Prediction of peripheral nerve intrinsic organisation: a pilot study

ABSTRACT

Traumatic peripheral nerve injuries (PNI) are debilitating and can leave patients severely limited with drastic changes in quality of life. PNI treatment options are varied; however, only 50% of those patients regain any useful function. The complex intrinsic organisation of nerves is largely contribute to this issue. This research aimed at creating a reliable way of isolating and predicting the location of peripheral nerve fascicles on microscopic images of nerve cross-sections via semantic segmentation and machine learning. Using the Anaconda Python platform, a machine learning algorithm was programmed using a random forest classifier. Training data resulted in a 91% accuracy rate with respect to correctly identifying image pixels as fascicle or not. When the trained model was tested in the real world with unseen data, the results varied. However, the model retained the ability to identify fascicles and isolate them from background. The development of clinically successful PNI treatments has long been hindered by intrinsic nerve structures that machine learning may be able to solve. With further development, this research has the potential to expand and form new frontiers in the field of clinical medicine and drastically improve outcomes of PNI.

KEYWORDS:

• Machine learning
• semantic segmentation
• peripheral nerve injury
• fascicles
• nerves

1. Introduction

1.1. Peripheral nerve injury overview

The peripheral nervous system is comprised of nerves that contain sensory or motor axons which run from the spinal cord to innervate different organs and tissues in the body (Gordon 2020). Nerves have a highly organised structure (Gordon 2020). There are four connective tissue structures that split the nerve into organisational strata. The endoneurium surrounds individual axons. Sensory or motor axons are grouped together into fascicles isolated by a perineurium. Finally, fascicular groups are separated from one another, yet held under the same visualised peripheral nerve by the epineurium and intra-fascicular epineurium. The fourth connective tissue carries the vascular supply to the nerve and is named the mesoneurium (Gordon 2020).

Peripheral nerve injury (PNI) occurs in approximately 2% of people who have had trauma involving the upper or lower limb. Although there are many treatments and procedures available for the treatment of PNI, such as conduits and other cable-like structures, the outcomes of these treatments are not ideal (Stewart 2003; Taylor et al. 2008; Bergmeister et al. 2020). Peripheral nerves are known to carry fascicles of nerve fibres to various destinations throughout the body (Gordon 2020). Fascicular organisation in peripheral nerves has been determined to have both a cable-like organisation, where fascicles run beside each other (distal in the body), as well as a plexiform organisation (proximal in the body), where the fascicles merge and separate multiple times, and in seemingly random order, with other fascicles throughout their length in a nerve (Stewart 2003; Gordon 2020; Bergmeister et al. 2020). However, it is still unclear how the nerve fascicle organisation could affect the level of functional recovery after a peripheral nerve allograft. Furthermore, understanding how fascicles arrange themselves throughout the length of a nerve and whether this arrangement differs between nerves is vital to the surgical approach of treating PNI with nerve allografts.

Only about 50% of patients who are treated for PNI with current treatment modalities will regain useful function (Stewart 2003; Taylor et al. 2008; Bergmeister et al. 2020). Treatments of such injuries are relatively few and limited. Most are related to approximating the two stumps of a transected nerve, isolating the nerve from surrounding scar tissue, or grafting a conduit or autologous nerve. All these forms of repair do not bypass the complex and prolonged innate mechanisms of peripheral nerve regeneration and repair. Most techniques used today provide structural support for sprouting axons to reach their original innervations. Therefore, more precise methods of recreating the intrinsic nerve structure can only benefit axonal sprouting and regeneration. New techniques developed in nerve grafting and stump approximation involves intra-nerve dissection and direct fascicle-to-fascicle suturing, which attempts to breach the issue of fascicular misalignment. However, results have not been promising due to excessive internal inflammation and scar tissue formation, which is counterintuitive to the initial purpose of this technique (Stewart 2003; Bergmeister et al. 2020).

Depending on the treatment or procedure being utilised, factors such as unknown fascicular topography and unstandardised surgical techniques lead to poor outcomes in patient functionality after treatment (Matsuyama et al. 2000; Taylor et al. 2008; Moore et al. 2011; Delgado-Martínez et al. 2016). Nerve graft transplantation is another treatment technique that warrants refinement and standardisation. Nerve grafts are used in large segment defects, and when other treatments are not available. Nerve grafts are essentially conduits whose axons undergo degeneration in hopes that the remaining endoneural network will aid the regenerating nerve stump (Delgado-Martínez et al. 2016). Procedures involving nerve grafts should have proper graft screening methods to determine which nerve graft will provide the best outcomes for patients. Features such as graft fascicle organisation, graft size, and graft laterality could help standardise nerve transplantation procedures.

