Deep Learning based Skin-layer Segmentation for Characterizing Cutaneous Wounds from Optical Coherence Tomography Images

The wound healing process is a complex procedure that involves four distinct stages: homeostasis, inflammation, proliferation, and remodeling. This process involves a variety of cells, growth factors, and cellular and molecular mechanisms, and its pathophysiology is still being studied to optimize treatment and improve patient outcomes. Chronic wounds, such as diabetic and venous ulcers, have stopped healing normally due to certain pathophysiologic conditions. The gold standard for precise tissue architecture assessment is still histological analysis, but this process is invasive and can cause additional trauma. Most wound care prediction and monitoring currently relies on visual evaluation, leaving a critical need for novel, non-invasive strategies for individualized diagnosis and care[1]. Imaging technologies can provide quick, objective, and non-invasive assessments of wound severity, potential for healing, and healing progress by measuring the optical characteristics of skin components. Various imaging modalities have been used to study cutaneous wounds, including near-infrared imaging, thermal imaging, optical coherence tomography (OCT), fluorescence imaging, and laser speckle imaging. OCT in particular provides high-resolution images of tissue microstructure and has been used to assess burn wound severity, monitor scar progression, and investigate chronic wounds. Since OCT can accurately distinguish between the epidermal and dermal layers, it can be utilized in clinical settings to track wound re-epithelization[2]. To establish OCT as a standalone tool for non-invasive wound evaluation, it is necessary to correlate its data with corresponding histological data[3] and validate it with a large sample set of optical data. The present pilot study aims to use OCT as a precise and accurate tool for non-invasive wound evaluation by developing a Convolutional Neural Network (CNN) based model that can automatically segment OCT images to annotate different layers of normal and wounded skin tissue and help quantitatively analyze the healing stage, epithelial thickness, regenerated tissue thickness, and wound closure dynamics. Previous research has focused on the use of deep learning and other techniques for the segmentation of the epidermis layer in OCT images, which is useful for analyzing wound epithelial reformation[4],[5],[6],[7],[8],[9],[10]. However, segmenting deeper layers such as the dermis and subcutaneous tissue is equally important to assess the dynamics of wound healing[11] fully,[12],[13]. Our study aims to fill this gap by developing a U-Net-based approach for segmenting three layers (epidermis, dermis, and subcutaneous) of wounded skin in OCT images. We selected the U-Net architecture because of its effectiveness in biomedical image segmentation[14].
In our work, we tackled the difficulties posed by autonomous systems, particularly in optical coherence tomography datasets, such as limited training data and image data distribution dependency. To overcome these obstacles, we utilized transfer learning, a common approach in the field. Using segmentation models with pre-trained models, we could apply Machine Learning (ML) models in situations where large, well-curated training datasets are difficult to obtain. In the case of optical coherence tomography, where creating high-quality annotated image datasets is both costly and technically challenging, reusing pre-trained models helps to implement ML models more efficiently. Our study utilized the base U-net and used pre-trained models such as VGG16, ResNet34, and InceptionV3.
Transfer learning helps address the problem of limited data, but uncertainty quantification is computationally expensive with single models, leading to an ensemble of pre-trained models. Ensemble learning is a method in artificial intelligence (AI)[15] where multiple models or experts are used to solve an AI problem. Suppose we have k models in an ensemble, denoted as $(m_{1},...,m_{k})$, and let $p(x=y_{i}|m_{j})$ denote the probability that input x is classified as $y_{i}$ under model $m_{j}$. Then, the ensemble prediction is given by:

The goal is to increase diversity among the models, accurate segmentation, good calibration, and reliable uncertainty quantification, and avoid overfitting. The final prediction is obtained by combining the outputs of the individual models. We also compared the performance of multiple pre-trained models as the encoder part of the U-Net. We evaluated the potential improvement obtained through the ensemble of those models. We calculated the wound bed area from the segmented output. The details of the ensemble design, transfer learning, and training dataset preparation are outlined in Section 2, and the significant results and discussion are provided in Section 3.

The results as demonstrated in Table I show that all models successfully segmented the different layers in the wound images. Figure 2 displays the progression of wound healing over time, demonstrating the ability of the models to accurately track the healing process, saving time and effort. This also allowed for the calculation of wound bed area and the development of a wound healing progression function (Figure 3). The use of a patch-based augmentation technique and small batch size resulted in a fast training time, as evidenced by the training graphs (1). Although all models performed similarly with minor differences in per-class segmentation. However, the ensemble model emerged as the best performer, combining the best results from all models. The wound healing progress (Figure 2) shows that the healing process is not linear but shows only slight changes during day 0 to day 4, and the healing process accelerates thereafter. This demonstrates the critical need of proper wound curation during the initial day post injury. The healing process from day 4 to day 12 (last imaging trial) shows rapid improvement and more than 98% wound are healed by day 12, unless there are specific aberrations or infection scenarios. However, such cases are beyond the scope of the current study. These results were further corroborated by using histopathology samples (excised wound region). In Figure 4 we show the differences between normal and healed wound regions using Hematoxylin and Eosin (HE) stained histopathology section. The histopathology studies clearly highlights the differences in the thickness of keratinized layer and the density of collagen in the granulation region. These features can be clearly be seen in the microscopic images and further characterized using image analysis methods to strengthen our understanding of the wound healing process [20, 21].

In summary, our study demonstrates the effectiveness of deep learning models, specifically U-Net and its variants, in segmenting different skin layers from OCT scans of dorsal cutaneous wounds in mice. We successfully demonstrated the effectiveness of automatic segmentation in calculating the wound bed area to annotate the different stages of wound healing. This is a novel domain of exploration in medical imaging and machine intelligence. The initial sets of experiments have been carried out on two Swiss albino mice, which shows significant feature variations in skin characteristics such as the mean thickness of skin layers, hair density, and concomitantly, high light absorption. However, wounds in mice heal typically by contraction, which gives us a simple mechanism to study for the initial assessment of our tool. We will expand the scope of our studies to include more animals and gradually shift to large animal models which are better candidates for studying wound healing in preclinical phases. In our work, we evaluated various U-Net models and demonstrated the impact of patch-based data augmentation and small batch size on the performance of deep-learning models in medical imaging. Our comparison results indicate that the ensemble model, which combines the best results from multiple models, was the top performer. The findings suggest that deep learning has the potential to be a valuable tool in analyzing wound healing progress [22]. Additionally, these results conclusively show the relevance of OCT and similar sub-surface imaging methods in studying wound healing dynamics and changes in tissue characteristics during the healing process. Such investigation can be useful in clinical cases of non-healing and infected wounds. Further, research and joint feature learning from multi-modal and multiscale imaging [23, 24] are needed to improve accuracy and address the limitations [25]. Further improvements of relevant techniques can usher in a new era of digital pathology[26], improve access to care in resource-limited settings, and alleviate the clinical and economic burden of wound care in the tropics[27].

The authors acknowledge the support of Dr. Shirin Dasgupta,Senior Resident (Pathology) at Dr. B. C. Roy Multi Speciality Medical Research Centre, Indian Institute of Technology, Kharagpur, toward her support for data annotation and study validation.