An artificial intelligence-based deep learning algorithm for the diagnosis of diabetic neuropathy using corneal confocal microscopy: a development and validation study

The prevalence of diabetic peripheral neuropathy (DPN) can be as high as 50% in an unselected population. Currently, screening for DPN most commonly relies on the 10 g monofilament test, which identifies individuals at risk of foot ulceration but is poor at identifying those with early neuropathy [1]. Screening methods such as clinical examination, questionnaires and vibration perception threshold do not provide direct quantification of small nerve fibres, which are the earliest site of injury. Skin biopsy enables direct visualisation of thinly myelinated and unmyelinated nerve fibres, which are the earliest affected in DPN, and can be used to diagnose small-fibre neuropathy (SFN) [2]. The assessment of intra-epidermal nerve fibres (IENFs) and IENF density is currently advocated in clinical practice in the USA [3] and recommended as an endpoint in clinical trials [4]. However, skin biopsy is invasive and requires specialised laboratory facilities for analysis. The cornea is the most densely innervated tissue of the human body, containing a network of unmyelinated axons (small nerve fibres) called the sub-basal nerve plexus (SBP).

Corneal confocal microscopy (CCM) has been used to image the SBP, which has been shown to be remarkably stable in healthy corneas over 3 years [5] but demonstrates early and progressive pathology in a range of peripheral and central neurodegenerative conditions [6‐11]. Figure 1a,b and Fig. 1c,d show examples from participants without and with diabetic neuropathy, respectively. Previous studies have demonstrated analytical validation by showing that CCM reliably quantifies early axonal damage in DPN [12, 13] with high sensitivity and specificity [14, 15] and closely correlates to the loss of IENFs [15, 16]. CCM also predicts incident diabetic neuropathy [17] and can detect corneal nerve regeneration in people with DPN [18]. CCM may also detect early nerve fibre loss before IENF loss in skin biopsy [19]. In some individuals, corneal nerve fibre (CNF) loss may be the first evidence of subclinical DPN [19]; Brines et al [20] have shown that determination of CNF area and width distribution may improve the diagnostic and predictive ability of CCM.

To accurately quantify CNF morphology, nerves must be distinguished from background and other cell structures accurately. A major limitation for wider clinical utilisation is the need for manual image analyses, which is highly labour-intensive and requires considerable expertise to quantify nerve pathology [21]. The development of methods for the objective, reliable and rapid analysis of corneal nerves is vital if CCM is to be adopted for screening and large clinical trial programmes. Furthermore, to be used as a diagnostic tool, it is essential to extract the measurements automatically with high reliability [21]. Dabbah et al [22] presented a dual-model automated detection method for CNFs using CCM, showing excellent correlation with manual ground-truth analysis (r = 0.92). They further refined this method, using the dual-model property in a multi-scale framework to generate feature vectors from localised information at every pixel, and achieved an even stronger correlation with the ground-truth (r = 0.95) [21]. This study, however, used neural networks without convolution layers [21], which necessitates pre-processing and encourages overfitting.

Kim and Markoulli [23] developed a nerve segmentation technique to delineate corneal nerve morphology in CCM. This involved processes ranging from filtering methods (with rapid implementation but low-contrast and imprecise focus) to more complex support vector machine approaches (which rely on features defined by the user). Chen et al [24] presented a method based on feature engineering, achieving state-of-the-art results, although its reliance on hand-crafted features increases the complexity to the user and can introduce user-bias, returning suboptimal results [25].

Recently, approaches based on machine learning have achieved excellent performance in computer vision and medical image analysis tasks. Deep learning and, particularly, convolutional neural networks (CNNs; a class of deep neural networks) have emerged as a highly effective branch of machine learning for image classification [25]. This approach allows for ‘end-to-end’ classification results to be achieved without the need for specifying or designing features or setting example-specific parameters. CNN design follows vision processing in living organisms [26], with the connectivity pattern between neurons resembling visual cortex organisation. Based on training with pre-annotated data, CNNs combine the traditionally separate machine-learning tasks of feature designing, learning and image classification in a single model, relieving the traditional machine-learning burden of designing hand-crafted features. More recently, this has extended beyond image-wise classification to efficient pixel-wise classification, allowing image segmentation to be achieved (i.e. pixels may be classed as belonging or not belonging to an object of interest). There has been a significant increase recently in the development of deep learning algorithms (DLAs) with CNNs, an approach that achieves excellent performance in many computer vision applications and has clinical utility in healthcare [27]. Compared with manual detection, accurate automated detection of corneal nerves using CCM has many potential benefits, including objectivity, increased efficiency and reproducibility, allowing enhanced early disease diagnostics and improved patient outcomes. Artificial intelligence-based DLAs have the added advantage of continual learning and refinement alongside concurrent analysis.

