Automatic analysis of diabetic peripheral neuropathy using multi-scale quantitative morphology of nerve fibres in corneal confocal microscopy imaging

Abstract

Diabetic peripheral neuropathy (DPN) is one of the most common long term complications of diabetes. Corneal confocal microscopy (CCM) image analysis is a novel non-invasive technique which quantifies corneal nerve fibre damage and enables diagnosis of DPN. This paper presents an automatic analysis and classification system for detecting nerve fibres in CCM images based on a multi-scale adaptive dual-model detection algorithm. The algorithm exploits the curvilinear structure of the nerve fibres and adapts itself to the local image information. Detected nerve fibres are then quantified and used as feature vectors for classification using random forest (RF) and neural networks (NNT) classifiers. We show, in a comparative study with other well known curvilinear detectors, that the best performance is achieved by the multi-scale dual model in conjunction with the NNT classifier. An evaluation of clinical effectiveness shows that the performance of the automated system matches that of ground-truth defined by expert manual annotation.

Introduction

According to numerous clinical reports (DiabetesUK, 2010), diabetes is among the most challenging chronic health problems. For example, in the UK it is estimated that one in twenty people has diabetes, whether diagnosed or undiagnosed, and by 2025 four million people will have the condition. Damage to the peripheral nerves (diabetic peripheral neuropathy, DPN) is one of the commonest long-term complications of diabetes occurring in at least 50% of patients with diabetes (Boulton, 2005). As a consequence, about one in six diabetic patients have chronic painful neuropathy, compared to one in 20 non-diabetic subjects (Daousi et al., 2004). It is the main initiating factor for foot ulceration, Charcot’s neuroarthropathy and lower extremity amputation. As 80% of amputations are preceded by foot ulceration, an effective means of detecting and treating neuropathy would have a major medical, social and economic impact. The development of new treatments to slow, arrest or reverse this condition is of paramount importance but is presently limited due to difficulties with end points employed in clinical trials (Dyck et al., 2007). Therefore accurate detection and quantification of DPN are important to define at-risk patients, anticipate deterioration, and assess new therapies. Current methods are unsatisfactory, lacking sensitivity and requiring expert assessment, and focus only on large fibres (neurophysiology) or are invasive (skin/nerve biopsy). Unfortunately, diabetic neuropathy lacks a non-invasive surrogate for nerve damage (Tesfaye et al., 2010).

Recent research (Malik et al., 2003, Kallinikos et al., 2004, Hossain et al., 2005) using corneal confocal microscopy (CCM) suggests that this non-invasive, and hence reiterative, test might be an ideal surrogate endpoint for human diabetic neuropathy. The establishment of CCM as a surrogate for early diagnosis and an early biomarker for diabetic neuropathy could identify those at risk and prompt more intense intervention including improved glycaemic, blood pressure and lipid control. Furthermore a sensitive surrogate endpoint would significantly lower hurdles to the development of disease-modifying therapeutics by enhancing the capacity to test therapeutic efficacy. The major advance of CCM is the entirely non-invasive and relatively rapid (≈2 min) acquisition of images of small nerve fibres in patients. However, the major limitation preventing extension of this technique to wider clinical practice is that analysis of the images using interactive image analysis is highly labour-intensive and requires considerable expertise to quantify nerve pathology. To be clinically useful as a diagnostic tool, it is essential that the measurements be extracted automatically.

If an automatic CCM image analysis system is to be applied clinically, especially to define early degeneration or regeneration, then a key step is the automatic detection of low-contrast nerve fibres among image noise (see Fig. 1). The literature on this topic is not extensive, although the problem has a superficial similarity to other, more widely investigated, applications, such as detection of blood-vessels in retinal images. Ruggeri et al., 2006, Scarpa et al., 2008 describe a heuristic method that was adapted from retinal analysis. A number of methods have been developed to enhance the contrast of such linear structures. In a previous study (Dabbah et al., 2009), we used the 2D Gabor filter (Jain and Farrokhnia, 1991) to detect nerve fibres in CCM images. The filter is a band-pass filter that consists of a sinusoidal plane wave with a certain orientation and frequency, modulated by a Gaussian envelope. This spatial domain enhancement is based on the convolution of the image with the even-symmetric Gabor filter that is tuned to the local nerve-fibre orientation. We subsequently extended this to form a dual-model detector (Dabbah et al., 2010), see Section 4.

The automated system of analysing CCM images presented in this paper is an extension of our previous single scale dual-model fibre detector (Dabbah et al., 2010). The new detection algorithm uses the dual-model property in a multi-scale framework to generate feature vectors from localised information at every pixel. These vectors are then used to classify pixels using random forests (RF) (Breiman, 2001) and neural networks (NNT) (Moller, 1993).

In the remainder of the paper we introduce CCM imaging, the image characteristics and the metrics that have been used to quantify the nerve morphology by interactive image analysis (Sections 2 Corneal confocal microscopy, 3 Linear-structure and feature detection). We describe the single-scale dual model filter (Dabbah et al., 2010) and its extension to multiple scales with pixel classification (Sections 4 Single-scale dual-model enhancement, 5 Multi-resolution dual-model enhancement, 6 Nerve fibre classification). In Section 7 we describe a comparative evaluation showing the improved performance of the multi-scale version over not only the single-scale filter but a number of other multi-scale detectors. We also demonstrate that the automatically detected fibres result in morphometric features equivalent to those generated by expert interactive analysis.