Content area
Full text
Abstract-Nailfold capillaroscopy (NC) is a non-invasive imaging technique employed to assess the condition of blood capillaries in the nailfold and is particularly useful for diagnosis of scleroderma spectrum disorders and Raynaud's phenomenon. Diagnosis is based on the identification of particular scleroderma patterns in the images which are typically grouped into early, active and late patterns. In this paper, we present a computer vision approach to recognising scleroderma patterns in NC images. Following a preprocessing step to enhance image quality, we extract texture information in a holistic way rather than trying to extract and measure individual capillaries. As texture features we employ multi-dimensional LPB variance descriptors which capture multi-resolution texture and local contrast information. Our experimental results confirm our approach to work well and to outperform an earlier approach.
Keywords: Medical imaging, nailfold capillaroscopy, texture, LBP, MD-LBPV.
(ProQuest: ... denotes formulae omitted.)
1. Introduction Nailfold capillaroscopy (NC) is a non-invasive imaging technique employed to assess the condition of blood capillaries in the nailfold. It is particularly useful for early detection of scleroderma spectrum disorders [1] and evaluation of Raynaud's phenomenon [2]. Diagnosis using NC images involves the classification into Early, Active and Late groups, also known as NC patterns or scleroderma (SD) patterns [3], [4] (see Fig. 1 for example images) based on the identification of enlarged or giant capillaries, haemorrhages, loss of capillaries, disorganisation of the vascular array, and ramified/bushy capillaries in the images [5].
While diagnosis based on NC is typically performed by manual inspection, computerised nailfold capillaroscopy can help to reduce the inherent ambiguity in human judgement while greatly reducing the time for diagnosis [6]. However, unfortunately the literature on computer aided approaches to NC image analysis is relatively sparse. Existing approaches [7], [8], [9], [10], [11] typically aim to segment capillaries and analyse the extracted structures.
In contrast, in [12] we have proposed a novel holistic approach for analysing NC images using texture analysis. In particular, we have shown that local binary pattern variance (LBPV) [13] texture features can be successfully employed to distinguish between different NC patterns. In this paper, we build upon this approach and show that by extracting these features in a multi-scale fashion while at the same time maintaining information between the scales, improved recognition performance...