1.2. Image segmentation applied to nerve morphology

Image segmentation is an image processing technique used to identify, quantify, or separate certain objects, features, or textures in an image (Pednekar et al. 2018). Image segmentation has long been used in countless fields, which include security, military targeting systems, materials engineering, biochemistry, and molecular biology (Pham et al. 2000). It is a highly efficient and reliable. Image segmentation can be classified further into different types (Pednekar et al. 2018). However, in this study, we used semantic segmentation, which is characterised by an ‘all or nothing’ technique to classify pixels in an image as either ‘object’ or ‘non-object’. Through software programing and the integration of machine learning, a program can be trained to identify objects in an input image. In this specific case, we developed and trained a program to identify fascicles on human nerve cross-sections.

Machine learning and artificial intelligence has been making headway in the field of diagnostic medicine (May 2021) and aiding in objectifying and standardising decision making with respect to certain interventions. The current literature contains a multitude of publications that speak to the inclusion of machine learning technology that spans the field of medicine.

Peiffer-Smadja et al. 2020 (Peiffer-Smadja et al. 2020) reviewed and discussed the available literature on the clinical use of machine learning in patient care with respect to identifying infectious diseases. Machine learning programs are being integrated into primary, secondary, and tertiary care settings to aid clinicians in identifying infectious and septic patients. Artificial intelligence (AI) is making leaps in the medical field and its reach is broadening. Adlung et al. 2021 (Adlung et al. 2021) reviewed the current uses of AI and machine learning in the clinical decision-making process. Machine learning has been used by Google to predict 3D structures of amino acid sequences to better classify targets for novel pharmacologic therapy. Machine learning is also being developed to better analyse and identify pathology and neoplasms on radiologic images (Alapat et al. 2022) and improve the diagnostics of these conditions in addition to becoming a step towards individualised precision medicine.

2. Materials and methods

2.1. Human peripheral nerve samples

This work was performed with the approval of IBC VT # 21–049 and VCOM IRB Record # 2021–023. Obturator nerve samples were gathered from 25 formalin-fixed whole-body donors. Grant’s dissector was used to locate and isolate the obturator nerves (Detton 2020). Nerves from both right and left lower extremities were harvested to produce a total of 50 samples. The proximal nerve end from each donor was marked to maintain orientation. Dissected nerves were placed in formaldehyde containers separately to maintain preservation. Containers were labelled with donor number and laterality (R or L obturator nerve). These samples were stored in a refrigerator until analysis could be performed.

2.2. Peripheral nerve imaging

To acquire cross-sectional images, each individual whole-nerve was divided into three sections (proximal, medial, and distal). Each section was further divided into three parts (1,2 and 3). This method yielded nine total cross sections per nerve. The cross section of each part was submerged in buffer solution and imaged using light microscopy with an OptixCamm Summit K2 camera attachment (Yount et al. 2023). The highest quality image was achieved at 4× magnification, and each image was saved for future steps.

2.3. Data set and semantic segmentation

We retrospectively created a database of a mix of nerve images, and we developed precise definitions of the nerve fascicles to be able to be visualise by the next procedure (Figure 1).

Figure 1. A simplified diagram that depicts the mechanism of segmenting a pixel within an image using a random forest classifier. The input data is split into test data and training data. Training data comprised of many pixels are individually run through an algorithm of different decision trees. Each branch of a tree is a feature used to extract information from that pixel. Labels are collected from multiple trees and the pixel is given a final label that agrees with the majority of decision trees or in some cases the final label is an average of all preliminary labels.

Semantic segmentation was performed using a random forest classifier on the Anaconda Python platform. Several image processing features and filters were applied to the images to best isolate the fascicles. To train the machine learning program, seven microscopic images were used to create masks by manually labelling the fascicles using the open-source platform LabelStudio. The program was trained using the seven masks and seven raw microscopic images. Predictions were later run using a new set of unseen microscopic images and a prediction code that utilises the trained model (Figures 1 and 2). Pixel accuracy and intersection over union (IoU) were calculated by creating separate masks for those untrained images. Images were selected based on the clarity of picture as well as the quality of the nerve (Figure 2).