The aim of this study was to develop and validate a DLA for corneal nerve segmentation in CCM images and to compare this with the widely used and validated automated image analysis software, ACCMetrics (Early Neuropathy Assessment [ENA] group, University of Manchester, Manchester, UK) [24].

Figure 3 shows four example testing images along with their ‘ground-truth’ manual annotations and segmentation results obtained by LCNN, LDLA and ACCM. LCNN and LDLA produced results that were more faithful to the manual annotations than did the ACCM, particularly in example 4 where the ACCM failed to detect nerves at the top middle and right bottom of the image. Overall, the segmentation performance was consistent and there was no obvious failed case. For illustration, electronic supplementary material [ESM] Fig. 1 shows all the first 30 images of dataset 1, ESM Fig. 2 shows the 12 randomly chosen images from dataset 2, and ESM Fig. 3 shows 12 randomly chosen images from dataset 3.

Analysis of datasets 1 and 2 shows that the mean total CNF length from the manual ‘ground-truth’ annotation was highest (2441.4±919.5 μm) compared with the three automated approaches (LCNN 2089.4±804.6 μm; LDLA 2260.3±835.3 μm; ACCM 2394.1±768.1 μm). Total CNF length was greater in the ACCM and was closer to the total length from the manual annotation than were the LCNN or LDLA results. However, ICC analysis (Table 1) demonstrated that the LCNN and LDLA both produced results more consistent with the manual annotations when compared with the ACCM. Furthermore, our two methods performed consistently better than the ACCM in terms of correct segment length, number of branching points and fractal numbers. Bland–Altman analysis (Fig. 4) further confirmed that the limits of agreement of the ACCM were greater than those of both the LCNN and LDLA, implying greater variability despite the mean total corneal lengths in the ACCM. In other words, although the results of the ACCM were closer to the manual annotation in this case, the variation due to over- and under-segmentation was much larger than either the LCNN or the LDLA; the ACCM may therefore have produced heterogeneous results.

Based on the 95% CI of the ICC estimate, a value of <0.5, 0.5–0.75, 0.75–0.9 and >0.90 is indicative of poor, moderate, good and excellent reliability, respectively [36].

Table 2 shows the comparisons of the root mean square error and SD of the derived measures v_i over each image i, against the manual annotations, in terms of number of branching points, number of terminal points, number of segments, total nerve fibre length, mean nerve fibre length, SD of nerve fibre length, and fractal number for each of the methods M using:

As shown in Table 2, LDLA had lower values for every measure, indicating closer agreement with the ground-truth annotation. For each measure, the LCNN had the second-lowest root mean squared error (RMSE) and the ACCM had the highest, indicating weaker agreement. The LDLA had the lowest SD for all measures except the number of terminal points, indicating more consistent agreement with the ground-truth over the set of images; the ACCM had the highest SD for all measures. From this, it can be concluded that both the LCNN and the LDLA outperform the ACCM and that the LDLA clearly has the best performance.

Given the convincing performance of the LDLA, which outperforms both the LCNN and the ACCM in each metric, it was applied to the third dataset and the results were used for clinical evaluation.

In this study, an artificial intelligence-based DLA has been developed for the analysis and quantification of corneal nerves in CCM images. To our knowledge, this is the first DLA for the analysis of corneal nerve morphology and pathology. This study validates our DLA and demonstrates its superior performance compared with ACCMetrics, the existing state-of-the-art system. In particular, there are more consistent results, as demonstrated by a superior intraclass correlation for a number of metrics including total CNF length. In addition to the total CNF length, this DLA is also capable of producing the number of branching and tail points, fractal numbers, tortuosity and segment length. As such, these quantitative variables may provide additional utility to diagnose diabetic neuropathy and neuropathic severity.

A fractal is a visual product of a non-linear system characterised by its complexity and by the quality of self-similarity or scale invariance. Fractal analysis of the corneal SBP has been proposed by several authors [38, 39]. We believe that the additional utility of fractal dimensions provide an additional means of differentiating individuals with early or subclinical DPN. CNF length is a robust measure of DPN and SFN. A large multicentre pooled concurrent diagnostic validity study revealed that CNF length was the optimal CCM variable [40]. CNF length has also been shown to be a measure of early small-fibre regeneration [41]. From published data, CNF length and density are the most robust measures of DPN. Our data confirms the validity of CNF length. However, we feel other metrics are also of importance and require further scientific interrogation in a real-world clinically oriented study.

