit substantial variation in appearance/location/pose as well as low contrast with surrounding anatomy on
The rights of this work are transferred to the extent transferable according to title 17 § 105 U.S.C.
c Springer International Publishing Switzerland 2015 (outside the US) N. Navab et al. (Eds.): MICCAI 2015, Part II, LNCS 9350, pp. 53–61, 2015. DOI: 10.1007/978-3-319-24571-3_7
54
A. Seff et al.
CT scans. Recent work on LN CADe has varied according to the feature types and learning algorithms used for training. [1, 8] utilize direct 3D information from CT scans, performing boosting-based feature selection over a pool of 50–60 thousand 3D Haar wavelet features. Due to the curse of dimensionality (analyzed in [14]), such approaches can result in systems with limited sensitivity (e.g. 60.9% at 6.1 FP/scan for mediastinal LNs in [8]). Circumventing 3D feature computation during LN classification, [14] implements a shallow hierarchy of linear models operating on 2D slices or views of LN candidate volumes of interest (VOIs) with histograms of oriented gradients (HOG) [3] features. Also
Data Loading...