Figure 2. a1 and b1 depict masks created by manually labeling fascicles on LabelStudio. These masks are used to train the program on identifying fascicles within a nerve. a2 and b2 are the respective raw microscopic images of the masks. They depict the nerve cross-sections under light microscopy with 4x magnification. Images were captured using OptixCamm Summit K2 camera attachment with an external source of light.

3. Results

The trained machine learning program resulted in a 91% pixel accuracy rate. The accuracy is derived from how successful the training model was able to match the masks that were manually created. It is derived by the sum of true positives and true negatives divided by the total number of pixels. The code initially set aside a 1/3 of the raw input data for accuracy testing. 2/3 of the pixels are then used to train the model using the masks as guidance. Once the programme has completed its training the final decision-making process was then tested using the remaining set-aside data and assigned an accuracy based on whether it made the correct choice. If one were to plug one of the training images back into the trained prediction model it will identify approximately 91% of the pixels correctly, which can be seen in Figure 3.

Figure 3. Raw microscopic image (left) that was used for training is plugged back into the prediction model which generates a semantic segmentation with an approximately 91% accuracy rate (right).

When we tested new and previously unseen raw images without pre-processing and reduction of background (Figure 4), it is evident that the accuracy is less than 91% when compared to Figure 3. However, it seems like the program is still able to identify fascicles well despite the artefacts. In Figure 4, the top images demonstrate the raw microscopic images, and the bottom images show the model’s ability to identify the fascicles. In Figure 4(c) one can note what appears to be a bifurcating or trifurcating fascicle at the top of the image. This structure was correctly segmented. Also, in Figure 4(a) the model correctly left out fat globules in the centre of the nerve which are sometimes even hard to differentiate with the naked eye. The prediction model was able to discern complex features and structures of the nerve. Table 1 shows the respective pixel accuracy (Pixel Accuracy = True Positive + True Negative/True Positive + True Negative + False Positive + False Negative) and intersection over union (IoU) of those 3 images. IoU is calculated by taking the true positives only and dividing it by the sum of true positive, false positive, and false negatives (IoU = TP/TP + FP + FN). It, therefore, compares what the prediction model identified as fascicles only and does not consider how successfully it left out background or intrinsic structures that were not fascicles.

Figure 4. a, b, and c show the results of the trained machine learning program on previously unseen data (bottom) and their respective raw microscopic images (top).

Table 1. Depicts the accuracy and intersection over union of the predicted segmentations in Figure 4 when compared to ground truth.

Image	Accuracy	IoU
4A	0.86	0.50
4B	0.86	0.46
4C	0.83	0.46

Lastly, we used our images in another available image segmentation tool (Figure 5). As it is possible to observe there is a reduction of accuracy in Figure 5(b) in comparation to our method Figure 5(a), which could be due to non-specificity, differing from our method that was trained to specifically identify nerve fascicles.

Figure 5. Comparison with other available images segmentation tools. We compared our outcomes with other available tools. (a) Our segmentation. (b-c) segmentation done by another tool.

4. Discussion

The purpose of this research was to find ways to improve graft and conduit selection in peripheral nerve transections. Grafts and conduits serve as structural support for sprouting axons that originate from the proximal stump of a transected nerve. They do not transplant axons but they act as a scaffold and bridge, which allows axons to regrow towards its original innervation sites through the distal stump. A big hindrance to this process is the variance of fascicular organisation between individual nerves. Therefore, grafting does not maintain the integrity of the microanatomical organisation and conduits do not provide internal structure. This leads to axonal loss and an inability reinnervate end-organ targets. Thus, an ability to match grafts and in-vivo nerves at the fascicular level can drastically improve outcomes in peripheral nerve injury repair.

Fascicular repair and bridging have been attempted and are still used in limited indications. However, the results were not better than simple epineural repair (Winn 2022). First, the use of intraoperative microscopes and the suturing of two nerve stumps in-vivo requires an extremely high level of dexterity. The time and effort required to visualise fascicles and attempt to align them is not only impractical, but also the risks associated with doing so can be futile and outweigh the potential benefits. Manipulation of the fascicles can and does inadvertently cause damage to the intraneural structures and requires much more suturing which forms obstructing scar tissue and inflammation (Winn 2022). Longer operative times also place the patient at risk for infection and increased risk of adverse effects of anaesthesia (Cheng et al. 2017). Furthermore, the correct evaluation of the nerve quality after the injury and post operatory is still a limitation (Luzhansky et al. 2019). This research aims at innovating a way to maximise fascicular overlap while minimising manipulation and damage efficiently and practically. It also allows the ability to further develop and innovate individualised nerve conduits or individualised selection of allografts at the microanatomical level, leveraging the use of tissue banks in peripheral nerve repair. This is the first step in a multi-step approach to improve the repair of peripheral nerve transactions. As a recent revision article in imaging in the repair of peripheral nerve injury showed, it has been extremely challenge to have on hands and standardised the most adequate imaging modality to evaluate peripheral nerve quality, injury level, and repair (Luzhansky et al. 2019).