In this study, the quantification of images demonstrates a reduction in total CNF length in individuals with diabetic neuropathy compared with healthy volunteers. This study is in keeping with other data on the utility of CNF length as a valid biomarker of diabetic neuropathy [11, 12, 15, 18]. The sensitivity and specificity of our DLA for gold-standard DPN diagnosis with CCM (using the Toronto criteria) is far superior to currently used clinical methods such as the 10 g monofilament and 128 Hz tuning fork [42] (with rudimentary clinical assessments), thus providing a strong rationale for its use in clinical screening/practice.

This study extends our preliminary work on 584 CCM images where the initial DLA demonstrated good localisation performance for the detection of corneal nerves [43]. Our preliminary model was refined to produce the ensemble (LDLA) model, now validated in large image datasets to diagnose diabetic neuropathy using CCM.

The strength of deep learning is echoed by Oakley et al [44] who used corneal nerve segmentation in CCM images from macaques. Deep-learning-based approaches make the segmentation task relatively easier for the end user compared with the conventional approaches employing various filters and graphs [23]. In particular, compared with conventional machine-learning methods, such as support vector machines (SVM), deep learning reduces the need and additional complexity of feature selection and extraction, allowing the computer to learn features alongside the segmentation. The training of deep learning approaches is computationally expensive (e.g. it takes approximately 30 min per epoch to train a single U-Net model). The advantage is that, once the model is trained, the segmentation is very fast, taking milliseconds to segment CCM images.

In recent years, CNNs and DLAs have been added to algorithms used to screen for diabetic retinopathy. DLAs promise to leverage the large number of images for physician-interpreted screening and learn from raw pixels. The high variance and low bias of these models will allow DLAs to diagnose diabetic neuropathy using CCM images without the pre-processing requirements and more-likely overfitting of earlier approaches [25]. This automated DLA for the detection of diabetic neuropathy offers a number of advantages including consistency of interpretation, high sensitivity and specificity, and near instantaneous reporting of results. In this study, good sensitivity and adequate specificities were achieved using our DLA.

This is the largest study, to date, of the development and validation of corneal nerve segmentation and supersedes the numbers in the study by Chen et al [24], who used 1088 images from 176 individuals, with 200 images for training and 888 for testing. Our study used a robust dataset; however, further development of this DLA requires use of a developmental set of images with large numbers (tens of thousands) of normal and abnormal pathologies. An area of further research is that of interrupted CNF segments, which have often proved challenging in the CNF segmentation results obtained using earlier methods [24]. This problem is mainly caused by non-uniform illumination and contrast variations of CNF in images. Since quantitative biomarkers like CNF length and density are important measures for computer-aided diagnosis, missing CNF segments may theoretically reduce the diagnostic reliability of any automated system. In our previous work, the automatic gap reconnection method proposed by Zhang et al [45] was employed to bridge the interrupted nerve fibre structures. The gap-filling task is achieved by enforcing line propagation using the stochastic contour completion process with iterative group convolutions. Geometric connectivity of local CNF structures can be easily recovered based on their contextual information [45]. However, this connection step was not included in this model as there was only a modest improvement in the quantification of CNF length despite extra computation time of about 1 min per image. This is an area for future development of the DLA. It will also be important to investigate the potential for bias to be introduced by factors such as camera type. The major advantage of this DLA over standard automated techniques is the continual learning and refinement of the algorithm.

Given that 420 million people worldwide have been diagnosed with diabetes mellitus [46] and that the prevalence of diabetic neuropathy is ~50% [47], there is a need for valid quantitative population screening for diabetic neuropathy to prevent or limit sequelae such as foot ulcers and amputations. Skin biopsy with quantification of IENFs has been considered the ‘reference standard’ test for the diagnosis of SFN [48]. It is an invasive test, needing specialist diagnostic facilities and repeated tests at the same site, which is not always feasible. CCM is a rapid non-invasive ophthalmic imaging modality, which quantifies early axonal damage in diabetic neuropathy with high sensitivity and specificity [12‐16, 49]. CCM also predicts incident neuropathy [17] and accurately detects CNF regeneration [18, 50]. The utility of CCM in diagnosing and monitoring the progression of diabetic neuropathy has been extensively evaluated [11, 12, 15, 16, 18, 38].

Further studies are required to determine the feasibility of applying this algorithm in the clinical setting and to compare outcomes with those obtained from currently used diabetic neuropathy screening methods typically having low sensitivity for detection except in advanced neuropathy. There is also a need to compare the diagnostic ability of this DLA with tests of small-fibre dysfunction (thermal thresholds/sudomotor/autonomic) and IENF density in skin biopsy in DPN and other peripheral neuropathies. The next key step is also to utilise the DLA alongside clinical neuropathy screening in a multicentre primary care study.