An Enhanced Harmonic Densely Connected Hybrid Transformer Network Architecture for Chronic Wound Segmentation Utilising Multi-Colour Space Tensor Merging

TL;DR

The paper tackles the problem of chronic wound segmentation across diverse skin tones by enhancing HarDNet through multi-colour space tensor merging and a suite of architectural refinements (IBN, SN, PReLU, harmonic block adjustments). It introduces RGB+eY colour-space merges and GAN- and animal-meat–based pretraining to boost learning, plus cross-domain weakly supervised training to expand limited data. Quantitatively, HarDNet-CWS improves IoU by $+0.1274$ and DSC by $+0.1221$ on a dark-skin test set with ground truth, and qualitative clinician ratings show notable improvements, underscoring the value of colour-space information and robust normalization. The work also engages in a comprehensive reliability study with expert raters and highlights the persistent need for standardized evaluation in chronic wound segmentation, especially for underrepresented skin tones, with implications for remote monitoring and equitable care.

Abstract

Chronic wounds and associated complications present ever growing burdens for clinics and hospitals world wide. Venous, arterial, diabetic, and pressure wounds are becoming increasingly common globally. These conditions can result in highly debilitating repercussions for those affected, with limb amputations and increased mortality risk resulting from infection becoming more common. New methods to assist clinicians in chronic wound care are therefore vital to maintain high quality care standards. This paper presents an improved HarDNet segmentation architecture which integrates a contrast-eliminating component in the initial layers of the network to enhance feature learning. We also utilise a multi-colour space tensor merging process and adjust the harmonic shape of the convolution blocks to facilitate these additional features. We train our proposed model using wound images from light-skinned patients and test the model on two test sets (one set with ground truth, and one without) comprising only darker-skinned cases. Subjective ratings are obtained from clinical wound experts with intraclass correlation coefficient used to determine inter-rater reliability. For the dark-skin tone test set with ground truth, we demonstrate improvements in terms of Dice similarity coefficient (+0.1221) and intersection over union (+0.1274). Qualitative analysis showed high expert ratings, with improvements of >3% demonstrated when comparing the baseline model with the proposed model. This paper presents the first study to focus on darker-skin tones for chronic wound segmentation using models trained only on wound images exhibiting lighter skin. Diabetes is highly prevalent in countries where patients have darker skin tones, highlighting the need for a greater focus on such cases. Additionally, we conduct the largest qualitative study to date for chronic wound segmentation.

Figures (12)

• Figure 1: Illustration of an image from the DFUC2022 training set and corresponding masks: (a) original image; (b) original mask based on clinician delineation; (c) original mask processed using active contour model; (d) original image with clinician delineation mask overlaid; and (e) original image with original mask processed using active contour model overlaid. Note that images were cropped for illustration purposes.
• Figure 2: Illustration of 3 cases from the DFUC2022 dataset where clinical experts determined the baseline model predictions (HarDNet-DFUS) to be superior to the ground truth labels. The first column shows the original images, the second column shows the ground truth, and the third column shows the model predictions.
• Figure 3: Illustration of 2 cases from test set B showing the Y channel and its alternate representations. Y - luminance, eY - exaggerated luminance, eYS-R&B - exaggerated luminance with swapped R and B coefficients. Note that the Difference images show the difference in features between the eY and eYS-R&B images. The first row image is from the Alzubaidi dataset, and the second row image is from the FUSC dataset.
• Figure 4: Illustration of (a) the original HarDNet-DFUS convolutional block design found in the encoder stem, and (b) our enhanced block design utilising instance-batch normalisation, PReLU activation, and Switchable Normalisation. BN - batch normalisaton, ReLU - rectified linear unit, IN - instance normalisation, PReLU - parametric rectified linear unit, SN - switchable normalisation.
• Figure 5: Illustration of (a) the original HarDNet-DFUS harmonic block design, and (b) our proposed HarDNet-CWS harmonic block which increases the density of the lower density blocks, and reduces the density of the higher density blocks which results in a reduction of the overall harmonic amplitude. L - number of layers in HarDNet convolution block.