An Automatic Tool for Quantification of Nerve Fibers in Corneal Confocal Microscopy Images

Abstract

Objective: We describe and evaluate an automated software tool for nerve-fiber detection and quantification in corneal confocal microscopy (CCM) images, combining sensitive nerve- fiber detection with morphological descriptors.

Method: We have evaluated the tool for quantification of Diabetic Sensorimotor Polyneuropathy (DSPN) using both new and previously published morphological features. The evaluation used 888 images from 176 subjects (84 controls and 92 patients with type 1 diabetes). The patient group was further subdivided into those with ( n = 63) and without ( n = 29) DSPN.

Results: We achieve improved nerve- fiber detection over previous results (91.7% sensitivity and specificity in identifying nerve-fiber pixels). Automatic quantification of nerve morphology shows a high correlation with previously reported, manually measured, features. Receiver Operating Characteristic (ROC) analysis of both manual and automatic measurement regimes resulted in similar results in distinguishing patients with DSPN from those without: AUC of about 0.77 and 72% sensitivity-specificity at the equal error rate point.

Conclusion: Automated quantification of corneal nerves in CCM images provides a sensitive tool for identification of DSPN. Its performance is equivalent to manual quantification, while improving speed and repeatability.

Significance: CCM is a novel in vivo imaging modality that has the potential to be a noninvasive and objective image biomarker for peripheral neuropathy. Automatic quantification of nerve morphology is a major step forward in the early diagnosis and assessment of progression, and, in particular, for use in clinical trials to establish therapeutic benefit in diabetic and other peripheral neuropathies.

Fig. 1

(a) Original CCM image. (b) Manually quantified CCM image. (c) Automatically quantified CCM image. Red lines represent main nerve fibres, blue lines are branches and green spots indicate branch points on the main nerve trunks. Refer to online coloured version.

Fig. 2

(a) Original CCM image (b) Response image after nerve-fibre detection and denoising (c) Nerve-fibre skeleton with highlighted weak connection segments (d) Nerve-fibre skeleton after assessment of weak connections. (e) Automatically detected end points (hollow circles) and intersection points (solid circles). (f) Final detected nerve fibres.

Fig. 3

(a) Original CCM image with a highlighted segment, a selection of orthogonal profile lines are indicated on the enlarged inset. Profiles are calculated at each pixel along the segment. (b) Average of all the profile lines along the whole fibre segment. (c) The symmetric profile of (b) is firstly calculated, and then normalised (Solid line). A Gaussian distribution is fitted for nerve-fibre width estimation (broken line). The final width equals 2.5 times the RMS width (σ) of the fitted Gaussian curve.

Fig. 4

ROC curves for nerve-fibre detection on dataset 2, using DMNN (Dual Model, Neural Network), DMRF (Dual Model, Random Forest), DTNN (Dual-Tree Wavelet, neural Network) and DTRF (Dual-Tree Wavelet, Random Forest) respectively.

Fig. 5

Boxplots of manually measured features for control, non-neuropathy and neuropathy groups (a) MCNFD (b) MCNFL (c) MCNBD.

Fig. 6

Boxplots of automatically measured features for control, non-neuropathy and neuropathy groups in dataset 2 (a) ACNFD (b) ACNFL (c) ACNBD (d) ACNFA (e) ASDOH (f) ASDWH.