The machine learning model that was created was able to identify fascicles and isolate them from surrounding structures. These results are an improvement upon the available options today for selecting nerve grafts and conduits. With continued research and development, a real alternative method of grafting peripheral nerves to accommodate for each individual patient anatomy is possible and this is the first step towards that process.

Limitations arose during the research and construction of this program. With limited sample sizes, overfitting is a potential confounder. Overfitting is an issue that is commonly encountered in machine learning, as a form of overestimation. The machine learning program gives weight to certain features of the data that is counterintuitive to the goal (Ying 2019).

This can be illustrated by a simple example: if a program is built with the desire to identify dogs in images, one would begin by acquiring training images. So, one can go to a dog park and take pictures of dogs at the park. Those images are used to train the program. The results look accurate at first. However, once this code is implemented with unencountered images things can change. You decide to identify a picture of a dog in your living room, and it does not see or identify the dog. You then and take a picture of a stray cat outside. The program identifies the cat as a dog. You realise that the machine learning program has associated greenery or the sky with dogs, which is not true.

This simplified explanation of overfitting can cause trouble with achieving accurate results and is not always so obvious with its various complexities. One possible way to avoid this is increasing training data. The more variable data that a machine learning program can use the more it is able to pick up on subtle differences and avoid generalisations or the use of inappropriate features to train itself. Artificial intelligence and machine learning is ‘data-hungry’. The more data there is to be trained the better and more intelligent the program gets. The limiting factor is most often computing power, and at this moment we have reached the capacity of our available hardware and donor sample size (Ying 2019).

When it comes to our findings, they are not unique to this research alone. Ge et al (Ge et al. 2023). published a paper on the exact phenomena we faced, which is generalisability. Generalisability refers to how well the machine learning model performs in the real world. The same authors referred to this difference between performance during evaluation and the real world as the ‘AI chasm’, which is prominent when it comes to machine learning and medicine (Ge et al. 2023). While building the program the feedback seems great and, as we experienced, shows a high accuracy rate and good results with training images. However, when the model is placed in a real-world setting and given unseen data with confounding variables the model underperforms.

Nonetheless, machine learning has the capability to change the field of medicine positively (Bilodeau et al. 2022). The results of this research still showed promise in the ability to use machine learning and image segmentation to identify fascicles. More work needs to be done to increase sample sizes and the quantity of training data as well as developing the machine learning program further to facilitate consistent and accurate results in the real world. Next steps will also integrate the use of high-resolution ultrasound imaging to better conform this work into the clinical setting and allow it to be used pre-operatively as well. Building upon this research can help form new techniques for individualised patient care and treatment in peripheral nerve injuries.

5. Conclusions

Machine learning has the potential to be a great addition to the diagnostic and intraoperative algorithm of peripheral nerve injury repair. Image segmentation of fascicles can lead to improvements in grafting success and restoring function in PNI by enhancing the accuracy of graft selection to best accommodate for the intrinsic organisation of nerves. Furthermore, when we contacted a surgeon that works with peripheral nerve repair to know his opinion about our pilot research, it has been clear that our tool could be used to improve the assessment of nerve quality before and after the repair, as well as it could be used for standardisation of nerve placement in the case of peripheral nerve transplantation. Thus, we believe that our research is a promising first step in the use of artificial intelligence and machine learning in peripheral nerve repair. Future research should be directed at expanding the scope of use of machine learning and image processing in peripheral nerve injuries. We see benefits in quantifying fascicular numbers and locations within nerves to identify patterns across different intrinsic factors and disease processes, as well as using more clinical-based imaging techniques like high-resolution ultrasounds.

Acknowledgements

We would like to thank Edward Via College of Osteopathic Medicine for all support. This research received intramural funding. We also would like to thank Dr. Hallan Bertelli for his feedback as a surgeon orthopedist.

Disclosure statement

No potential conflict of interest was reported by the author(s).